[ { "path": ".github/ISSUE_TEMPLATE/bug_report.yml", "content": "name: \"Bug report\"\ndescription: \"Something isn't working right with NexaSDK\"\nlabels: [\"bug\"]\nbody:\n - type: markdown\n attributes:\n value: |\n Thanks for reporting a bug in NexaSDK 🙏\n\n Before filing:\n - Search existing issues to see if this has already been reported.\n - Try the latest release, if possible.\n\n - type: checkboxes\n id: checks\n attributes:\n label: \"Pre-checks\"\n options:\n - label: \"I searched existing issues\"\n required: true\n - label: \"I’m using the latest NexaSDK release\"\n required: false\n\n - type: textarea\n id: what-happened\n attributes:\n label: \"What happened?\"\n description: \"Describe the bug and what you expected to happen.\"\n placeholder: |\n When I do sth with model X on backend Y, it does Z.\n I expected it to ...\n render: text\n validations:\n required: true\n\n - type: textarea\n id: steps-to-reproduce\n attributes:\n label: \"Steps to reproduce\"\n description: \"Minimal steps or code to trigger the issue.\"\n placeholder: |\n 1. Install NexaSDK using ...\n 2. Run: ...\n 3. See error: ...\n render: markdown\n validations:\n required: true\n\n - type: textarea\n id: logs\n attributes:\n label: \"Logs & stack traces\"\n description: \"Paste any relevant logs or error messages.\"\n placeholder: |\n ```text\n \n ```\n render: text\n\n - type: textarea\n id: models\n attributes:\n label: \"Model(s) and quantization\"\n description: \"Which model(s) and quantization are you using?\"\n placeholder: |\n e.g. Ministral-3-3B-Instruct-2512 Q4_K_M, AutoNeural-SA8295P int4, Qwen2.5 fp16\n render: text\n\n - type: input\n id: nexasdk-version\n attributes:\n label: \"NexaSDK version\"\n description: 'Output of `nexa version` or `python -c \"import nexa; print(nexa.__version__)\"` or similar.'\n placeholder: |\n e.g.\n NexaSDK Bridge Version: v1.0.31\n NexaSDK CLI Version: v0.2.60\n validations:\n required: true\n\n - type: input\n id: install-method\n attributes:\n label: \"Install method\"\n description: \"How did you install NexaSDK?\"\n placeholder: \"pip, conda, docker, from source, other\"\n validations:\n required: true\n\n - type: input\n id: os\n attributes:\n label: \"OS and version\"\n placeholder: \"e.g. Windows 11, Ubuntu 22.04, macOS 15\"\n validations:\n required: true\n\n - type: input\n id: hardware\n attributes:\n label: \"Hardware / accelerator\"\n description: \"CPU/GPU/NPU and device details.\"\n placeholder: \"e.g. RTX 4090, Snapdragon X Elite NPU, SA8295P, M4 Max\"\n\n - type: textarea\n id: extra\n attributes:\n label: \"Additional context\"\n description: \"Anything else that might help (configs, screenshots, links).\"\n render: markdown\n" }, { "path": ".github/ISSUE_TEMPLATE/feature_request.yml", "content": "name: \"Feature request\"\ndescription: \"Request a new feature or improvement for NexaSDK\"\nlabels: [\"enhancement\"]\nbody:\n - type: markdown\n attributes:\n value: |\n Thanks for helping improve NexaSDK 🙌\n\n Use this template for concrete improvements to the SDK.\n\n - type: checkboxes\n id: checks\n attributes:\n label: \"Pre-checks\"\n options:\n - label: \"I searched existing issues and discussions\"\n required: true\n\n - type: textarea\n id: problem\n attributes:\n label: \"What problem are you trying to solve?\"\n description: \"Describe your use case and what’s painful today.\"\n placeholder: |\n I’m building X (e.g. local copilot on Snapdragon PC) and I’m blocked by Y in NexaSDK.\n render: text\n validations:\n required: true\n\n - type: textarea\n id: solution\n attributes:\n label: \"What would you like NexaSDK to do?\"\n description: \"Describe your ideal solution (APIs, CLI flags, configs, etc.).\"\n placeholder: |\n Example:\n - Add a `nexa run --backend npu` option that ...\n - Provide a Python helper `nexa.auto_backend(...)` that ...\n render: markdown\n validations:\n required: true\n\n - type: textarea\n id: alternatives\n attributes:\n label: \"Alternatives you've considered\"\n description: \"Workarounds or other tools you’ve tried.\"\n placeholder: |\n I can currently do this by ..., but it’s slow / complex / fragile.\n render: text\n\n - type: textarea\n id: impact\n attributes:\n label: \"Who does this help, and how much?\"\n description: \"Rough sense of impact (devx, perf, device coverage, etc.).\"\n placeholder: |\n This would be useful for all Windows NPU developers using NexaSDK because ...\n render: text\n\n - type: textarea\n id: extra\n attributes:\n label: \"Additional context\"\n description: \"Links, mock APIs, diagrams, prior art, etc.\"\n render: markdown\n" }, { "path": ".github/ISSUE_TEMPLATE/model_request.yml", "content": "name: \"Model request\"\ndescription: \"Request support for a new model or variant in NexaSDK\"\nlabels: [\"model-request\"]\nbody:\n - type: markdown\n attributes:\n value: |\n Request support for a new model to be added to NexaSDK.\n\n Please include as much detail as you can so we can prioritize properly.\n\n - type: checkboxes\n id: checks\n attributes:\n label: \"Pre-checks\"\n options:\n - label: \"I searched existing issues and discussions for this model\"\n required: true\n\n - type: input\n id: model-name\n attributes:\n label: \"Model name\"\n description: \"Name of the model you’d like supported.\"\n placeholder: \"e.g. Ministral-8x7B, Qwen3-VL-32B, etc.\"\n validations:\n required: true\n\n - type: input\n id: model-source\n attributes:\n label: \"Model source / link\"\n description: \"Link to the model (Hugging Face, GitHub, vendor page, etc.).\"\n placeholder: \"https://huggingface.co/...\"\n validations:\n required: true\n\n - type: textarea\n id: use-cases\n attributes:\n label: \"Primary use cases\"\n description: \"What are you planning to build with this model?\"\n placeholder: |\n e.g. On-device coding assistant, in-car multimodal assistant, document RAG, etc.\n render: text\n validations:\n required: true\n\n - type: textarea\n id: hardware\n attributes:\n label: \"Target hardware / backend\"\n description: \"Where do you want this model to run?\"\n placeholder: |\n e.g. Snapdragon X Elite NPU, SA8295P, RTX 4080, Apple M4, CPU-only, etc.\n render: text\n validations:\n required: true\n\n - type: textarea\n id: extra\n attributes:\n label: \"Additional context\"\n description: \"Benchmarks, papers, licensing notes, or anything else that helps us evaluate.\"\n render: markdown\n" }, { "path": ".github/workflows/build-linux.yml", "content": "name: NexaCLI on Linux\n\non:\n push:\n tags:\n - \"v*\"\n\njobs:\n build-cli:\n name: Build ${{ matrix.runner }}\n runs-on: ${{ matrix.runner }}\n strategy:\n fail-fast: false\n matrix:\n include:\n - runner: ubuntu-22.04\n arch: x86_64\n - runner: ubuntu-22.04-arm\n arch: arm64\n\n env:\n VERSION: ${{ github.ref_name }}\n\n steps:\n - name: Checkout repository\n uses: actions/checkout@v4\n\n - name: Setup Go\n uses: actions/setup-go@v5\n with:\n go-version: 1.25.4\n cache-dependency-path: runner/go.sum\n\n - name: Download runtime\n run: |\n cd runner\n make download\n\n - name: Build Nexa Cli\n run: |\n cd runner\n make build\n rm build/ml.h\n\n - name: Upload Artifact\n uses: actions/upload-artifact@v4\n with:\n name: \"nexa-cli_linux_${{ matrix.arch }}\"\n path: runner/build\n include-hidden-files: true\n\n package-release:\n name: Package Release ${{ matrix.runner }}\n needs: build-cli\n runs-on: ${{ matrix.runner }}\n strategy:\n fail-fast: false\n matrix:\n include:\n - runner: ubuntu-22.04\n arch: x86_64\n - runner: ubuntu-22.04-arm\n arch: arm64\n\n env:\n VERSION: ${{ github.ref_name }}\n\n steps:\n - name: Checkout repository\n uses: actions/checkout@v4\n\n - name: Download build artifacts\n uses: actions/download-artifact@v4\n with:\n name: \"nexa-cli_linux_${{ matrix.arch }}\"\n path: artifacts\n\n - name: Create Self-Contained Installer Script\n id: package\n run: |\n chmod +x runner/release/linux/package.sh\n ./runner/release/linux/package.sh artifacts \\\n runner/release/linux/install.sh \\\n nexa-cli_linux_${{ matrix.arch }}.sh\n\n - name: Create GitHub Release\n uses: softprops/action-gh-release@v2\n with:\n repository: NexaAI/nexa-sdk\n tag_name: ${{ env.VERSION }}\n files: |\n nexa-cli_linux_${{ matrix.arch }}.sh\n draft: ${{ contains(github.ref, '-rc') }}\n generate_release_notes: ${{ matrix.arch == 'x86_64' }}\n make_latest: ${{ !contains(github.ref, '-rc') }}\n env:\n GITHUB_TOKEN: ${{ secrets.ZACK_ACCESS_TOKEN }}\n\n - name: Configure AWS credentials\n if: ${{ !contains(github.ref, '-rc') }}\n uses: aws-actions/configure-aws-credentials@v4\n with:\n aws-access-key-id: ${{ secrets.AWS_ACCESS_KEY_ID }}\n aws-secret-access-key: ${{ secrets.AWS_SECRET_ACCESS_KEY }}\n aws-region: us-west-1\n\n - name: Sync to S3\n if: ${{ !contains(github.ref, '-rc') }}\n shell: bash\n run: |\n aws s3 cp \"nexa-cli_linux_${{ matrix.arch }}.sh\" \"s3://nexa-model-hub-bucket/public/nexa_sdk/downloads/nexa-cli_linux_${{ matrix.arch }}.sh\"\n\n docker-release:\n name: Docker Release\n needs: build-cli\n runs-on: ubuntu-22.04\n\n env:\n VERSION: ${{ github.ref_name }}\n\n steps:\n - name: Checkout repository\n uses: actions/checkout@v4\n\n - name: Download build artifacts\n uses: actions/download-artifact@v4\n with:\n path: artifacts\n\n - name: Prepare Build Context\n run: |\n mkdir -p runner/release/linux\n mv artifacts/nexa-cli_linux_x86_64 runner/release/linux/nexa_sdk.amd64\n mv artifacts/nexa-cli_linux_arm64 runner/release/linux/nexa_sdk.arm64\n chmod +x runner/release/linux/*/nexa\n chmod +x runner/release/linux/*/nexa-cli\n\n - name: Login to Docker Hub\n uses: docker/login-action@v3\n with:\n username: ${{ secrets.DOCKERHUB_USERNAME }}\n password: ${{ secrets.DOCKERHUB_TOKEN }}\n\n - name: Set up QEMU\n uses: docker/setup-qemu-action@v3\n\n - name: Set up Docker Buildx\n uses: docker/setup-buildx-action@v3\n\n - name: Build and Push Docker Image (Base)\n uses: docker/build-push-action@v6\n with:\n context: runner/release/linux\n push: true\n platforms: linux/amd64,linux/arm64\n tags: |\n nexa4ai/nexasdk${{ contains(github.ref, '-rc') && '-test' || '' }}:${{ env.VERSION }}\n ${{ !contains(github.ref, '-rc') && 'nexa4ai/nexasdk:latest' || '' }}\n\n - name: Build and Push Docker Image (Cuda)\n uses: docker/build-push-action@v6\n with:\n context: runner/release/linux\n push: true\n platforms: linux/amd64\n build-args: |\n BASE_IMAGE=nvidia/cuda:12.9.1-runtime-ubuntu22.04\n tags: |\n nexa4ai/nexasdk${{ contains(github.ref, '-rc') && '-test' || '' }}:${{ env.VERSION }}-cuda\n ${{ !contains(github.ref, '-rc') && 'nexa4ai/nexasdk:latest-cuda' || '' }}\n" }, { "path": ".github/workflows/build-macos.yml", "content": "name: NexaCLI on MacOS\n\non:\n push:\n tags:\n - \"v*\"\n\njobs:\n build-cli:\n name: macos_${{ matrix.arch }}\n runs-on: ${{ matrix.runner }}\n strategy:\n fail-fast: false\n matrix:\n include:\n - runner: macos-15-intel\n arch: x86_64\n - runner: macos-15\n arch: arm64\n\n env:\n VERSION: ${{ github.ref_name }}\n\n steps:\n - name: Checkout repository\n uses: actions/checkout@v4\n\n - uses: actions/setup-go@v5\n with:\n go-version: 1.25.4\n cache-dependency-path: runner/go.sum\n\n - name: Download runtime\n run: |\n cd runner\n make download\n\n - name: Build Nexa Cli\n run: |\n cd runner\n make build\n rm build/ml.h\n\n - name: Upload Artifact\n uses: actions/upload-artifact@v4\n with:\n name: nexa-cli_macos_${{ matrix.arch }}\n path: runner/build\n include-hidden-files: true\n\n package-and-release:\n name: Package for macos_${{ matrix.arch }}\n needs: build-cli\n runs-on: ${{ matrix.runner }}\n strategy:\n fail-fast: false\n matrix:\n include:\n - runner: macos-15-intel\n arch: x86_64\n - runner: macos-15\n arch: arm64\n env:\n VERSION: ${{ github.ref_name }}\n\n steps:\n - name: Checkout repository\n uses: actions/checkout@v4\n\n - name: Download build artifacts\n uses: actions/download-artifact@v4\n with:\n name: nexa-cli_macos_${{ matrix.arch }}\n path: artifacts\n\n - name: Prepare file structure for packaging\n id: prep_files\n run: |\n chmod +x runner/release/darwin/scripts/prepackage.sh\n bash runner/release/darwin/scripts/prepackage.sh \"${{ env.VERSION }}\"\n\n - name: Fix dylib Linkages (RPATH)\n run: |\n RESOURCES_PATH=\"${{ env.APP_PATH }}/Contents/Resources\"\n install_name_tool -add_rpath \"@loader_path\" \"${RESOURCES_PATH}/nexa-cli\"\n\n - name: Import Code Signing Certificates\n env:\n APP_CERTIFICATE_BASE64: ${{ secrets.APPLE_CERTIFICATE_BASE64 }}\n APP_CERTIFICATE_PASSWORD: ${{ secrets.APPLE_CERTIFICATE_PASSWORD }}\n INSTALLER_CERTIFICATE_BASE64: ${{ secrets.APPLE_INSTALLER_CERTIFICATE_BASE64 }}\n INSTALLER_CERTIFICATE_PASSWORD: ${{ secrets.APPLE_CERTIFICATE_PASSWORD }}\n run: |\n echo $APP_CERTIFICATE_BASE64 | base64 --decode > app_certificate.p12\n echo $INSTALLER_CERTIFICATE_BASE64 | base64 --decode > installer_certificate.p12\n security create-keychain -p \"\" build.keychain\n security default-keychain -s build.keychain\n security unlock-keychain -p \"\" build.keychain\n security import app_certificate.p12 -k build.keychain -P $APP_CERTIFICATE_PASSWORD -T /usr/bin/codesign\n security import installer_certificate.p12 -k build.keychain -P $INSTALLER_CERTIFICATE_PASSWORD -T /usr/bin/productsign\n security set-key-partition-list -S apple-tool:,apple:,codesign: -s -k \"\" build.keychain\n\n - name: Sign binaries and libraries\n run: |\n chmod +x runner/release/darwin/scripts/sign.sh\n bash runner/release/darwin/scripts/sign.sh \"${{ env.APP_PATH }}\" \"${{ secrets.APPLE_ID_APPLICATION }}\"\n\n - name: Build PKG\n run: |\n pkgbuild --root \"${{ steps.prep_files.outputs.STAGING_DIR }}\" \\\n --scripts \"${{ steps.prep_files.outputs.SCRIPTS_DIR }}\" \\\n --identifier \"com.nexaai.nexa-sdk\" \\\n --version \"${{ env.VERSION }}\" \\\n --install-location / \\\n \"artifacts/nexa-cli_macos_${{ matrix.arch }}-unsigned.pkg\"\n\n - name: Productsign PKG\n run: |\n productsign --sign \"${{ secrets.APPLE_ID_INSTALLER }}\" --timestamp \"artifacts/nexa-cli_macos_${{ matrix.arch }}-unsigned.pkg\" \"artifacts/nexa-cli_macos_${{ matrix.arch }}.pkg\"\n pkgutil --check-signature \"artifacts/nexa-cli_macos_${{ matrix.arch }}.pkg\"\n rm \"artifacts/nexa-cli_macos_${{ matrix.arch }}-unsigned.pkg\"\n\n - name: Notarize & Staple PKG\n run: |\n chmod +x runner/release/darwin/scripts/notarize.sh\n bash runner/release/darwin/scripts/notarize.sh \\\n \"artifacts/nexa-cli_macos_${{ matrix.arch }}.pkg\" \\\n \"${{ secrets.APPLE_ID }}\" \\\n \"${{ secrets.APPLE_ID_PASSWORD }}\" \\\n \"${{ secrets.APPLE_TEAM_ID }}\"\n\n - name: Create GitHub Release\n uses: softprops/action-gh-release@v2\n with:\n repository: NexaAI/nexa-sdk\n tag_name: ${{ env.VERSION }}\n files: |\n artifacts/nexa-cli_macos_${{ matrix.arch }}.pkg\n draft: ${{ contains(github.ref, '-rc') }}\n generate_release_notes: false\n make_latest: ${{ !contains(github.ref, '-rc') }}\n env:\n GITHUB_TOKEN: ${{ secrets.ZACK_ACCESS_TOKEN }}\n\n - name: Configure AWS credentials\n if: ${{ !contains(github.ref, '-rc') }}\n uses: aws-actions/configure-aws-credentials@v4\n with:\n aws-access-key-id: ${{ secrets.AWS_ACCESS_KEY_ID }}\n aws-secret-access-key: ${{ secrets.AWS_SECRET_ACCESS_KEY }}\n aws-region: us-west-1\n\n - name: Sync to S3\n if: ${{ !contains(github.ref, '-rc') }}\n shell: bash\n run: |\n aws s3 cp \"artifacts/nexa-cli_macos_${{ matrix.arch }}.pkg\" \"s3://nexa-model-hub-bucket/public/nexa_sdk/downloads/nexa-cli_macos_${{ matrix.arch }}.pkg\"\n" }, { "path": ".github/workflows/build-windows.yml", "content": "name: NexaCLI on Windows\n\non:\n push:\n tags:\n - \"v*\"\n\njobs:\n build-cli:\n name: windows_${{ matrix.arch }}\n runs-on: ${{ matrix.runner }}\n strategy:\n fail-fast: false\n matrix:\n include:\n - runner: sdk-windows-8-core-2\n arch: x86_64\n - runner: windows-11-arm\n arch: arm64\n\n env:\n VERSION: ${{ github.ref_name }}\n\n steps:\n - name: Checkout repository\n uses: actions/checkout@v4\n\n - name: Setup Go\n uses: actions/setup-go@v5\n with:\n go-version: 1.25.4\n cache-dependency-path: runner/go.sum\n\n - name: Setup MSYS2 (ARM64)\n if: matrix.arch == 'arm64'\n uses: msys2/setup-msys2@v2\n with:\n msystem: CLANGARM64\n location: C:/Tools\n update: true\n install: mingw-w64-clang-aarch64-clang\n\n - name: Setup Compile Environment (ARM64)\n if: matrix.arch == 'arm64'\n shell: bash\n run: |\n echo \"CC=C:/Tools/msys64/clangarm64/bin/clang.exe\" >> $GITHUB_ENV\n echo \"CXX=C:/Tools/msys64/clangarm64/bin/clang++.exe\" >> $GITHUB_ENV\n\n - name: Download runtime\n shell: bash\n run: |\n cd runner\n make download ARCH=${{ matrix.arch }}\n ls build\n\n - name: Build Nexa Cli\n shell: bash\n run: |\n cd runner\n make build ARCH=${{ matrix.arch }}\n rm -f build/ml.h\n\n - name: Upload Artifact\n uses: actions/upload-artifact@v4\n with:\n name: nexa-cli_windows_${{ matrix.arch }}\n path: runner/build\n include-hidden-files: true\n\n build-installer:\n name: Build Windows Installer & Release\n needs: build-cli\n runs-on: ${{ matrix.runner }}\n strategy:\n fail-fast: false\n matrix:\n include:\n - runner: sdk-windows-8-core-2\n arch: x86_64\n - runner: windows-11-arm\n arch: arm64\n env:\n VERSION: ${{ github.ref_name }}\n ARCH: ${{ matrix.arch }}\n\n steps:\n - name: Checkout repository\n uses: actions/checkout@v4\n\n - name: Download build artifacts\n uses: actions/download-artifact@v4\n with:\n name: nexa-cli_windows_${{ matrix.arch }}\n path: artifacts\n\n - name: Build launcher with PyInstaller\n run: |\n pip install pyinstaller\n pyinstaller --onefile --noconsole --distpath artifacts --name \"nexa-cli-launcher\" --icon=\"runner/release/windows/nexa_logo.ico\" runner/release/windows/nexa_launcher.py\n\n # === Install Inno Setup (before compiling installers) ===\n - name: Install Inno Setup\n shell: pwsh\n run: |\n if (Get-Command ISCC.exe -ErrorAction SilentlyContinue) { return }\n if (Get-Command choco -ErrorAction SilentlyContinue) {\n choco install innosetup -y --no-progress\n } else {\n winget install --id JRSoftware.InnoSetup -e --accept-source-agreements --accept-package-agreements\n }\n $inno = \"${env:ProgramFiles(x86)}\\Inno Setup 6\"\n if (Test-Path \"$inno\\ISCC.exe\") {\n \"$inno\" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append\n }\n\n - name: Sign executables with AzureSignTool\n shell: pwsh\n run: |\n dotnet tool install --global AzureSignTool\n\n Get-ChildItem -Path \"artifacts\" -Recurse -Include *.exe, *.dll | ForEach-Object {\n azuresigntool sign `\n -kvu \"${{ secrets.AZURE_KEY_VAULT_URI }}\" `\n -kvc \"${{ secrets.AZURE_CERT_NAME }}\" `\n -kvi \"${{ secrets.AZURE_CLIENT_ID }}\" `\n -kvs \"${{ secrets.AZURE_CLIENT_SECRET }}\" `\n --azure-key-vault-tenant-id \"${{ secrets.AZURE_TENANT_ID }}\" `\n -tr http://timestamp.globalsign.com/tsa/advanced `\n -td sha256 `\n $_.FullName\n }\n\n - name: Prepare Installer Files (X86_64)\n if: matrix.arch == 'x86_64'\n shell: bash\n run: |\n cp -r artifacts artifacts_cuda\n rm -r artifacts_cuda/nexaai*\n rm -r artifacts/nexa_cuda*\n\n - name: Compile Installer (CUDA) (X86_64)\n if: matrix.arch == 'x86_64'\n uses: Minionguyjpro/Inno-Setup-Action@v1.2.6\n with:\n path: runner/release/windows/nexa_installer.iss\n options: /O+\n env:\n ORT: \"_cuda\"\n\n - name: Compile Installer\n uses: Minionguyjpro/Inno-Setup-Action@v1.2.6\n with:\n path: runner/release/windows/nexa_installer.iss\n options: /O+\n env:\n ORT: \"\"\n\n - name: Sign Installer (ARM64)\n if: matrix.arch == 'arm64'\n shell: pwsh\n run: |\n azuresigntool sign `\n -kvu \"${{ secrets.AZURE_KEY_VAULT_URI }}\" `\n -kvc \"${{ secrets.AZURE_CERT_NAME }}\" `\n -kvi \"${{ secrets.AZURE_CLIENT_ID }}\" `\n -kvs \"${{ secrets.AZURE_CLIENT_SECRET }}\" `\n --azure-key-vault-tenant-id \"${{ secrets.AZURE_TENANT_ID }}\" `\n -tr http://timestamp.globalsign.com/tsa/advanced `\n -td sha256 `\n \"nexa-cli_windows_${{ matrix.arch }}.exe\"\n\n - name: Sign Installer (X86_64)\n if: matrix.arch == 'x86_64'\n run: |\n azuresigntool sign `\n -kvu \"${{ secrets.AZURE_KEY_VAULT_URI }}\" `\n -kvc \"${{ secrets.AZURE_CERT_NAME }}\" `\n -kvi \"${{ secrets.AZURE_CLIENT_ID }}\" `\n -kvs \"${{ secrets.AZURE_CLIENT_SECRET }}\" `\n --azure-key-vault-tenant-id \"${{ secrets.AZURE_TENANT_ID }}\" `\n -tr http://timestamp.globalsign.com/tsa/advanced `\n -td sha256 `\n \"nexa-cli_windows_${{ matrix.arch }}_cuda.exe\"\n\n azuresigntool sign `\n -kvu \"${{ secrets.AZURE_KEY_VAULT_URI }}\" `\n -kvc \"${{ secrets.AZURE_CERT_NAME }}\" `\n -kvi \"${{ secrets.AZURE_CLIENT_ID }}\" `\n -kvs \"${{ secrets.AZURE_CLIENT_SECRET }}\" `\n --azure-key-vault-tenant-id \"${{ secrets.AZURE_TENANT_ID }}\" `\n -tr http://timestamp.globalsign.com/tsa/advanced `\n -td sha256 `\n \"nexa-cli_windows_${{ matrix.arch }}.exe\"\n\n - name: Create GitHub Release\n uses: softprops/action-gh-release@v2\n with:\n repository: NexaAI/nexa-sdk\n tag_name: ${{ env.VERSION }}\n files: |\n nexa-cli_windows_${{ matrix.arch }}_cuda.exe\n nexa-cli_windows_${{ matrix.arch }}.exe\n draft: ${{ contains(github.ref, '-rc') }}\n generate_release_notes: false\n make_latest: ${{ !contains(github.ref, '-rc') }}\n env:\n GITHUB_TOKEN: ${{ secrets.ZACK_ACCESS_TOKEN }}\n\n - name: Configure AWS credentials\n if: ${{ !contains(github.ref, '-rc') }}\n uses: aws-actions/configure-aws-credentials@v4\n with:\n aws-access-key-id: ${{ secrets.AWS_ACCESS_KEY_ID }}\n aws-secret-access-key: ${{ secrets.AWS_SECRET_ACCESS_KEY }}\n aws-region: us-west-1\n\n - name: Sync to S3\n if: ${{ !contains(github.ref, '-rc') }}\n shell: bash\n run: |\n aws s3 cp \"nexa-cli_windows_${{ matrix.arch }}.exe\" \"s3://nexa-model-hub-bucket/public/nexa_sdk/downloads/nexa-cli_windows_${{ matrix.arch }}.exe\"\n\n - name: Sync to S3 (CUDA) (X86_64)\n if: ${{ !contains(github.ref, '-rc') && matrix.arch == 'x86_64' }}\n shell: bash\n run: |\n aws s3 cp \"nexa-cli_windows_${{ matrix.arch }}_cuda.exe\" \"s3://nexa-model-hub-bucket/public/nexa_sdk/downloads/nexa-cli_windows_${{ matrix.arch }}_cuda.exe\"\n" }, { "path": ".github/workflows/lint.yml", "content": "name: Lint\n\non:\n pull_request:\n branches: [main]\n paths:\n - 'runner/**'\n - '.github/workflows/lint.yml'\n\njobs:\n lint:\n runs-on: ubuntu-22.04\n steps:\n - uses: actions/checkout@v4\n - uses: actions/setup-go@v5\n with:\n go-version: '1.25'\n cache-dependency-path: runner/go.sum\n - name: go fmt\n working-directory: runner\n run: |\n unformatted=$(gofmt -s -l .)\n [ -z \"$unformatted\" ] || (echo \"Unformatted: $unformatted\"; exit 1)\n - name: Download runtime\n working-directory: runner\n run: make download\n - name: go vet\n working-directory: runner\n run: go vet ./...\n" }, { "path": ".github/workflows/test.yml", "content": "name: Test\n\non:\n # pull_request:\n # branches: [main, main-test]\n workflow_dispatch:\n\njobs:\n test-llm:\n name: ${{ matrix.os }} / ${{ matrix.backend }}\n runs-on: ${{ matrix.os }}\n strategy:\n fail-fast: false\n matrix:\n os: [macos-15]\n backend: [llama-cpp-metal]\n steps:\n - name: Checkout code\n uses: actions/checkout@v4\n\n - name: Setup Go\n uses: actions/setup-go@v5\n with:\n go-version: 1.24.3\n cache-dependency-path: runner/go.sum\n\n - name: Download backend runtime\n run: make download BRIDGE_BACKEND=${{ matrix.backend }}\n\n - name: Cache Nexa model files\n uses: actions/cache@v4\n with:\n path: ~/.cache/models/Qwen3-0.6B-Q8_0.gguf\n key: Qwen3-0.6B-Q8_0.gguf\n restore-keys: |\n Qwen3-0.6B-Q8_0.gguf\n\n - name: Download model\n shell: bash\n run: |\n MODEL_PATH=\"$HOME/.cache/models/Qwen3-0.6B-Q8_0.gguf\"\n if [ ! -f \"$MODEL_PATH\" ]; then\n mkdir -p \"$(dirname \"$MODEL_PATH\")\"\n wget -nv -O \"$MODEL_PATH\" https://huggingface.co/nexaml/Qwen3-0.6B/resolve/main/Qwen3-0.6B-Q8_0.gguf\n else\n echo \"Model already cached, skip download.\"\n fi\n\n - name: Test LLM Multi-Turn Chat\n shell: bash\n run: |\n if [[ \"$RUNNER_OS\" == \"macOS\" ]]; then\n export DYLD_LIBRARY_PATH=\"$PWD/build/lib/${{ matrix.backend }}\"\n elif [[ \"$RUNNER_OS\" == \"Linux\" ]]; then\n export LD_LIBRARY_PATH=\"$PWD/build/lib/${{ matrix.backend }}\"\n elif [[ \"$RUNNER_OS\" == \"Windows\" ]]; then\n export PATH=\"$PWD/build/lib/${{ matrix.backend }};$PATH\"\n fi\n export CGO_LDFLAGS=\"-L$PWD/build/lib/${{ matrix.backend }}\"\n cd runner/nexa-sdk && go test -v -run \"Test.*LLM|TestChatMultiTurn\"\n" }, { "path": ".gitignore", "content": ".DS_Store\n.vscode/\ntempnexaai-env/\n" }, { "path": "LICENSE", "content": " Apache License\n Version 2.0, January 2004\n http://www.apache.org/licenses/\n\n TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION\n\n 1. Definitions.\n\n \"License\" shall mean the terms and conditions for use, reproduction,\n and distribution as defined by Sections 1 through 9 of this document.\n\n \"Licensor\" shall mean the copyright owner or entity authorized by\n the copyright owner that is granting the License.\n\n \"Legal Entity\" shall mean the union of the acting entity and all\n other entities that control, are controlled by, or are under common\n control with that entity. For the purposes of this definition,\n \"control\" means (i) the power, direct or indirect, to cause the\n direction or management of such entity, whether by contract or\n otherwise, or (ii) ownership of fifty percent (50%) or more of the\n outstanding shares, or (iii) beneficial ownership of such entity.\n\n \"You\" (or \"Your\") shall mean an individual or Legal Entity\n exercising permissions granted by this License.\n\n \"Source\" form shall mean the preferred form for making modifications,\n including but not limited to software source code, documentation\n source, and configuration files.\n\n \"Object\" form shall mean any form resulting from mechanical\n transformation or translation of a Source form, including but\n not limited to compiled object code, generated documentation,\n and conversions to other media types.\n\n \"Work\" shall mean the work of authorship, whether in Source or\n Object form, made available under the License, as indicated by a\n copyright notice that is included in or attached to the work\n (an example is provided in the Appendix below).\n\n \"Derivative Works\" shall mean any work, whether in Source or Object\n form, that is based on (or derived from) the Work and for which the\n editorial revisions, annotations, elaborations, or other modifications\n represent, as a whole, an original work of authorship. For the purposes\n of this License, Derivative Works shall not include works that remain\n separable from, or merely link (or bind by name) to the interfaces of,\n the Work and Derivative Works thereof.\n\n \"Contribution\" shall mean any work of authorship, including\n the original version of the Work and any modifications or additions\n to that Work or Derivative Works thereof, that is intentionally\n submitted to Licensor for inclusion in the Work by the copyright owner\n or by an individual or Legal Entity authorized to submit on behalf of\n the copyright owner. For the purposes of this definition, \"submitted\"\n means any form of electronic, verbal, or written communication sent\n to the Licensor or its representatives, including but not limited to\n communication on electronic mailing lists, source code control systems,\n and issue tracking systems that are managed by, or on behalf of, the\n Licensor for the purpose of discussing and improving the Work, but\n excluding communication that is conspicuously marked or otherwise\n designated in writing by the copyright owner as \"Not a Contribution.\"\n\n \"Contributor\" shall mean Licensor and any individual or Legal Entity\n on behalf of whom a Contribution has been received by Licensor and\n subsequently incorporated within the Work.\n\n 2. Grant of Copyright License. Subject to the terms and conditions of\n this License, each Contributor hereby grants to You a perpetual,\n worldwide, non-exclusive, no-charge, royalty-free, irrevocable\n copyright license to reproduce, prepare Derivative Works of,\n publicly display, publicly perform, sublicense, and distribute the\n Work and such Derivative Works in Source or Object form.\n\n 3. Grant of Patent License. Subject to the terms and conditions of\n this License, each Contributor hereby grants to You a perpetual,\n worldwide, non-exclusive, no-charge, royalty-free, irrevocable\n (except as stated in this section) patent license to make, have made,\n use, offer to sell, sell, import, and otherwise transfer the Work,\n where such license applies only to those patent claims licensable\n by such Contributor that are necessarily infringed by their\n Contribution(s) alone or by combination of their Contribution(s)\n with the Work to which such Contribution(s) was submitted. If You\n institute patent litigation against any entity (including a\n cross-claim or counterclaim in a lawsuit) alleging that the Work\n or a Contribution incorporated within the Work constitutes direct\n or contributory patent infringement, then any patent licenses\n granted to You under this License for that Work shall terminate\n as of the date such litigation is filed.\n\n 4. Redistribution. You may reproduce and distribute copies of the\n Work or Derivative Works thereof in any medium, with or without\n modifications, and in Source or Object form, provided that You\n meet the following conditions:\n\n (a) You must give any other recipients of the Work or\n Derivative Works a copy of this License; and\n\n (b) You must cause any modified files to carry prominent notices\n stating that You changed the files; and\n\n (c) You must retain, in the Source form of any Derivative Works\n that You distribute, all copyright, patent, trademark, and\n attribution notices from the Source form of the Work,\n excluding those notices that do not pertain to any part of\n the Derivative Works; and\n\n (d) If the Work includes a \"NOTICE\" text file as part of its\n distribution, then any Derivative Works that You distribute must\n include a readable copy of the attribution notices contained\n within such NOTICE file, excluding those notices that do not\n pertain to any part of the Derivative Works, in at least one\n of the following places: within a NOTICE text file distributed\n as part of the Derivative Works; within the Source form or\n documentation, if provided along with the Derivative Works; or,\n within a display generated by the Derivative Works, if and\n wherever such third-party notices normally appear. The contents\n of the NOTICE file are for informational purposes only and\n do not modify the License. You may add Your own attribution\n notices within Derivative Works that You distribute, alongside\n or as an addendum to the NOTICE text from the Work, provided\n that such additional attribution notices cannot be construed\n as modifying the License.\n\n You may add Your own copyright statement to Your modifications and\n may provide additional or different license terms and conditions\n for use, reproduction, or distribution of Your modifications, or\n for any such Derivative Works as a whole, provided Your use,\n reproduction, and distribution of the Work otherwise complies with\n the conditions stated in this License.\n\n 5. Submission of Contributions. Unless You explicitly state otherwise,\n any Contribution intentionally submitted for inclusion in the Work\n by You to the Licensor shall be under the terms and conditions of\n this License, without any additional terms or conditions.\n Notwithstanding the above, nothing herein shall supersede or modify\n the terms of any separate license agreement you may have executed\n with Licensor regarding such Contributions.\n\n 6. Trademarks. This License does not grant permission to use the trade\n names, trademarks, service marks, or product names of the Licensor,\n except as required for reasonable and customary use in describing the\n origin of the Work and reproducing the content of the NOTICE file.\n\n 7. Disclaimer of Warranty. Unless required by applicable law or\n agreed to in writing, Licensor provides the Work (and each\n Contributor provides its Contributions) on an \"AS IS\" BASIS,\n WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or\n implied, including, without limitation, any warranties or conditions\n of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A\n PARTICULAR PURPOSE. You are solely responsible for determining the\n appropriateness of using or redistributing the Work and assume any\n risks associated with Your exercise of permissions under this License.\n\n 8. Limitation of Liability. In no event and under no legal theory,\n whether in tort (including negligence), contract, or otherwise,\n unless required by applicable law (such as deliberate and grossly\n negligent acts) or agreed to in writing, shall any Contributor be\n liable to You for damages, including any direct, indirect, special,\n incidental, or consequential damages of any character arising as a\n result of this License or out of the use or inability to use the\n Work (including but not limited to damages for loss of goodwill,\n work stoppage, computer failure or malfunction, or any and all\n other commercial damages or losses), even if such Contributor\n has been advised of the possibility of such damages.\n\n 9. Accepting Warranty or Additional Liability. While redistributing\n the Work or Derivative Works thereof, You may choose to offer,\n and charge a fee for, acceptance of support, warranty, indemnity,\n or other liability obligations and/or rights consistent with this\n License. However, in accepting such obligations, You may act only\n on Your own behalf and on Your sole responsibility, not on behalf\n of any other Contributor, and only if You agree to indemnify,\n defend, and hold each Contributor harmless for any liability\n incurred by, or claims asserted against, such Contributor by reason\n of your accepting any such warranty or additional liability.\n\n END OF TERMS AND CONDITIONS\n\n APPENDIX: How to apply the Apache License to your work.\n\n To apply the Apache License to your work, attach the following\n boilerplate notice, with the fields enclosed by brackets \"[]\"\n replaced with your own identifying information. (Don't include\n the brackets!) The text should be enclosed in the appropriate\n comment syntax for the file format. We also recommend that a\n file or class name and description of purpose be included on the\n same \"printed page\" as the copyright notice for easier\n identification within third-party archives.\n\n Copyright 2024-2026 Nexa AI, Inc.\n\n Licensed under the Apache License, Version 2.0 (the \"License\");\n you may not use this file except in compliance with the License.\n You may obtain a copy of the License at\n\n http://www.apache.org/licenses/LICENSE-2.0\n\n Unless required by applicable law or agreed to in writing, software\n distributed under the License is distributed on an \"AS IS\" BASIS,\n WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n See the License for the specific language governing permissions and\n limitations under the License.\n" }, { "path": "NOTICE", "content": "# Third-Party Licenses\n\nThis document provides information about third-party libraries used in the NexaSDK project and their respective licenses. All libraries listed below are licensed under permissive open-source licenses (MIT, Apache 2.0, BSD, or equivalent) that are compatible with NexaSDK's Apache 2.0 license.\n\n---\n\n## Go Dependencies\n\n### gin-gonic/gin\n\n- **License**: MIT License\n- **URL**: https://github.com/gin-gonic/gin\n- **Description**: HTTP web framework written in Go\n\n### spf13/cobra\n\n- **License**: Apache License 2.0\n- **URL**: https://github.com/spf13/cobra\n- **Description**: Commander for modern Go CLI applications\n\n### spf13/viper\n\n- **License**: MIT License\n- **URL**: https://github.com/spf13/viper\n- **Description**: Go configuration with fangs\n\n### AWS SDK for Go v2\n\n- **License**: Apache License 2.0\n- **URL**: https://github.com/aws/aws-sdk-go-v2\n- **Description**: AWS SDK for the Go programming language\n- **Packages**: aws-sdk-go-v2, aws-sdk-go-v2/config, aws-sdk-go-v2/service/s3\n\n### bytedance/sonic\n\n- **License**: Apache License 2.0\n- **URL**: https://github.com/bytedance/sonic\n- **Description**: Blazingly fast JSON serializing & deserializing library\n\n### charmbracelet/huh\n\n- **License**: MIT License\n- **URL**: https://github.com/charmbracelet/huh\n- **Description**: Build terminal forms and prompts\n\n### charmbracelet/lipgloss\n\n- **License**: MIT License\n- **URL**: https://github.com/charmbracelet/lipgloss\n- **Description**: Style definitions for nice terminal layouts\n\n### charmbracelet/bubbletea\n\n- **License**: MIT License\n- **URL**: https://github.com/charmbracelet/bubbletea\n- **Description**: Powerful little TUI framework\n\n### dustin/go-humanize\n\n- **License**: MIT License\n- **URL**: https://github.com/dustin/go-humanize\n- **Description**: Formatters for units to human friendly sizes\n\n### gofrs/flock\n\n- **License**: BSD 3-Clause License\n- **URL**: https://github.com/gofrs/flock\n- **Description**: Thread-safe file locking library\n\n### chzyer/readline\n\n- **License**: MIT License\n- **URL**: https://github.com/chzyer/readline\n- **Description**: Pure Go readline library with support for command line editing and history\n\n### openai/openai-go\n\n- **License**: Apache License 2.0\n- **URL**: https://github.com/openai/openai-go\n- **Description**: OpenAI Go API library\n\n### schollz/progressbar\n\n- **License**: MIT License\n- **URL**: https://github.com/schollz/progressbar\n- **Description**: Simple command line progress bar\n\n### valyala/fasthttp\n\n- **License**: MIT License\n- **URL**: https://github.com/valyala/fasthttp\n- **Description**: Fast HTTP implementation for Go\n\n---\n\n## C/C++ Dependencies\n\n### ggml\n\n- **License**: MIT License\n- **Copyright**: Copyright (c) 2023-2024 The ggml authors\n- **URL**: https://github.com/ggml-org/ggml\n- **Description**: Tensor library for machine learning\n- **License File**: https://github.com/ggml-org/ggml/blob/master/LICENSE\n\n### pybind11\n\n- **License**: BSD 3-Clause License\n- **Copyright**: Copyright (c) 2016 Wenzel Jakob\n- **URL**: https://github.com/pybind/pybind11\n- **Description**: Seamless operability between C++11 and Python\n- **License File**: [nexasdk-bridge/third-party/pybind11/LICENSE](https://github.com/pybind/pybind11/blob/master/LICENSE)\n\n### mlx-audio\n\n- **License**: MIT License\n- **Copyright**: Copyright (c) 2024 Prince Canuma\n- **URL**: https://github.com/Blaizzy/mlx-audio\n- **Description**: Audio processing with MLX\n- **License File**: [nexasdk-bridge/third-party/mlx-audio/LICENSE](https://github.com/Blaizzy/mlx-audio/blob/main/LICENSE)\n\n### mlx-vlm\n\n- **License**: MIT License\n- **Copyright**: Copyright © 2023 Apple Inc.\n- **URL**: https://github.com/Blaizzy/mlx-vlm\n- **Description**: Vision-language models with MLX\n- **License File**: [nexasdk-bridge/third-party/mlx-vlm/LICENSE](https://github.com/Blaizzy/mlx-vlm/blob/main/LICENSE)\n\n### xtensor\n\n- **License**: BSD 3-Clause License\n- **Copyright**: Copyright (c) 2016 Johan Mabille, Sylvain Corlay and Wolf Vollprecht, Copyright (c) 2016 QuantStack\n- **URL**: https://github.com/xtensor-stack/xtensor\n- **Description**: C++ tensors with broadcasting and lazy computing\n- **License File**: https://github.com/xtensor-stack/xtensor/blob/master/LICENSE\n\n### cpp-httplib\n\n- **License**: MIT License\n- **Copyright**: Copyright (c) 2017 yhirose\n- **URL**: https://github.com/yhirose/cpp-httplib\n- **Description**: C++ single-file header-only HTTP/HTTPS library\n- **Used in**: llama.cpp\n- **License File**: [nexasdk-bridge/third-party/llama.cpp/licenses/LICENSE-httplib](https://github.com/ggerganov/llama.cpp/blob/master/licenses/LICENSE-httplib)\n\n### nlohmann/json\n\n- **License**: MIT License\n- **Copyright**: Copyright (c) 2013-2025 Niels Lohmann\n- **URL**: https://github.com/nlohmann/json\n- **Description**: JSON for Modern C++\n- **Used in**: llama.cpp\n- **License File**: [nexasdk-bridge/third-party/llama.cpp/licenses/LICENSE-jsonhpp](https://github.com/ggerganov/llama.cpp/blob/master/licenses/LICENSE-jsonhpp)\n\n---\n\n## Python Dependencies\n\n### gradio\n\n- **License**: Apache License 2.0\n- **URL**: https://github.com/gradio-app/gradio\n- **Description**: Build machine learning web apps\n\n### requests\n\n- **License**: Apache License 2.0\n- **URL**: https://github.com/psf/requests\n- **Description**: HTTP library for Python\n\n### Flask\n\n- **License**: BSD 3-Clause License\n- **URL**: https://github.com/pallets/flask\n- **Description**: Lightweight WSGI web application framework\n\n### faiss-cpu\n\n- **License**: MIT License\n- **URL**: https://github.com/facebookresearch/faiss\n- **Description**: Library for efficient similarity search and clustering of dense vectors\n\n### sentence-transformers\n\n- **License**: Apache License 2.0\n- **URL**: https://github.com/UKPLab/sentence-transformers\n- **Description**: Compute dense vector representations for sentences and paragraphs\n\n### langchain\n\n- **License**: MIT License\n- **URL**: https://github.com/langchain-ai/langchain\n- **Description**: Building applications with LLMs through composability\n\n### pydantic\n\n- **License**: MIT License\n- **URL**: https://github.com/pydantic/pydantic\n- **Description**: Data validation using Python type hints\n\n### python-docx\n\n- **License**: MIT License\n- **URL**: https://github.com/python-openxml/python-docx\n- **Description**: Create and modify Word documents with Python\n\n### psutil\n\n- **License**: BSD 3-Clause License\n- **URL**: https://github.com/giampaolo/psutil\n- **Description**: Cross-platform library for process and system monitoring\n\n### openai\n\n- **License**: Apache License 2.0\n- **URL**: https://github.com/openai/openai-python\n- **Description**: Python library for the OpenAI API\n\n### mcp\n\n- **License**: MIT License\n- **URL**: https://github.com/modelcontextprotocol/python-sdk\n- **Description**: Model Context Protocol Python SDK\n\n---\n\n## Android/Java Dependencies (from bindings/android)\n\n### AndroidX Libraries\n\n- **License**: Apache License 2.0\n- **URL**: https://developer.android.com/jetpack/androidx\n- **Description**: Android Jetpack libraries\n- **Packages**: core-ktx, lifecycle-runtime-ktx, activity-compose, compose-bom, appcompat, constraintlayout\n\n### Kotlin Standard Library\n\n- **License**: Apache License 2.0\n- **URL**: https://github.com/JetBrains/kotlin\n- **Description**: Kotlin programming language standard library\n\n### AWS Android SDK S3\n\n- **License**: Apache License 2.0\n- **URL**: https://github.com/aws/aws-sdk-android\n- **Description**: AWS SDK for Android - S3 support\n\n### OkHttp\n\n- **License**: Apache License 2.0\n- **URL**: https://github.com/square/okhttp\n- **Description**: HTTP client for Android and Java\n\n### Gson\n\n- **License**: Apache License 2.0\n- **URL**: https://github.com/google/gson\n- **Description**: Java serialization/deserialization library to convert Java Objects into JSON\n\n### Glide\n\n- **License**: BSD, MIT, and Apache License 2.0 (Multi-license)\n- **URL**: https://github.com/bumptech/glide\n- **Description**: Fast and efficient image loading library for Android\n\n### Kotlinx Serialization\n\n- **License**: Apache License 2.0\n- **URL**: https://github.com/Kotlin/kotlinx.serialization\n- **Description**: Multiplatform serialization for Kotlin\n\n### Kotlinx Coroutines\n\n- **License**: Apache License 2.0\n- **URL**: https://github.com/Kotlin/kotlinx.coroutines\n- **Description**: Library support for Kotlin coroutines\n\n### Material Components for Android\n\n- **License**: Apache License 2.0\n- **URL**: https://github.com/material-components/material-components-android\n- **Description**: Modular and customizable Material Design UI components\n\n---\n\n## Summary\n\nAll third-party libraries used in NexaSDK are licensed under permissive open-source licenses that allow for commercial and non-commercial use, modification, and distribution. The primary licenses used are:\n\n- **MIT License**: Highly permissive, allows almost unrestricted use\n- **Apache License 2.0**: Permissive license with patent grant and attribution requirements\n- **BSD 3-Clause License**: Permissive license with attribution requirements\n" }, { "path": "README.md", "content": "
\n \"Nexa\n

\n 简体中文 \n |\n English \n

\n

🤝 Supported chipmakers

\n \n \n \n \n \n

\n

\n \n \"NexaSDK\n \n \n \"NexaAI/nexa-sdk\n \n

\n

\n \n \"Documentation\"\n \n \n \"Vote\n \n \"X\n \n \"Join\n \n \n \"Join\n \n

\n
\n\n# NexaSDK\n\n**NexaSDK lets you build the smartest and fastest on-device AI with minimum energy.** It is a highly performant local inference framework that runs the latest multimodal AI models locally on NPU, GPU, and CPU - across Android, Windows, Linux, macOS, and iOS devices with a few lines of code.\n\nNexaSDK supports latest models **weeks or months before anyone else** — Qwen3-VL, DeepSeek-OCR, Gemma3n (Vision), and more.\n\n> ⭐ **Star this repo** to keep up with exciting updates and new releases about latest on-device AI capabilities.\n\n## 🏆 Recognized Milestones\n\n- **Qualcomm** featured us **3 times** in official blogs.\n - [Innovating Multimodal AI on Qualcomm Hexagon NPU](https://www.qualcomm.com/developer/blog/2025/09/omnineural-4b-nexaml-qualcomm-hexagon-npu).\n - [First-ever Day-0 model support on Qualcomm Hexagon NPU for compute and mobile platforms, Auto and IoT](https://www.qualcomm.com/developer/blog/2025/10/granite-4-0-to-the-edge-on-device-ai-for-real-world-performance).\n - [A simple way to bring on-device AI to smartphones with Snapdragon](https://www.qualcomm.com/developer/blog/2025/11/nexa-ai-for-android-simple-way-to-bring-on-device-ai-to-smartphones-with-snapdragon)\n- **Qwen** featured us for [Day-0 Qwen3-VL support on NPU, GPU, and CPU](https://x.com/Alibaba_Qwen/status/1978154384098754943). We were 3 weeks ahead of Ollama and llama.cpp on GGUF support, and no one else supports it on NPU to date.\n- **IBM** featured our NexaML inference engine alongside vLLM, llama.cpp, and MLX in [official IBM blog](https://www.ibm.com/new/announcements/ibm-granite-4-0-hyper-efficient-high-performance-hybrid-models) and also for Day-0 Granite 4.0 support.\n- **Google** featured us for [EmbeddingGemma Day-0 NPU support](https://x.com/googleaidevs/status/1969188152049889511).\n- **AMD** featured us for [enabling SDXL-turbo image generation on AMD NPU](https://www.amd.com/en/developer/resources/technical-articles/2025/advancing-ai-with-nexa-ai--image-generation-on-amd-npu-with-sdxl.html).\n- **NVIDIA** featured Hyperlink, a viral local AI app powered by NexaSDK, in their [official blog](https://blogs.nvidia.com/blog/rtx-ai-garage-nexa-hyperlink-local-agent/).\n- **Microsoft** presented us on stage at Microsoft Ignite 2025 as [official partner](https://www.linkedin.com/posts/mixen_excited-to-celebrate-our-developer-partnerships-activity-7396601602327007232-AmCR?utm_source=share&utm_medium=member_desktop&rcm=ACoAAChXnS8B4gqbBLUlWfwt-ck0XAv472NzT4k).\n- **Intel** featured us for [Intel NPU support in NexaSDK](https://www.linkedin.com/posts/intel-software_ai-ondeviceai-nexasdk-activity-7376337062087667712-xw7i?utm_source=share&utm_medium=member_desktop&rcm=ACoAAChXnS8B4gqbBLUlWfwt-ck0XAv472NzT4k).\n\n## 🚀 Quick Start\n\n| Platform | Links |\n| --------------- | ----------------------------------------------------------------------------------------- |\n| 🖥️ CLI | [Quick Start](#-cli) ｜ [Docs](https://docs.nexa.ai/en/nexa-sdk-go/NexaCLI) |\n| 🐍 Python | [Quick Start](#-python-sdk) ｜ [Docs](https://docs.nexa.ai/en/nexa-sdk-python/overview) |\n| 🤖 Android | [Quick Start](#-android-sdk) ｜ [Docs](https://docs.nexa.ai/en/nexa-sdk-android/overview) |\n| 🐳 Linux Docker | [Quick Start](#-linux-docker) ｜ [Docs](https://docs.nexa.ai/en/nexa-sdk-docker/overview) |\n| 🍎 iOS | [Quick Start](#-ios-sdk) ｜ [Docs](https://docs.nexa.ai/en/nexa-sdk-ios/overview) |\n\n---\n\n### 🖥️ CLI\n\n**Download:**\n\n| Windows | macOS | Linux |\n| -------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------- | -------------------------------------------------------------------------------------------- |\n| [arm64 (Qualcomm NPU)](https://public-storage.nexa4ai.com/nexa_sdk/downloads/nexa-cli_windows_arm64.exe) | [arm64 (Apple Silicon)](https://public-storage.nexa4ai.com/nexa_sdk/downloads/nexa-cli_macos_arm64.pkg) | [arm64](https://github.com/NexaAI/nexa-sdk/releases/latest/download/nexa-cli_linux_arm64.sh) |\n| [x64 (Intel/AMD NPU)](https://public-storage.nexa4ai.com/nexa_sdk/downloads/nexa-cli_windows_x86_64.exe) | [x64](https://public-storage.nexa4ai.com/nexa_sdk/downloads/nexa-cli_macos_x86_64.pkg) | [x64](https://github.com/NexaAI/nexa-sdk/releases/latest/download/nexa-cli_linux_x86_64.sh) |\n\n**Run your first model:**\n\n```bash\n# Chat with Qwen3\nnexa infer ggml-org/Qwen3-1.7B-GGUF\n\n# Multimodal: drag images into the CLI\nnexa infer NexaAI/Qwen3-VL-4B-Instruct-GGUF\n\n# NPU (Windows arm64 with Snapdragon X Elite)\nnexa infer NexaAI/OmniNeural-4B\n```\n\n- **Models:** LLM, Multimodal, ASR, OCR, Rerank, Object Detection, Image Generation, Embedding\n- **Formats:** GGUF, MLX, NEXA\n- **NPU Models:** [Model Hub](https://sdk.nexa.ai/model)\n- 📖 [CLI Reference Docs](https://docs.nexa.ai/en/nexa-sdk-go/NexaCLI)\n\n---\n\n### 🐍 Python SDK\n\n```bash\npip install nexaai\n```\n\n```python\nfrom nexaai import LLM, GenerationConfig, ModelConfig, LlmChatMessage\n\nllm = LLM.from_(model=\"NexaAI/Qwen3-0.6B-GGUF\", config=ModelConfig())\n\nconversation = [\n LlmChatMessage(role=\"user\", content=\"Hello, tell me a joke\")\n]\nprompt = llm.apply_chat_template(conversation)\nfor token in llm.generate_stream(prompt, GenerationConfig(max_tokens=100)):\n print(token, end=\"\", flush=True)\n```\n\n- **Models:** LLM, Multimodal, ASR, OCR, Rerank, Object Detection, Image Generation, Embedding\n- **Formats:** GGUF, MLX, NEXA\n- **NPU Models:** [Model Hub](https://sdk.nexa.ai/model)\n- 📖 [Python SDK Docs](https://docs.nexa.ai/en/nexa-sdk-python/quickstart)\n\n---\n\n### 🤖 Android SDK\n\nAdd to your `app/AndroidManifest.xml`\n\n```xml\n\n```\n\nAdd to your `build.gradle.kts`:\n\n```kotlin\ndependencies {\n implementation(\"ai.nexa:core:0.0.19\")\n}\n```\n\n```kotlin\n// Initialize SDK\nNexaSdk.getInstance().init(this)\n\n// Load and run model\nVlmWrapper.builder()\n .vlmCreateInput(VlmCreateInput(\n model_name = \"omni-neural\",\n model_path = \"/data/data/your.app/files/models/OmniNeural-4B/files-1-1.nexa\",\n plugin_id = \"npu\",\n config = ModelConfig()\n ))\n .build()\n .onSuccess { vlm ->\n vlm.generateStreamFlow(\"Hello!\", GenerationConfig()).collect { print(it) }\n }\n```\n\n- **Requirements:** Android minSdk 27, Qualcomm Snapdragon 8 Gen 4 Chip\n- **Models:** LLM, Multimodal, ASR, OCR, Rerank, Embedding\n- **NPU Models:** [Supported Models](https://docs.nexa.ai/en/nexa-sdk-android/overview#supported-models)\n- 📖 [Android SDK Docs](https://docs.nexa.ai/en/nexa-sdk-android/quickstart)\n\n---\n\n### 🐳 Linux Docker\n\n```bash\ndocker pull nexa4ai/nexasdk:latest\n\nexport NEXA_TOKEN=\"your_token_here\"\ndocker run --rm -it --privileged \\\n -e NEXA_TOKEN \\\n nexa4ai/nexasdk:latest infer NexaAI/Granite-4.0-h-350M-NPU\n```\n\n- **Requirements:** Qualcomm Dragonwing IQ9, ARM64 systems\n- **Models:** LLM, VLM, ASR, CV, Rerank, Embedding\n- **NPU Models:** [Supported Models](https://docs.nexa.ai/en/nexa-sdk-docker/overview#supported-models)\n- 📖 [Linux Docker Docs](https://docs.nexa.ai/en/nexa-sdk-docker/quickstart)\n\n---\n\n### 🍎 iOS SDK\n\nDownload [NexaSdk.xcframework](https://nexa-model-hub-bucket.s3.us-west-1.amazonaws.com/public/ios/latest/NexaSdk.xcframework.zip) and add to your Xcode project.\n\n```swift\nimport NexaSdk\n\n// Example: Speech Recognition\nlet asr = try Asr(plugin: .ane)\ntry await asr.load(from: modelURL)\n\nlet result = try await asr.transcribe(options: .init(audioPath: \"audio.wav\"))\nprint(result.asrResult.transcript)\n```\n\n- **Requirements:** iOS 17.0+ / macOS 15.0+, Swift 5.9+\n- **Models:** LLM, ASR, OCR, Rerank, Embedding\n- **ANE Models:** [Apple Neural Engine Models](https://huggingface.co/collections/NexaAI/apple-neural-engine)\n- 📖 [iOS SDK Docs](https://docs.nexa.ai/en/nexa-sdk-ios/quickstart)\n\n## ⚙️ Features & Comparisons\n\n
\n\n| Features | **NexaSDK** | **Ollama** | **llama.cpp** | **LM Studio** |\n| ---------------------------------------- | ---------------------------------------------------------- | ---------- | ------------- | ------------- |\n| NPU support | ✅ NPU-first | ❌ | ❌ | ❌ |\n| Android/iOS SDK support | ✅ NPU/GPU/CPU support | ⚠️ | ⚠️ | ❌ |\n| Linux support (Docker image) | ✅ | ✅ | ✅ | ❌ |\n| Day-0 model support in GGUF, MLX, NEXA | ✅ | ❌ | ⚠️ | ❌ |\n| Full multimodality support | ✅ Image, Audio, Text, Embedding, Rerank, ASR, TTS | ⚠️ | ⚠️ | ⚠️ |\n| Cross-platform support | ✅ Desktop, Mobile (Android, iOS), Automotive, IoT (Linux) | ⚠️ | ⚠️ | ⚠️ |\n| One line of code to run | ✅ | ✅ | ⚠️ | ✅ |\n| OpenAI-compatible API + Function calling | ✅ | ✅ | ✅ | ✅ |\n\n

\n \n Legend:\n ✅ Supported   |  \n ⚠️ Partial or limited support   |  \n ❌ No\n \n

\n
\n\n## 🙏 Acknowledgements\n\nWe would like to thank the following projects:\n\n- [ggml](https://github.com/ggml-org/ggml)\n- [mlx-lm](https://github.com/ml-explore/mlx-lm)\n- [mlx-vlm](https://github.com/Blaizzy/mlx-vlm)\n- [mlx-audio](https://github.com/Blaizzy/mlx-audio)\n\n## 📄 License\n\nNexaSDK uses a dual licensing model:\n\n### CPU/GPU Components\n\nLicensed under [Apache License 2.0](LICENSE).\n\n### NPU Components\n\n- **Personal Use**: Free license key available from [Nexa AI Model Hub](https://sdk.nexa.ai/model). Each key activates 1 device for NPU usage.\n- **Commercial Use**: Contact [hello@nexa.ai](mailto:hello@nexa.ai) for licensing.\n\n## 🤝 Contact & Community Support\n\n### Business Inquiries\n\nFor model launching partner, business inquiries, or any other questions, please schedule a call with us [here](https://nexa.ai/book-a-call).\n\n### Community & Support\n\nWant more model support, backend support, device support or other features? We'd love to hear from you!\n\nFeel free to [submit an issue](https://github.com/NexaAI/nexa-sdk/issues) on our GitHub repository with your requests, suggestions, or feedback. Your input helps us prioritize what to build next.\n\nJoin our community:\n\n- [Discord](https://discord.gg/thRu2HaK4D)\n- [Slack](https://join.slack.com/t/nexaai/shared_invite/zt-30a8yfv8k-1JqAXv~OjKJKLqvbKqHJxA)\n- **[Nexa Wishlist](https://sdk.nexa.ai/wishlist)** — Request and vote for the models you want to run on-device.\n\n## 🏆 Nexa × Qualcomm On-Device Bounty Program\n\nRound 1: Build a working Android AI app that runs fully on-device on Qualcomm Hexagon NPU with NexaSDK.\n\nTimeline (PT): Jan 15 → Feb 15\nPrizes: $6,500 cash prize, Qualcomm official spotlight, flagship Snapdragon device, expert mentorship, and more\n\n👉 Join & details: [https://sdk.nexa.ai/bounty](https://sdk.nexa.ai/bounty)\n" }, { "path": "README_zh.md", "content": "
\n\t\"Nexa\n\t

\n\t\t 简体中文 \n\t\t|\n\t\t English \n\t

\n\t

🤝 支持的芯片厂商

\n\t\t\n\t\t\t\n\t\t\t\n\t\t\t\n\t\t\n\t

\n\t

\n\t\t\n\t\t\t\t\"NexaSDK\n\t\t\n\t\t\n\t\t\t\t\"NexaAI/nexa-sdk\n\t\t\n\t

\n\t

\n\t\t\n\t\t\t\t\"Documentation\"\n\t\t\n\t\t\n\t\t\t\t\"Vote\n\t\t\n\t\t\"X\n\t\t\n\t\t\t\t\"Join\n\t\t\n\t\t\n\t\t\t\t\"Join\n\t\t\n\t

\n
\n\n# NexaSDK\n\n**NexaSDK 让你用极低能耗打造最快、最聪明的本地 AI。** 它是一套高性能本地推理框架，只需几行代码即可在 Android、Windows、Linux、macOS 与 iOS 的 NPU、GPU、CPU 上运行最新的多模态 AI 模型。\n\nNexaSDK 往往能比其他人提前数周甚至数月支持最新模型 —— Qwen3-VL、DeepSeek-OCR、Gemma3n（视觉版）等。\n\n> ⭐ **star本仓库**，及时获取最新的本地 AI 能力更新与发布。\n\n## 🏆 重要里程碑\n\n- **Qualcomm** 官方博客中 3 次重点介绍我们。\n\t- [在 Qualcomm Hexagon NPU 上创新多模态 AI](https://www.qualcomm.com/developer/blog/2025/09/omnineural-4b-nexaml-qualcomm-hexagon-npu)。\n\t- [Qualcomm Hexagon NPU 在计算与移动平台（汽车与 IoT）上的首个 Day-0 模型支持](https://www.qualcomm.com/developer/blog/2025/10/granite-4-0-to-the-edge-on-device-ai-for-real-world-performance)。\n\t- [在 Snapdragon 手机上体验端侧 AI 的简单方式](https://www.qualcomm.com/developer/blog/2025/11/nexa-ai-for-android-simple-way-to-bring-on-device-ai-to-smartphones-with-snapdragon)\n- **Qwen** 为我们在 NPU、GPU、CPU 上 [Day-0 支持 Qwen3-VL](https://x.com/Alibaba_Qwen/status/1978154384098754943) 做了官方推荐。我们在 GGUF 支持上领先 Ollama 和 llama.cpp 3 周，并且目前只有我们支持 NPU。\n- **IBM** 在[官方博客](https://www.ibm.com/new/announcements/ibm-granite-4-0-hyper-efficient-high-performance-hybrid-models)中，将 NexaML 推理引擎与 vLLM、llama.cpp、MLX 并列介绍，感谢我们 Day-0 支持 Granite 4.0。\n- **Google** 认可我们对 EmbeddingGemma Day-0 的 NPU 支持（[官方致谢](https://x.com/googleaidevs/status/1969188152049889511)）。\n- **AMD** 在[官方博客](https://www.amd.com/en/developer/resources/technical-articles/2025/advancing-ai-with-nexa-ai--image-generation-on-amd-npu-with-sdxl.html)报道我们在 AMD NPU 上实现 SDXL-turbo 图像生成。\n- **NVIDIA** 在[官方博客](https://blogs.nvidia.com/blog/rtx-ai-garage-nexa-hyperlink-local-agent/)中介绍了由 NexaSDK 支撑的火爆本地 AI 应用 Hyperlink。\n- **Microsoft** 在 Microsoft Ignite 2025 上台展示了我们作为[官方合作伙伴](https://www.linkedin.com/posts/mixen_excited-to-celebrate-our-developer-partnerships-activity-7396601602327007232-AmCR?utm_source=share&utm_medium=member_desktop&rcm=ACoAAChXnS8B4gqbBLUlWfwt-ck0XAv472NzT4k)。\n- **Intel** 在[官方帖子](https://www.linkedin.com/posts/intel-software_ai-ondeviceai-nexasdk-activity-7376337062087667712-xw7i?utm_source=share&utm_medium=member_desktop&rcm=ACoAAChXnS8B4gqbBLUlWfwt-ck0XAv472NzT4k)中提到我们对 Intel NPU 的支持。\n\n## 🚀 快速开始\n\n| 平台 | 链接 |\n| --------------- | ---------------------------------------------------------------------------------------------- |\n| 🖥️ CLI | [快速开始](#-cli) ｜ [文档](https://docs.nexa.ai/en/nexa-sdk-go/NexaCLI) |\n| 🐍 Python | [快速开始](#-python-sdk) ｜ [文档](https://docs.nexa.ai/en/nexa-sdk-python/overview) |\n| 🤖 Android | [快速开始](#-android-sdk) ｜ [文档](https://docs.nexa.ai/en/nexa-sdk-android/overview) |\n| 🐳 Linux Docker | [快速开始](#-linux-docker) ｜ [文档](https://docs.nexa.ai/en/nexa-sdk-docker/overview) |\n| 🍎 iOS | [快速开始](#-ios-sdk) ｜ [文档](https://docs.nexa.ai/en/nexa-sdk-ios/overview) |\n\n---\n\n### 🖥️ CLI\n\n**下载：**\n\n| Windows | macOS | Linux |\n| -------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------- | -------------------------------------------------------------------------------------------- |\n| [arm64 (Qualcomm NPU)](https://public-storage.nexa4ai.com/nexa_sdk/downloads/nexa-cli_windows_arm64.exe) | [arm64 (Apple Silicon)](https://public-storage.nexa4ai.com/nexa_sdk/downloads/nexa-cli_macos_arm64.pkg) | [arm64](https://github.com/NexaAI/nexa-sdk/releases/latest/download/nexa-cli_linux_arm64.sh) |\n| [x64 (Intel/AMD NPU)](https://public-storage.nexa4ai.com/nexa_sdk/downloads/nexa-cli_windows_x86_64.exe) | [x64](https://public-storage.nexa4ai.com/nexa_sdk/downloads/nexa-cli_macos_x86_64.pkg) | [x64](https://github.com/NexaAI/nexa-sdk/releases/latest/download/nexa-cli_linux_x86_64.sh) |\n\n**运行你的第一个模型：**\n\n```bash\n# 与 Qwen3 对话\nnexa infer ggml-org/Qwen3-1.7B-GGUF\n\n# 多模态：在 CLI 中拖入图片\nnexa infer NexaAI/Qwen3-VL-4B-Instruct-GGUF\n\n# NPU（Windows arm64，Snapdragon X Elite）\nnexa infer NexaAI/OmniNeural-4B\n```\n\n- **模型类型：** LLM、多模态、ASR、OCR、Rerank、目标检测、图像生成、Embedding\n- **格式：** GGUF、MLX、NEXA\n- **NPU 模型：** [Model Hub](https://sdk.nexa.ai/model)\n- 📖 [CLI 参考文档](https://docs.nexa.ai/en/nexa-sdk-go/NexaCLI)\n\n---\n\n### 🐍 Python SDK\n\n```bash\npip install nexaai\n```\n\n```python\nfrom nexaai import LLM, GenerationConfig, ModelConfig, LlmChatMessage\n\nllm = LLM.from_(model=\"NexaAI/Qwen3-0.6B-GGUF\", config=ModelConfig())\n\nconversation = [\n\t\tLlmChatMessage(role=\"user\", content=\"Hello, tell me a joke\")\n]\nprompt = llm.apply_chat_template(conversation)\nfor token in llm.generate_stream(prompt, GenerationConfig(max_tokens=100)):\n\t\tprint(token, end=\"\", flush=True)\n```\n\n- **模型类型：** LLM、多模态、ASR、OCR、Rerank、目标检测、图像生成、Embedding\n- **格式：** GGUF、MLX、NEXA\n- **NPU 模型：** [Model Hub](https://sdk.nexa.ai/model)\n- 📖 [Python SDK 文档](https://docs.nexa.ai/en/nexa-sdk-python/quickstart)\n\n---\n\n### 🤖 Android SDK\n\n在 `app/AndroidManifest.xml` 中添加：\n\n```xml\n\n```\n\n在 `build.gradle.kts` 中添加：\n\n```kotlin\ndependencies {\n\t\timplementation(\"ai.nexa:core:0.0.19\")\n}\n```\n\n```kotlin\n// 初始化 SDK\nNexaSdk.getInstance().init(this)\n\n// 加载并运行模型\nVlmWrapper.builder()\n\t\t.vlmCreateInput(VlmCreateInput(\n\t\t\t\tmodel_name = \"omni-neural\",\n\t\t\t\tmodel_path = \"/data/data/your.app/files/models/OmniNeural-4B/files-1-1.nexa\",\n\t\t\t\tplugin_id = \"npu\",\n\t\t\t\tconfig = ModelConfig()\n\t\t))\n\t\t.build()\n\t\t.onSuccess { vlm ->\n\t\t\t\tvlm.generateStreamFlow(\"Hello!\", GenerationConfig()).collect { print(it) }\n\t\t}\n```\n\n- **要求：** Android minSdk 27，Qualcomm Snapdragon 8 Gen 4 芯片\n- **模型类型：** LLM、多模态、ASR、OCR、Rerank、Embedding\n- **NPU 模型：** [Supported Models](https://docs.nexa.ai/en/nexa-sdk-android/overview#supported-models)\n- 📖 [Android SDK 文档](https://docs.nexa.ai/en/nexa-sdk-android/quickstart)\n\n---\n\n### 🐳 Linux Docker\n\n```bash\ndocker pull nexa4ai/nexasdk:latest\n\nexport NEXA_TOKEN=\"your_token_here\"\ndocker run --rm -it --privileged \\\n\t-e NEXA_TOKEN \\\n\tnexa4ai/nexasdk:latest infer NexaAI/Granite-4.0-h-350M-NPU\n```\n\n- **要求：** Qualcomm Dragonwing IQ9，ARM64 系统\n- **模型类型：** LLM、VLM、ASR、CV、Rerank、Embedding\n- **NPU 模型：** [Supported Models](https://docs.nexa.ai/en/nexa-sdk-docker/overview#supported-models)\n- 📖 [Linux Docker 文档](https://docs.nexa.ai/en/nexa-sdk-docker/quickstart)\n\n---\n\n### 🍎 iOS SDK\n\n下载 [NexaSdk.xcframework](https://nexa-model-hub-bucket.s3.us-west-1.amazonaws.com/public/ios/latest/NexaSdk.xcframework.zip) 并添加到 Xcode 项目。\n\n```swift\nimport NexaSdk\n\n// 示例：语音识别\nlet asr = try Asr(plugin: .ane)\ntry await asr.load(from: modelURL)\n\nlet result = try await asr.transcribe(options: .init(audioPath: \"audio.wav\"))\nprint(result.asrResult.transcript)\n```\n\n- **要求：** iOS 17.0+ / macOS 15.0+，Swift 5.9+\n- **模型类型：** LLM、ASR、OCR、Rerank、Embedding\n- **ANE 模型：** [Apple Neural Engine Models](https://huggingface.co/collections/NexaAI/apple-neural-engine)\n- 📖 [iOS SDK 文档](https://docs.nexa.ai/en/nexa-sdk-ios/quickstart)\n\n## ⚙️ 功能与对比\n\n
\n\n| 功能 | **NexaSDK** | **Ollama** | **llama.cpp** | **LM Studio** |\n| -------------------------------------- | ---------------------------------------------------------- | ---------- | ------------- | ------------- |\n| NPU 支持 | ✅ NPU 优先 | ❌ | ❌ | ❌ |\n| Android/iOS SDK 支持 | ✅ NPU/GPU/CPU 支持 | ⚠️ | ⚠️ | ❌ |\n| Linux 支持（Docker 镜像） | ✅ | ✅ | ✅ | ❌ |\n| Day-0 支持 GGUF、MLX、NEXA | ✅ | ❌ | ⚠️ | ❌ |\n| 完整多模态支持 | ✅ 图像、音频、文本、Embedding、Rerank、ASR、TTS | ⚠️ | ⚠️ | ⚠️ |\n| 跨平台支持 | ✅ 桌面、移动（Android、iOS）、车载、IoT（Linux） | ⚠️ | ⚠️ | ⚠️ |\n| 一行代码即可运行 | ✅ | ✅ | ⚠️ | ✅ |\n| OpenAI 兼容 API + Function calling | ✅ | ✅ | ✅ | ✅ |\n\n

\n\t\n\t\t\t图例：\n\t\t\t✅ Supported   |  \n\t\t\t⚠️ Partial or limited support   |  \n\t\t\t❌ No\n\t\n

\n
\n\n## 🙏 致谢\n\n我们感谢以下项目：\n\n- [ggml](https://github.com/ggml-org/ggml)\n- [mlx-lm](https://github.com/ml-explore/mlx-lm)\n- [mlx-vlm](https://github.com/Blaizzy/mlx-vlm)\n- [mlx-audio](https://github.com/Blaizzy/mlx-audio)\n\n## 📄 许可证\n\nNexaSDK 采用双重许可模式：\n\n### CPU/GPU 组件\n\n基于 [Apache License 2.0](LICENSE)。\n\n### NPU 组件\n\n- **个人使用**：可从 [Nexa AI Model Hub](https://sdk.nexa.ai/model) 免费获取许可密钥。每个密钥激活 1 台设备的 NPU 使用。\n- **商业使用**：联系 [hello@nexa.ai](mailto:hello@nexa.ai) 获取授权。\n\n## 🤝 联系与社区支持\n\n### 商务合作\n\n如需模型发布合作、商务洽谈或其他问题，请在[此处](https://nexa.ai/book-a-call)安排会议。\n\n### 社区与支持\n\n想要更多模型支持、后端支持、设备支持或新功能？我们很乐意听到你的声音！\n\n欢迎在 GitHub [提交 issue](https://github.com/NexaAI/nexa-sdk/issues)，提出你的需求、建议或反馈。你的意见帮助我们确定优先级。\n\n加入社区：\n\n- [Discord](https://discord.gg/thRu2HaK4D)\n- [Slack](https://join.slack.com/t/nexaai/shared_invite/zt-30a8yfv8k-1JqAXv~OjKJKLqvbKqHJxA)\n- **[Nexa Wishlist](https://sdk.nexa.ai/wishlist)** —— 提交并为你想要运行在本地的模型投票。\n" }, { "path": "bindings/android/.gitignore", "content": "# Gradle files\n.gradle/\nbuild/\narm64-v8a/\n.DS_Store\n# Local configuration file (sdk path, etc)\nlocal.properties\n\n# Log/OS Files\n*.log\n\n# Android Studio generated files and folders\ncaptures/\n.externalNativeBuild/\n.cxx/\n*.aab\n*.apk\noutput-metadata.json\n\n# IntelliJ\n*.iml\n.idea/\nmisc.xml\ndeploymentTargetDropDown.xml\nrender.experimental.xml\n\n# Keystore files\n*.jks\n*.keystore\n\n# Google Services (e.g. APIs or Firebase)\ngoogle-services.json\n\n# Android Profiling\n*.hprof\napp/libs/app-release.aar\napp/libs/app-debug.aar\napp/release/\nrepo/" }, { "path": "bindings/android/README.md", "content": "# Nexa Android SDK Demo App\n\n[![Product Hunt](https://api.producthunt.com/widgets/embed-image/v1/top-post-badge.svg?post_id=1049998&theme=dark&period=daily&t=1765991451976)](https://www.producthunt.com/products/nexasdk-for-mobile)\n\n> 📣 **NexaSDK for Android** is highlighted by [Qualcomm blog](https://www.qualcomm.com/developer/blog/2025/11/nexa-ai-for-android-simple-way-to-bring-on-device-ai-to-smartphones-with-snapdragon) as \"a simple way to bring on-device AI to smartphones with Snapdragon\"\n\n## Overview\n\nThe Nexa AI Android SDK enables on-device AI inference for Android applications with NPU acceleration. Run Large Language Models (LLMs), Vision-Language Models (VLMs), Embeddings, Speech Recognition (ASR), Reranking, and Computer Vision models on Android devices with support for NPU, GPU, and CPU inference. \n\n📖 For full documentation, see [Android SDK Doc](https://docs.nexa.ai/en/nexa-sdk-android/overview).\n\n## Device Compatibility\n\n### Supported Hardware\n\n| Component | Requirement |\n|-----------|-------------|\n| **NPU** | Qualcomm Snapdragon 8 Gen 4 (optimized) |\n| **GPU** | Qualcomm Adreno GPU |\n| **CPU** | ARM64-v8a |\n| **RAM** | 4GB+ recommended |\n| **Storage** | 100MB - 4GB (varies by model) |\n\n### Minimum Requirements\n\n- Android API Level 27+ (Android 8.1 Oreo)\n- Architecture: ARM64-v8a\n\n## Quick Start with APK\n\n### Standard Demo App\n\nDownload and install the pre-built APK:\n\n```bash\n# Download: https://nexa-model-hub-bucket.s3.us-west-1.amazonaws.com/public/android-demo-release/nexaai-demo-app.apk\nadb install nexaai-demo-app.apk\n```\n\n### GPT-OSS NPU Demo\n\nFor running GPT-OSS model on Qualcomm NPU:\n\n```bash\n# Download: https://nexa-model-hub-bucket.s3.us-west-1.amazonaws.com/public/nexa_sdk/huggingface-models/gpt-oss-android-demo/nexaai-gpt-oss-npu.apk\nadb install nexaai-gpt-oss-npu.apk\n```\n\n## Build from Source\n\n> 🎬 Watch the [tutorial video](https://www.linkedin.com/feed/update/urn:li:activity:7394055404562935809) showing how to run the sample app in 40 seconds.\n\n### Step-by-Step Instructions\n\n1. **Clone the repository**\n\n ```bash\n git clone https://github.com/NexaAI/nexa-sdk/\n ```\n\n2. **Open in Android Studio**\n \n Open the `bindings/android` folder in Android Studio.\n\n3. **Download a model**\n \n Follow the [Android SDK Doc](https://docs.nexa.ai/en/nexa-sdk-android/overview) to download a model. Below are some examples to download:\n - [Granite-4.0-h-350M-NPU](https://huggingface.co/NexaAI/Granite-4.0-h-350M-NPU-mobile)\n - [Granite-4-Micro-NPU](https://huggingface.co/NexaAI/Granite-4-Micro-NPU-mobile)\n - [OmniNeural-4B](https://huggingface.co/NexaAI/OmniNeural-4B-mobile)\n \n Place the model in the app's data directory:\n ```\n /data/data/com.nexa.demo/files/models/\n ```\n\n4. **Build and run** the app in Android Studio\n\n## Tutorial: LFM2-24B-A2B-Preview-GGUF\n\nThis walkthrough uses the [LFM2-24B-A2B-Preview-GGUF](https://huggingface.co/NexaAI/LFM2-24B-A2B-Preview-GGUF) model in the demo app.\n\n1. **Install the app** \n Install the demo app (APK or build from source as above).\n\n2. **Select the model** \n Open the model selector (dropdown next to the model name) and choose **LFM2-24B-A2B-Preview-GGUF**.\n\n \"LFM2-24B-A2B-Preview-GGUF\n\n3. **Download** \n Tap **Download** to fetch the model to your device. Wait until the download finishes.\n\n4. **Load** \n Tap **Load**. A **load model config** dialog appears: choose **CPU**, **GPU**, or **NPU** (for Qualcomm NPU), then tap **SURE**. Once the model is loaded, the chat area becomes available.\n\n \"load\n\n5. **Chat** \n Type your message in the input field at the bottom, then tap **Send** to get a response. Use **Clear** to clear the input or conversation as needed.\n\n \"Chat\n" }, { "path": "bindings/android/README_zh.md", "content": "# Nexa Android SDK 演示 App\n\n[![Product Hunt](https://api.producthunt.com/widgets/embed-image/v1/top-post-badge.svg?post_id=1049998&theme=dark&period=daily&t=1765991451976)](https://www.producthunt.com/products/nexasdk-for-mobile)\n\n> 📣 **NexaSDK for Android** 被 [Qualcomm 博客](https://www.qualcomm.com/developer/blog/2025/11/nexa-ai-for-android-simple-way-to-bring-on-device-ai-to-smartphones-with-snapdragon) 评价为\"将端侧 AI 引入 Snapdragon 智能手机的简易方案\"\n\n## 概述\n\nNexa AI Android SDK 让 Android 应用可以在设备端使用 NPU 加速进行 AI 推理。支持在 Android 设备上运行大型语言模型（LLMs）、视觉语言模型（VLMs）、嵌入模型、语音识别（ASR）、重排序和计算机视觉模型，并可在 NPU、GPU 或 CPU 上进行推理。\n\n📖 完整文档请参考 [Android SDK 文档](https://docs.nexa.ai/cn/nexa-sdk-android/overview)。\n\n## 设备兼容性\n\n### 支持的硬件\n\n| 组件 | 要求 |\n|-----------|-------------|\n| **NPU** | 高通骁龙 8 Gen 4（已优化） |\n| **GPU** | 高通 Adreno GPU |\n| **CPU** | ARM64-v8a |\n| **RAM** | 推荐 4GB 及以上 |\n| **存储** | 100MB - 4GB（具体取决于模型） |\n\n### 最低要求\n\n- Android API 等级 27 及以上（Android 8.1 Oreo）\n- 架构：ARM64-v8a\n\n## 快速开始（APK 安装）\n\n### 标准演示 App\n\n下载并安装预编译的 APK：\n\n```bash\n# 下载地址: https://nexa-model-hub-bucket.s3.us-west-1.amazonaws.com/public/android-demo-release/nexaai-demo-app.apk\nadb install nexaai-demo-app.apk\n```\n\n### GPT-OSS NPU 演示\n\n在高通 NPU 上运行 GPT-OSS 模型：\n\n```bash\n# 下载地址: https://nexa-model-hub-bucket.s3.us-west-1.amazonaws.com/public/nexa_sdk/huggingface-models/gpt-oss-android-demo/nexaai-gpt-oss-npu.apk\nadb install nexaai-gpt-oss-npu.apk\n```\n\n## 从源码编译\n\n> 🎬 观看 [教程视频](https://www.linkedin.com/feed/update/urn:li:activity:7394055404562935809)，只需 40 秒即可体验如何运行示例 App。\n\n### 详细步骤\n\n1. **克隆仓库**\n\n ```bash\n git clone https://github.com/NexaAI/nexa-sdk/\n ```\n\n2. **用 Android Studio 打开**\n \n 用 Android Studio 打开 `bindings/android` 文件夹。\n\n3. **下载模型**\n \n 按照 [Android SDK 文档](https://docs.nexa.ai/cn/nexa-sdk-android/overview) 的说明下载模型。以下是一些可下载的示例：\n - [Granite-4.0-h-350M-NPU](https://huggingface.co/NexaAI/Granite-4.0-h-350M-NPU-mobile)\n - [Granite-4-Micro-NPU](https://huggingface.co/NexaAI/Granite-4-Micro-NPU-mobile)\n - [OmniNeural-4B](https://huggingface.co/NexaAI/OmniNeural-4B-mobile)\n \n 将模型放到 App 的数据目录：\n ```\n /data/data/com.nexa.demo/files/models/\n ```\n\n4. **编译并运行** 在 Android Studio 中编译并运行应用\n" }, { "path": "bindings/android/app/.gitignore", "content": "/build\n/src/main/assets/model_list_backup.json\n/src/main/assets/model_list_local.json" }, { "path": "bindings/android/app/build.gradle.kts", "content": "// Copyright 2024-2026 Nexa AI, Inc.\n//\n// Licensed under the Apache License, Version 2.0 (the \"License\");\n// you may not use this file except in compliance with the License.\n// You may obtain a copy of the License at\n//\n// http://www.apache.org/licenses/LICENSE-2.0\n//\n// Unless required by applicable law or agreed to in writing, software\n// distributed under the License is distributed on an \"AS IS\" BASIS,\n// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n// See the License for the specific language governing permissions and\n// limitations under the License.\n\nplugins {\n alias(libs.plugins.android.application)\n alias(libs.plugins.kotlin.android)\n alias(libs.plugins.kotlin.compose)\n id(\"org.jetbrains.kotlin.plugin.serialization\") version \"1.9.23\"\n}\n\nandroid {\n namespace = \"com.nexa.demo\"\n compileSdk = 36\n\n signingConfigs {\n create(\"release\") {\n // Note: For production builds, use environment variables or local.properties\n // Example: storePassword = System.getenv(\"KEYSTORE_PASSWORD\") ?: \"\"\n storeFile = file(\"test\")\n storePassword = project.findProperty(\"KEYSTORE_PASSWORD\")?.toString() ?: \"123456\"\n keyAlias = \"test\"\n keyPassword = project.findProperty(\"KEY_PASSWORD\")?.toString() ?: \"123456\"\n }\n }\n\n defaultConfig {\n applicationId = \"com.nexa.demo\"\n minSdk = 27\n targetSdk = 36\n versionCode = 1\n versionName = \"1.0\"\n\n testInstrumentationRunner = \"androidx.test.runner.AndroidJUnitRunner\"\n }\n\n buildTypes {\n release {\n isMinifyEnabled = false\n proguardFiles(\n getDefaultProguardFile(\"proguard-android-optimize.txt\"),\n \"proguard-rules.pro\"\n )\n signingConfig = signingConfigs.getByName(\"release\")\n }\n\n debug {\n isMinifyEnabled = false\n proguardFiles(\n getDefaultProguardFile(\"proguard-android-optimize.txt\"),\n \"proguard-rules.pro\"\n )\n signingConfig = signingConfigs.getByName(\"release\")\n }\n }\n compileOptions {\n sourceCompatibility = JavaVersion.VERSION_11\n targetCompatibility = JavaVersion.VERSION_11\n }\n kotlinOptions {\n jvmTarget = \"11\"\n }\n// sourceSets {\n// getByName(\"main\") {\n// jniLibs.srcDirs(\"src/main/jniLibs\")\n// }\n// }\n packagingOptions {\n jniLibs.useLegacyPackaging = true\n }\n\n buildFeatures {\n viewBinding = true\n dataBinding = true\n compose = true\n buildConfig = true\n }\n}\n\nval bridgePathExist = gradle.extra[\"bridgePathExist\"] as Boolean\nprint(\"bridgePathExist: $bridgePathExist\\n\")\n\ndependencies {\n\n // ===== NEXA CLOUD SDK =====\n // Using cloud SDK instead of local bridge - latest version\n implementation(\"ai.nexa:core:+\")\n // ===== NEXA CLOUD SDK END =====\n implementation(project(\":transform\"))\n implementation(\":okdownload-core@aar\")\n implementation(\":okdownload-sqlite@aar\")\n implementation(\":okdownload-okhttp@aar\")\n implementation(\":okdownload-ktx@aar\")\n implementation(kotlin(\"reflect\"))\n implementation(libs.glide)\n implementation(libs.gson)\n implementation(libs.markwon.core)\n implementation(libs.markwon.strikethrough)\n implementation(libs.markwon.tables)\n implementation(libs.markwon.linkify)\n implementation(libs.markwon.latex)\n implementation(libs.markwon.inline.parser)\n implementation(libs.recyclerview)\n implementation(libs.toaster)\n implementation(libs.material)\n implementation(libs.imm.bar)\n implementation(libs.imm.bar.ktx)\n implementation(libs.auto.size)\n implementation(libs.okhttp)\n implementation(libs.kotlinx.serialization.json)\n implementation(libs.androidx.core.ktx)\n implementation(libs.androidx.lifecycle.runtime.ktx)\n implementation(libs.androidx.activity.compose)\n implementation(platform(libs.androidx.compose.bom))\n implementation(libs.androidx.ui)\n implementation(libs.androidx.ui.graphics)\n implementation(libs.androidx.ui.tooling.preview)\n implementation(libs.androidx.material3)\n implementation(libs.androidx.appcompat)\n implementation(libs.androidx.activity)\n implementation(libs.androidx.constraintlayout)\n testImplementation(libs.junit)\n androidTestImplementation(libs.androidx.junit)\n androidTestImplementation(libs.androidx.espresso.core)\n androidTestImplementation(platform(libs.androidx.compose.bom))\n androidTestImplementation(libs.androidx.ui.test.junit4)\n debugImplementation(libs.androidx.ui.tooling)\n debugImplementation(libs.androidx.ui.test.manifest)\n}" }, { "path": "bindings/android/app/proguard-rules.pro", "content": "# Copyright 2024-2026 Nexa AI, Inc.\n#\n# Licensed under the Apache License, Version 2.0 (the \"License\");\n# you may not use this file except in compliance with the License.\n# You may obtain a copy of the License at\n#\n# http://www.apache.org/licenses/LICENSE-2.0\n#\n# Unless required by applicable law or agreed to in writing, software\n# distributed under the License is distributed on an \"AS IS\" BASIS,\n# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n# See the License for the specific language governing permissions and\n# limitations under the License.\n\n# Add project specific ProGuard rules here.\n# You can control the set of applied configuration files using the\n# proguardFiles setting in build.gradle.\n#\n# For more details, see\n# http://developer.android.com/guide/developing/tools/proguard.html\n\n# If your project uses WebView with JS, uncomment the following\n# and specify the fully qualified class name to the JavaScript interface\n# class:\n#-keepclassmembers class fqcn.of.javascript.interface.for.webview {\n# public *;\n#}\n\n# Uncomment this to preserve the line number information for\n# debugging stack traces.\n#-keepattributes SourceFile,LineNumberTable\n\n# If you keep the line number information, uncomment this to\n# hide the original source file name.\n#-renamesourcefileattribute SourceFile" }, { "path": "bindings/android/app/src/androidTest/java/com/nexa/demo/ExampleInstrumentedTest.kt", "content": "// Copyright 2024-2026 Nexa AI, Inc.\n//\n// Licensed under the Apache License, Version 2.0 (the \"License\");\n// you may not use this file except in compliance with the License.\n// You may obtain a copy of the License at\n//\n// http://www.apache.org/licenses/LICENSE-2.0\n//\n// Unless required by applicable law or agreed to in writing, software\n// distributed under the License is distributed on an \"AS IS\" BASIS,\n// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n// See the License for the specific language governing permissions and\n// limitations under the License.\n\npackage com.nexa.demo\n\nimport androidx.test.platform.app.InstrumentationRegistry\nimport androidx.test.ext.junit.runners.AndroidJUnit4\n\nimport org.junit.Test\nimport org.junit.runner.RunWith\n\nimport org.junit.Assert.*\n\n/**\n * Instrumented test, which will execute on an Android device.\n *\n * See [testing documentation](http://d.android.com/tools/testing).\n */\n@RunWith(AndroidJUnit4::class)\nclass ExampleInstrumentedTest {\n @Test\n fun useAppContext() {\n // Context of the app under test.\n val appContext = InstrumentationRegistry.getInstrumentation().targetContext\n assertEquals(\"com.nexa.demo\", appContext.packageName)\n }\n}" }, { "path": "bindings/android/app/src/main/AndroidManifest.xml", "content": "\n\n\n\n \n \n \n \n \n \n\n \n \n\n \n \n\n \n \n \n\n \n \n \n\n \n \n \n \n\n" }, { "path": "bindings/android/app/src/main/assets/model_list.json", "content": "[\n {\n \"id\": \"Llama3.2-3B-NPU-Turbo-NPU\",\n \"displayName\": \"Llama3.2-3B-NPU-Turbo-NPU\",\n \"modelName\": \"files-1-2.nexa\",\n \"versionCode\": 1,\n \"baseUrl\": \"https://nexa-model-hub-bucket.s3.us-west-1.amazonaws.com/public/nexa_sdk/huggingface-models/Llama3.2-3B-NPU-Turbo-NPU-mobile/\",\n \"modelUrl\": \"files-1-2.nexa\"\n },\n {\n \"id\": \"Llama-3.2-3B-Instruct-GGUF\",\n \"displayName\": \"Llama-3.2-3B-Instruct-GGUF\",\n \"modelName\": \"Llama-3.2-3B-Instruct-Q4_0.gguf\",\n \"type\": \"chat\",\n \"versionCode\": 1,\n \"pluginIds\": 17,\n \"modelUrl\": \"https://huggingface.co/bartowski/Llama-3.2-3B-Instruct-GGUF/resolve/main/Llama-3.2-3B-Instruct-Q4_0.gguf\"\n },\n {\n \"id\": \"Qwen3-4B-Instruct-2507-npu\",\n \"displayName\": \"Qwen3-4B-Instruct-2507-npu\",\n \"modelName\": \"files-1-1.nexa\",\n \"versionCode\": 1,\n \"baseUrl\": \"https://nexa-model-hub-bucket.s3.us-west-1.amazonaws.com/public/nexa_sdk/huggingface-models/Qwen3-4B-Instruct-2507-npu-mobile/\",\n \"modelUrl\": \"files-1-1.nexa\"\n },\n {\n \"id\": \"gpt-oss-20b-GGUF\",\n \"displayName\": \"gpt-oss-20b-GGUF\",\n \"modelName\": \"gpt-oss-20b-Q4_0.gguf\",\n \"type\": \"chat\",\n \"versionCode\": 1,\n \"pluginIds\": 17,\n \"modelUrl\": \"https://huggingface.co/unsloth/gpt-oss-20b-GGUF/resolve/main/gpt-oss-20b-Q4_0.gguf\"\n },\n {\n \"id\": \"Qwen3-4B-GGUF\",\n \"displayName\": \"Qwen3-4B-GGUF\",\n \"modelName\": \"Qwen3-4B-Q4_0.gguf\",\n \"type\": \"chat\",\n \"versionCode\": 1,\n \"pluginIds\": 17,\n \"modelUrl\": \"https://huggingface.co/unsloth/Qwen3-4B-GGUF/resolve/main/Qwen3-4B-Q4_0.gguf\"\n },\n {\n \"id\": \"Granite-4-Micro-NPU\",\n \"displayName\": \"Granite-4.0-Micro 3B\",\n \"modelName\": \"files-1-2.nexa\",\n \"versionCode\": 1,\n \"baseUrl\": \"https://nexa-model-hub-bucket.s3.us-west-1.amazonaws.com/public/nexa_sdk/huggingface-models/Granite-4-Micro-NPU-mobile/\",\n \"modelUrl\": \"files-1-2.nexa\"\n },\n {\n \"id\": \"LFM2-1.2B-npu\",\n \"displayName\": \"LFM2-1.2B-npu\",\n \"modelName\": \"files-1-2.nexa\",\n \"versionCode\": 1,\n \"baseUrl\": \"https://nexa-model-hub-bucket.s3.us-west-1.amazonaws.com/public/nexa_sdk/huggingface-models/LFM2-1.2B-npu-mobile/\",\n \"modelUrl\": \"files-1-2.nexa\"\n },\n {\n \"id\": \"LFM2.5-1.2B-mobile\",\n \"displayName\": \"LFM2.5-1.2B-npu\",\n \"modelName\": \"files-1-2.nexa\",\n \"versionCode\": 1,\n \"baseUrl\": \"https://nexa-model-hub-bucket.s3.us-west-1.amazonaws.com/public/nexa_sdk/huggingface-models/LFM2.5-1.2B-mobile/\",\n \"modelUrl\": \"files-1-2.nexa\"\n },\n {\n \"id\": \"OmniNeural-4B\",\n \"displayName\": \"OmniNeural-4B\",\n \"modelName\": \"files-1-1.nexa\",\n \"versionCode\": 1,\n \"baseUrl\": \"https://nexa-model-hub-bucket.s3.us-west-1.amazonaws.com/public/nexa_sdk/huggingface-models/OmniNeural-4B-mobile/\",\n \"modelUrl\": \"files-1-1.nexa\"\n },\n {\n \"id\": \"paddleocr-npu\",\n \"displayName\": \"paddleocr-npu\",\n \"modelName\": \"weights-1-1.nexa\",\n \"versionCode\": 1,\n \"baseUrl\": \"https://nexa-model-hub-bucket.s3.us-west-1.amazonaws.com/public/nexa_sdk/huggingface-models/paddleocr-npu-mobile/\",\n \"modelUrl\": \"weights-1-1.nexa\"\n },\n {\n \"id\": \"yolo26x-npu\",\n \"displayName\": \"yolo26x-npu\",\n \"modelName\": \"weights-1-1.nexa\",\n \"versionCode\": 1,\n \"baseUrl\": \"https://nexa-model-hub-bucket.s3.us-west-1.amazonaws.com/public/nexa_sdk/huggingface-models/yolo26x-npu-mobile/\",\n \"modelUrl\": \"weights-1-1.nexa\"\n },\n {\n \"id\": \"depth-anything-v2-npu\",\n \"displayName\": \"depth-anything-v2-npu\",\n \"modelName\": \"weights-1-1.nexa\",\n \"versionCode\": 1,\n \"baseUrl\": \"https://nexa-model-hub-bucket.s3.us-west-1.amazonaws.com/public/nexa_sdk/huggingface-models/depth-anything-v2-npu-mobile/\",\n \"modelUrl\": \"weights-1-1.nexa\"\n },\n {\n \"id\": \"embeddinggemma-300m-npu\",\n \"displayName\": \"embeddinggemma-300m-npu\",\n \"modelName\": \"files-1-2.nexa\",\n \"versionCode\": 1,\n \"baseUrl\": \"https://nexa-model-hub-bucket.s3.us-west-1.amazonaws.com/public/nexa_sdk/huggingface-models/embeddinggemma-300m-npu-mobile/\",\n \"modelUrl\": \"files-1-2.nexa\"\n },\n {\n \"id\": \"LFM2-1.2B-GGUF-GGUF\",\n \"displayName\": \"LFM2-1.2B-GGUF-GGUF\",\n \"modelName\": \"LFM2-1.2B-Q4_0.gguf\",\n \"type\": \"chat\",\n \"versionCode\": 1,\n \"pluginIds\": 17,\n \"modelUrl\": \"https://huggingface.co/LiquidAI/LFM2-1.2B-GGUF/resolve/main/LFM2-1.2B-Q4_0.gguf\"\n },\n {\n \"id\": \"Llama-3.2-1B-Instruct-GGUF\",\n \"displayName\": \"Llama-3.2-1B-Instruct-GGUF\",\n \"modelName\": \"Llama-3.2-1B-Instruct-Q4_0.gguf\",\n \"type\": \"chat\",\n \"versionCode\": 1,\n \"pluginIds\": 17,\n \"modelUrl\": \"https://huggingface.co/bartowski/Llama-3.2-1B-Instruct-GGUF/resolve/main/Llama-3.2-1B-Instruct-Q4_0.gguf\"\n },\n {\n \"id\": \"LFM2-24B-A2B-Preview-GGUF\",\n \"displayName\": \"LFM2-24B-A2B-Preview-GGUF\",\n \"modelName\": \"LFM2-24B-A2B-Preview-Q4_0.gguf\",\n \"type\": \"chat\",\n \"versionCode\": 1,\n \"pluginIds\": 17,\n \"modelUrl\": \"https://huggingface.co/NexaAI/LFM2-24B-A2B-Preview-GGUF/resolve/main/LFM2-24B-A2B-Preview-Q4_0.gguf\"\n }\n]\n" }, { "path": "bindings/android/app/src/main/java/com/nexa/demo/ChatAdapter.kt", "content": "// Copyright 2024-2026 Nexa AI, Inc.\n//\n// Licensed under the Apache License, Version 2.0 (the \"License\");\n// you may not use this file except in compliance with the License.\n// You may obtain a copy of the License at\n//\n// http://www.apache.org/licenses/LICENSE-2.0\n//\n// Unless required by applicable law or agreed to in writing, software\n// distributed under the License is distributed on an \"AS IS\" BASIS,\n// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n// See the License for the specific language governing permissions and\n// limitations under the License.\n\npackage com.nexa.demo\n\nimport android.content.Intent\nimport android.graphics.BitmapFactory\nimport android.net.Uri\nimport android.text.method.LinkMovementMethod\nimport android.view.LayoutInflater\nimport android.view.View\nimport android.view.ViewGroup\nimport android.widget.ImageView\nimport android.widget.LinearLayout\nimport android.widget.TextView\nimport androidx.recyclerview.widget.RecyclerView\nimport io.noties.markwon.Markwon\nimport io.noties.markwon.ext.latex.JLatexMathPlugin\nimport io.noties.markwon.ext.strikethrough.StrikethroughPlugin\nimport io.noties.markwon.ext.tables.TablePlugin\nimport io.noties.markwon.inlineparser.MarkwonInlineParserPlugin\nimport io.noties.markwon.linkify.LinkifyPlugin\nimport java.io.File\n\n\ndata class Message(\n val content: String,\n val type: MessageType,\n val images: List = emptyList(),\n val audio: List = emptyList()\n)\n\nenum class MessageType(val value: Int) {\n USER(0),\n ASSISTANT(1),\n PROFILE(2),\n IMAGES(3),\n ASSISTANT_IMAGES(4);\n\n companion object {\n fun from(value: Int): MessageType =\n entries.firstOrNull { it.value == value } ?: PROFILE\n }\n}\n\nclass ChatAdapter(private val messages: List) :\n RecyclerView.Adapter() {\n\n override fun getItemViewType(position: Int): Int {\n val message = messages[position]\n return message.type.value\n }\n\n override fun onCreateViewHolder(parent: ViewGroup, viewType: Int): RecyclerView.ViewHolder {\n val inflater = LayoutInflater.from(parent.context)\n val type = MessageType.from(viewType)\n return when (type) {\n MessageType.USER -> UserViewHolder(inflater.inflate(R.layout.item_user_message, parent, false))\n MessageType.ASSISTANT -> AiViewHolder(inflater.inflate(R.layout.item_ai_message, parent, false))\n MessageType.IMAGES -> ImagesViewHolder(inflater.inflate(R.layout.item_image_message, parent, false))\n MessageType.ASSISTANT_IMAGES -> ImagesViewHolder(\n inflater.inflate(\n R.layout.item_assistant_image_message,\n parent,\n false\n )\n )\n\n else -> ProfileViewHolder(inflater.inflate(R.layout.item_profile_message, parent, false))\n }\n }\n\n override fun onBindViewHolder(holder: RecyclerView.ViewHolder, position: Int) {\n val message = messages[position]\n if (holder is UserViewHolder) holder.bind(message)\n if (holder is AiViewHolder) holder.bind(message)\n if (holder is ImagesViewHolder) holder.bind(message)\n if (holder is ProfileViewHolder) holder.bind(message)\n }\n\n override fun getItemCount() = messages.size\n\n class UserViewHolder(itemView: View) : RecyclerView.ViewHolder(itemView) {\n private val tvMessage: TextView = itemView.findViewById(R.id.tv_message)\n fun bind(message: Message) {\n tvMessage.text = message.content\n }\n }\n\n class AiViewHolder(itemView: View) : RecyclerView.ViewHolder(itemView) {\n private val tvMessage: TextView = itemView.findViewById(R.id.tv_message)\n private val markwon: Markwon = Markwon.builder(itemView.context)\n .usePlugin(StrikethroughPlugin.create())\n .usePlugin(TablePlugin.create(itemView.context))\n .usePlugin(LinkifyPlugin.create())\n .usePlugin(MarkwonInlineParserPlugin.create())\n .usePlugin(JLatexMathPlugin.create(tvMessage.textSize) { builder ->\n builder.inlinesEnabled(true)\n builder.blocksEnabled(true)\n })\n .build()\n\n fun bind(message: Message) {\n markwon.setMarkdown(tvMessage, message.content.trim())\n tvMessage.movementMethod = LinkMovementMethod.getInstance()\n }\n }\n\n class ProfileViewHolder(itemView: View) : RecyclerView.ViewHolder(itemView) {\n private val tvMessage: TextView = itemView.findViewById(R.id.tv_message)\n\n fun bind(message: Message) {\n tvMessage.text = message.content\n }\n\n private fun dpToPx(dp: Int, context: android.content.Context): Int {\n return (dp * context.resources.displayMetrics.density).toInt()\n }\n }\n\n class ImagesViewHolder(itemView: View) : RecyclerView.ViewHolder(itemView) {\n private val imageContainer: LinearLayout = itemView.findViewById(R.id.image_container)\n fun bind(message: Message) {\n val savedImageFiles = message.images\n imageContainer.removeAllViews()\n val context = itemView.context\n\n for (file in savedImageFiles) {\n val itemView = LayoutInflater.from(context)\n .inflate(R.layout.item_image_item_message, imageContainer, false)\n val ivImage = itemView.findViewById(R.id.iv_image)\n val bitmap = BitmapFactory.decodeFile(file.absolutePath)\n if (bitmap != null) {\n ivImage.setImageBitmap(bitmap)\n }\n imageContainer.addView(itemView)\n }\n }\n }\n}\n" }, { "path": "bindings/android/app/src/main/java/com/nexa/demo/FileConfig.kt", "content": "// Copyright 2024-2026 Nexa AI, Inc.\n//\n// Licensed under the Apache License, Version 2.0 (the \"License\");\n// you may not use this file except in compliance with the License.\n// You may obtain a copy of the License at\n//\n// http://www.apache.org/licenses/LICENSE-2.0\n//\n// Unless required by applicable law or agreed to in writing, software\n// distributed under the License is distributed on an \"AS IS\" BASIS,\n// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n// See the License for the specific language governing permissions and\n// limitations under the License.\n\npackage com.nexa.demo\n\nimport android.content.Context\nimport java.io.File\n\nclass FileConfig {\n companion object {\n val modelsDir = fun(context: Context): File {\n return File(context.filesDir, \"models\").apply { if (!exists()) mkdirs() }\n }\n }\n}" }, { "path": "bindings/android/app/src/main/java/com/nexa/demo/GenerationConfigSample.kt", "content": "// Copyright 2024-2026 Nexa AI, Inc.\n//\n// Licensed under the Apache License, Version 2.0 (the \"License\");\n// you may not use this file except in compliance with the License.\n// You may obtain a copy of the License at\n//\n// http://www.apache.org/licenses/LICENSE-2.0\n//\n// Unless required by applicable law or agreed to in writing, software\n// distributed under the License is distributed on an \"AS IS\" BASIS,\n// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n// See the License for the specific language governing permissions and\n// limitations under the License.\n\npackage com.nexa.demo\n\nimport com.nexa.sdk.bean.GenerationConfig\nimport com.nexa.sdk.bean.SamplerConfig\n\n// Configuration sample for generation with defaults compatible with bridge\n// maxTokens: 0 = no limit, generates until model's natural stopping point\ndata class GenerationConfigSample(\n var maxTokens: Int = 2048,\n var stopWords: List? = null,\n var stopCount: Int = 0,\n var nPast: Int = 0,\n var imagePaths: List? = null,\n var imageCount: Int = 0,\n var audioPaths: List? = null,\n var audioCount: Int = 0,\n var systemPrompt: String = \"\"\n) {\n // Convert to GenerationConfig with minimal sampler setup for bridge compatibility\n // Sampler config uses bridge defaults (no custom parameters applied)\n fun toGenerationConfig(grammarString:String? = null): GenerationConfig {\n return GenerationConfig(\n maxTokens = this.maxTokens,\n stopWords = this.stopWords?.toTypedArray(),\n stopCount = this.stopCount,\n nPast = this.nPast,\n //samplerConfig = SamplerConfig(\n // grammarString = grammarString\n // All other sampler parameters use bridge defaults\n // No temperature, topK, topP, penalties applied\n //),\n imagePaths = this.imagePaths?.toTypedArray(),\n imageCount = this.imageCount,\n audioPaths = this.audioPaths?.toTypedArray(),\n audioCount = this.audioCount\n )\n }\n}" }, { "path": "bindings/android/app/src/main/java/com/nexa/demo/MainActivity.kt", "content": "// Copyright 2024-2026 Nexa AI, Inc.\n//\n// Licensed under the Apache License, Version 2.0 (the \"License\");\n// you may not use this file except in compliance with the License.\n// You may obtain a copy of the License at\n//\n// http://www.apache.org/licenses/LICENSE-2.0\n//\n// Unless required by applicable law or agreed to in writing, software\n// distributed under the License is distributed on an \"AS IS\" BASIS,\n// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n// See the License for the specific language governing permissions and\n// limitations under the License.\n\npackage com.nexa.demo\n\nimport android.Manifest\nimport android.app.Activity\nimport android.content.Context\nimport android.content.DialogInterface\nimport android.content.Intent\nimport android.content.SharedPreferences\nimport android.content.pm.PackageManager\nimport android.graphics.Bitmap\nimport android.graphics.BitmapFactory\nimport android.graphics.Canvas\nimport android.graphics.Color\nimport android.graphics.Paint\nimport android.graphics.RectF\nimport android.net.Uri\nimport android.os.Bundle\nimport android.os.Environment\nimport android.provider.MediaStore\nimport android.system.Os\nimport android.util.Log\nimport android.view.LayoutInflater\nimport android.view.View\nimport android.view.inputmethod.InputMethodManager\nimport android.widget.AdapterView\nimport android.widget.Button\nimport android.widget.EditText\nimport android.widget.HorizontalScrollView\nimport android.widget.ImageButton\nimport android.widget.ImageView\nimport android.widget.LinearLayout\nimport android.widget.PopupWindow\nimport android.widget.ProgressBar\nimport android.widget.SimpleAdapter\nimport android.widget.Spinner\nimport android.widget.TextView\nimport android.widget.Toast\nimport androidx.appcompat.app.AlertDialog\nimport androidx.core.app.ActivityCompat\nimport androidx.core.content.ContextCompat\nimport androidx.core.content.FileProvider\nimport androidx.fragment.app.FragmentActivity\nimport androidx.recyclerview.widget.RecyclerView\nimport com.google.android.material.bottomsheet.BottomSheetDialog\nimport com.gyf.immersionbar.ktx.immersionBar\nimport com.hjq.toast.Toaster\nimport com.liulishuo.okdownload.DownloadContext\nimport com.liulishuo.okdownload.DownloadTask\nimport com.liulishuo.okdownload.OkDownload\nimport com.liulishuo.okdownload.core.cause.EndCause\nimport com.liulishuo.okdownload.core.connection.DownloadOkHttp3Connection\nimport com.liulishuo.okdownload.kotlin.listener.createDownloadContextListener\nimport com.liulishuo.okdownload.kotlin.listener.createListener1\nimport com.nexa.demo.bean.DownloadableFile\nimport com.nexa.demo.bean.DownloadableFileWithFallback\nimport com.nexa.demo.bean.DownloadState\nimport com.nexa.demo.bean.ModelData\nimport com.nexa.demo.bean.downloadableFiles\nimport com.nexa.demo.bean.downloadableFilesWithFallback\nimport com.nexa.demo.bean.getNexaManifest\nimport com.nexa.demo.bean.getNonExistModelFile\nimport com.nexa.demo.bean.getSupportPluginIds\nimport com.nexa.demo.bean.isNpuModel\nimport com.nexa.demo.bean.mmprojTokenFile\nimport com.nexa.demo.bean.modelDir\nimport com.nexa.demo.bean.modelFile\nimport com.nexa.demo.bean.tokenFile\nimport com.nexa.demo.bean.withFallbackUrls\nimport com.nexa.demo.utils.ModelFileListingUtil\nimport com.nexa.demo.databinding.ActivityMainBinding\nimport com.nexa.demo.databinding.DialogSelectPluginIdBinding\nimport com.nexa.demo.listeners.CustomDialogInterface\nimport com.nexa.demo.utils.ExecShell\nimport com.nexa.demo.utils.ImgUtil\nimport com.nexa.demo.utils.WavRecorder\nimport com.nexa.demo.utils.inflate\nimport com.nexa.sdk.AsrWrapper\nimport com.nexa.sdk.CvWrapper\nimport com.nexa.sdk.EmbedderWrapper\nimport com.nexa.sdk.LlmWrapper\nimport com.nexa.sdk.NexaSdk\nimport com.nexa.sdk.RerankerWrapper\nimport com.nexa.sdk.VlmWrapper\nimport com.nexa.sdk.bean.AsrCreateInput\nimport com.nexa.sdk.bean.AsrTranscribeInput\nimport com.nexa.sdk.bean.CVCapability\nimport com.nexa.sdk.bean.CVCreateInput\nimport com.nexa.sdk.bean.CVModelConfig\nimport com.nexa.sdk.bean.ChatMessage\nimport com.nexa.sdk.bean.EmbedderCreateInput\nimport com.nexa.sdk.bean.EmbeddingConfig\nimport com.nexa.sdk.bean.LlmCreateInput\nimport com.nexa.sdk.bean.LlmStreamResult\nimport com.nexa.sdk.bean.ModelConfig\nimport com.nexa.sdk.bean.RerankConfig\nimport com.nexa.sdk.bean.RerankerCreateInput\nimport com.nexa.sdk.bean.VlmChatMessage\nimport com.nexa.sdk.bean.VlmContent\nimport com.nexa.sdk.bean.VlmCreateInput\nimport com.nexa.sdk.bean.DeviceIdValue\nimport kotlinx.coroutines.CoroutineScope\nimport kotlinx.coroutines.Dispatchers\nimport kotlinx.coroutines.async\nimport kotlinx.coroutines.awaitAll\nimport kotlinx.coroutines.launch\nimport kotlinx.serialization.json.Json\nimport okhttp3.OkHttpClient\nimport okhttp3.Request\nimport java.io.File\nimport java.io.FileNotFoundException\nimport java.io.FileOutputStream\nimport java.security.SecureRandom\nimport java.security.cert.CertificateException\nimport java.security.cert.X509Certificate\nimport javax.net.ssl.SSLContext\nimport javax.net.ssl.SSLSession\nimport javax.net.ssl.SSLSocketFactory\nimport javax.net.ssl.TrustManager\nimport javax.net.ssl.X509TrustManager\n\nclass MainActivity : FragmentActivity() {\n\n private val binding: ActivityMainBinding by inflate()\n private var downloadContext: DownloadContext? = null\n private var downloadState = DownloadState.IDLE\n private var downloadingModelData: ModelData? = null\n private lateinit var spDownloaded: SharedPreferences\n private lateinit var llDownloading: LinearLayout\n private lateinit var tvDownloadProgress: TextView\n private lateinit var pbDownloading: ProgressBar\n private lateinit var spModelList: Spinner\n private lateinit var btnDownload: Button\n private lateinit var btnLoadModel: Button\n private lateinit var btnUnloadModel: Button\n private lateinit var btnStop: Button\n private lateinit var etInput: EditText\n private lateinit var btnSend: Button\n private lateinit var btnClearHistory: Button\n private lateinit var btnAddImage: Button\n private lateinit var btnAudioRecord: Button\n\n private lateinit var recyclerView: RecyclerView\n private lateinit var adapter: ChatAdapter\n private lateinit var bottomPanel: LinearLayout\n private lateinit var btnAudioDone: Button\n private lateinit var btnAudioCancel: Button\n\n private lateinit var scrollImages: HorizontalScrollView\n private lateinit var topScrollContainer: LinearLayout\n private lateinit var llLoading: LinearLayout\n private lateinit var vTip: View\n\n private lateinit var llmWrapper: LlmWrapper\n private lateinit var vlmWrapper: VlmWrapper\n var embedderWrapper: EmbedderWrapper? = null\n private lateinit var rerankerWrapper: RerankerWrapper\n private lateinit var cvWrapper: CvWrapper\n private lateinit var asrWrapper: AsrWrapper\n private val modelScope = CoroutineScope(Dispatchers.IO)\n\n private val chatList = arrayListOf()\n private lateinit var llmSystemPrompt: ChatMessage\n private val vlmChatList = arrayListOf()\n private lateinit var vlmSystemPrompty: VlmChatMessage\n private lateinit var modelList: List\n private var selectModelId = \"\"\n\n // ADD: Track which model type is loaded\n private var isLoadLlmModel = false\n private var isLoadVlmModel = false\n private var isLoadEmbedderModel = false\n private var isLoadRerankerModel = false\n private var isLoadCVModel = false\n private var isLoadAsrModel = false\n\n private var enableThinking = false\n\n private var wavRecorder: WavRecorder? = null\n private var audioFile: File? = null\n\n private val savedImageFiles = mutableListOf()\n private val messages = arrayListOf()\n\n override fun onCreate(savedInstanceState: Bundle?) {\n super.onCreate(savedInstanceState)\n immersionBar {\n statusBarColorInt(Color.WHITE)\n statusBarDarkFont(true)\n }\n requestPermissions(arrayOf(Manifest.permission.RECORD_AUDIO), 1002)\n okdownload()\n initData()\n initView()\n setListeners()\n }\n\n private fun resetLoadState() {\n isLoadLlmModel = false\n isLoadVlmModel = false\n isLoadEmbedderModel = false\n isLoadRerankerModel = false\n isLoadCVModel = false\n isLoadAsrModel = false\n }\n\n private fun initView() {\n adapter = ChatAdapter(messages)\n binding.rvChat.adapter = adapter\n\n llDownloading = findViewById(R.id.ll_downloading)\n tvDownloadProgress = findViewById(R.id.tv_download_progress)\n pbDownloading = findViewById(R.id.pb_downloading)\n spModelList = findViewById(R.id.sp_model_list)\n spModelList.adapter = object : SimpleAdapter(this, modelList.map {\n val map = mutableMapOf()\n map[\"displayName\"] = it.displayName\n map\n }, R.layout.item_model, arrayOf(\"displayName\"), intArrayOf(R.id.tv_model_id)) {\n\n }\n spModelList.onItemSelectedListener = object : AdapterView.OnItemSelectedListener {\n override fun onItemSelected(\n parent: AdapterView<*>?, view: View?, position: Int, id: Long\n ) {\n selectModelId = modelList[position].id\n\n messages.clear()\n adapter.notifyDataSetChanged()\n binding.rvChat.scrollTo(0, 0)\n }\n\n override fun onNothingSelected(parent: AdapterView<*>?) {\n selectModelId = \"\"\n }\n }\n btnDownload = findViewById(R.id.btn_download)\n btnLoadModel = findViewById(R.id.btn_load_model)\n btnUnloadModel = findViewById(R.id.btn_unload_model)\n btnStop = findViewById(R.id.btn_stop)\n etInput = findViewById(R.id.et_input)\n btnAddImage = findViewById(R.id.btn_add_image)\n btnAudioRecord = findViewById(R.id.btn_voice)\n\n bottomPanel = findViewById(R.id.bottom_panel)\n btnAudioCancel = findViewById(R.id.btn_audio_cancel)\n btnAudioDone = findViewById(R.id.btn_audio_done)\n\n btnSend = findViewById(R.id.btn_send)\n btnClearHistory = findViewById(R.id.btn_clear_history)\n scrollImages = findViewById(R.id.scroll_images)\n topScrollContainer = findViewById(R.id.ll_images_container)\n llLoading = findViewById(R.id.ll_loading)\n vTip = findViewById(R.id.v_tip)\n\n btnAudioCancel.setOnClickListener {\n stopRecord(true)\n }\n\n btnAudioDone.setOnClickListener {\n stopRecord(false)\n }\n\n findViewById\n \n\n \n
\n\n
\n
Hi!
\n
How can I help?
\n
\n\n \n
\n\n \n
\n
\n
\n \n
\n
\n \n \n
\n \n
\n
\n
\n \n\n \n\n\n" }, { "path": "cookbook/PC/function-calling/main.py", "content": "# Copyright 2024-2026 Nexa AI, Inc.\n#\n# Licensed under the Apache License, Version 2.0 (the \"License\");\n# you may not use this file except in compliance with the License.\n# You may obtain a copy of the License at\n#\n# http://www.apache.org/licenses/LICENSE-2.0\n#\n# Unless required by applicable law or agreed to in writing, software\n# distributed under the License is distributed on an \"AS IS\" BASIS,\n# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n# See the License for the specific language governing permissions and\n# limitations under the License.\n\n#!/usr/bin/env python3\n\"\"\"NexaAI VLM Function Call Demo with Google Calendar MCP\"\"\"\n\nimport asyncio\nimport json\nimport os\nimport sys\nimport argparse\nimport re\nfrom dataclasses import dataclass\nfrom typing import List, Dict, Any, Optional\n\nfrom nexaai import GenerationConfig, ModelConfig, VlmChatMessage, VlmContent, setup_logging\nfrom nexaai.vlm import VLM\nfrom mcp import ClientSession, StdioServerParameters\nfrom mcp.client.stdio import stdio_client\n\n\ndef _convert_schema_property(prop_schema: Dict[str, Any]) -> Dict[str, Any]:\n \"\"\"Recursively convert a schema property, handling nested objects.\"\"\"\n result = {\n \"type\": prop_schema.get(\"type\", \"string\"),\n }\n \n if \"description\" in prop_schema:\n result[\"description\"] = prop_schema[\"description\"]\n \n # Handle nested objects\n if prop_schema.get(\"type\") == \"object\" and \"properties\" in prop_schema:\n nested_props = {}\n nested_required = []\n \n for nested_name, nested_schema in prop_schema[\"properties\"].items():\n nested_props[nested_name] = _convert_schema_property(nested_schema)\n if nested_name in prop_schema.get(\"required\", []):\n nested_required.append(nested_name)\n \n result[\"properties\"] = nested_props\n if nested_required:\n result[\"required\"] = nested_required\n \n # Handle arrays of objects\n if prop_schema.get(\"type\") == \"array\" and \"items\" in prop_schema:\n result[\"items\"] = _convert_schema_property(prop_schema[\"items\"])\n \n return result\n\n\ndef mcp_tool_to_openai_format(tool) -> Dict[str, Any]:\n \"\"\"Convert MCP tool to OpenAI function calling format.\"\"\"\n properties = {}\n required = []\n \n if tool.inputSchema and \"properties\" in tool.inputSchema:\n for prop_name, prop_schema in tool.inputSchema[\"properties\"].items():\n properties[prop_name] = _convert_schema_property(prop_schema)\n if tool.inputSchema.get(\"required\") and prop_name in tool.inputSchema[\"required\"]:\n required.append(prop_name)\n \n return {\n \"type\": \"function\",\n \"function\": {\n \"name\": tool.name,\n \"description\": tool.description or \"\",\n \"parameters\": {\n \"type\": \"object\",\n \"properties\": properties,\n \"required\": required,\n }\n }\n }\n\n\nasync def get_mcp_tools(session: ClientSession) -> List[Dict[str, Any]]:\n \"\"\"Get tools from MCP server and convert to OpenAI format.\"\"\"\n result = await session.list_tools()\n return [mcp_tool_to_openai_format(tool) for tool in result.tools]\n\n\ndef normalize_tool_name(tool_name: str, available_tools: List[Dict[str, Any]]) -> str:\n \"\"\"Normalize tool name to match available tools.\"\"\"\n name_mappings = {\n \"create_calendar_event\": \"create-event\",\n \"create-event\": \"create-event\",\n \"list_calendar_events\": \"list-events\",\n \"list-events\": \"list-events\",\n \"update_calendar_event\": \"update-event\",\n \"update-event\": \"update-event\",\n \"delete_calendar_event\": \"delete-event\",\n \"delete-event\": \"delete-event\",\n \"get_current_time\": \"get-current-time\",\n \"get-current-time\": \"get-current-time\",\n }\n \n tool_names = [t.get(\"function\", {}).get(\"name\", \"\") for t in available_tools]\n \n if tool_name in name_mappings:\n normalized = name_mappings[tool_name]\n if normalized in tool_names:\n return normalized\n \n if tool_name in tool_names:\n return tool_name\n \n normalized = tool_name.replace(\"_\", \"-\")\n return normalized if normalized in tool_names else tool_name\n\n\nasync def execute_mcp_tool(session: ClientSession, tool_name: str, arguments: Dict[str, Any], \n available_tools: Optional[List[Dict[str, Any]]] = None) -> str:\n \"\"\"Execute a tool call via MCP server.\"\"\"\n try:\n if available_tools:\n tool_name = normalize_tool_name(tool_name, available_tools)\n result = await session.call_tool(tool_name, arguments=arguments)\n return result.model_dump_json(indent=2)\n except Exception as e:\n return f\"Error: {str(e)}\"\n\n\ndef create_calendar_server(credentials: str) -> StdioServerParameters:\n \"\"\"Create Google Calendar MCP server parameters.\"\"\"\n if not os.path.exists(credentials):\n raise FileNotFoundError(\n f\"Credentials file not found: {credentials}\\n\"\n f\"Please create the OAuth credentials file at: {os.path.abspath(credentials)}\"\n )\n return StdioServerParameters(\n command=\"npx\",\n args=[\"-y\", \"@cocal/google-calendar-mcp\"],\n env={\"GOOGLE_OAUTH_CREDENTIALS\": os.path.abspath(credentials)},\n )\n\n\ndef extract_function_call(text: str):\n \"\"\"Extract function call JSON from LLM response.\"\"\"\n if not text:\n return None\n \n text = re.sub(r\"<\\|[^|]+\\|>\", \"\", text.strip())\n \n try:\n parsed = json.loads(text)\n if isinstance(parsed, dict) and \"name\" in parsed:\n return parsed.get(\"name\"), parsed.get(\"arguments\", {})\n except json.JSONDecodeError:\n json_start = text.find('{')\n if json_start == -1:\n return None\n \n brace_count = 0\n for i in range(json_start, len(text)):\n if text[i] == '{':\n brace_count += 1\n elif text[i] == '}':\n brace_count -= 1\n if brace_count == 0:\n json_str = text[json_start:i + 1]\n try:\n parsed = json.loads(json_str)\n if isinstance(parsed, dict) and \"name\" in parsed:\n return parsed.get(\"name\"), parsed.get(\"arguments\", {})\n except json.JSONDecodeError:\n pass\n break\n \n return None\n\n\ndef _format_nested_properties(props: Dict[str, Any], required: List[str], prefix: str = \"\", indent: int = 2) -> List[str]:\n \"\"\"Recursively format nested object properties with required field indicators.\"\"\"\n param_list = []\n indent_str = \" \" * indent\n \n for param_name, param_info in props.items():\n param_type = param_info.get('type', 'string')\n param_desc = param_info.get('description', '')\n is_required = param_name in required\n full_param_name = f\"{prefix}.{param_name}\" if prefix else param_name\n \n # Handle nested objects\n if param_type == 'object' and 'properties' in param_info:\n nested_props = param_info.get('properties', {})\n nested_required = param_info.get('required', [])\n req_mark = \" (REQUIRED)\" if is_required else \"\"\n \n if param_desc:\n param_list.append(f\"{indent_str}{param_name} (object){req_mark}: {param_desc}\")\n else:\n param_list.append(f\"{indent_str}{param_name} (object){req_mark}\")\n \n # Recursively format nested properties\n nested_params = _format_nested_properties(\n nested_props, nested_required, full_param_name, indent + 2\n )\n param_list.extend(nested_params)\n else:\n req_mark = \" (REQUIRED)\" if is_required else \"\"\n if param_desc:\n param_list.append(f\"{indent_str}{param_name} ({param_type}){req_mark}: {param_desc}\")\n else:\n param_list.append(f\"{indent_str}{param_name} ({param_type}){req_mark}\")\n \n return param_list\n\n\ndef build_system_prompt(tools: list) -> str:\n \"\"\"Build system prompt from tool schemas.\"\"\"\n tools_descriptions = []\n for t in tools:\n func = t.get('function', {})\n name = func.get('name', '')\n desc = func.get('description', '')\n params = func.get('parameters', {})\n props = params.get('properties', {})\n required = params.get('required', [])\n \n # Format parameters with nested object support\n param_list = _format_nested_properties(props, required)\n \n params_str = \"\\n\".join(param_list) if param_list else \" (no parameters)\"\n \n # Highlight required parameters at the top\n required_params = [p for p in required]\n required_str = f\"\\n REQUIRED parameters: {', '.join(required_params)}\" if required_params else \"\"\n \n tools_descriptions.append(f\"{name}: {desc}{required_str}\\n{params_str}\")\n \n tools_list = \"\\n\\n\".join([f\"{i+1}. {td}\" for i, td in enumerate(tools_descriptions)])\n \n # Add example for create-event\n example_json = \"\"\"{\n \"name\": \"create-event\",\n \"arguments\": {\n \"calendarId\": \"primary\",\n \"summary\": \"Meeting\",\n \"start\": \"2025-01-01T10:00:00\",\n \"end\": \"2025-01-01T11:00:00\"\n }\n}\"\"\"\n \n return f\"\"\"You are a calendar assistant. When the user requests calendar actions, respond with ONLY a JSON object in this format:\n\n{{\"name\": \"function_name\", \"arguments\": {{\"param\": \"value\"}}}}\n\nCRITICAL RULES:\n- You MUST include ALL required parameters (marked as REQUIRED)\n- For nested objects, ALL required fields within the object must be included\n- If a parameter is marked as REQUIRED, it cannot be omitted\n- Output ONLY valid JSON, no other text before or after\n- Use exact function and parameter names (case-sensitive)\n\nExample for create-event:\n{example_json}\n\nAvailable functions:\n{tools_list}\n\nIMPORTANT: Before creating events, you may need to call get-current-time first to get accurate date/time context. Always include calendarId=\"primary\" for create-event unless specified otherwise.\n\"\"\"\n\n\ndef _handle_function_call_error(error_text: str, func_args: Dict[str, Any]) -> bool:\n \"\"\"Handle function call errors and auto-fix parameters. Returns True if should retry.\"\"\"\n # Auto-fix account errors\n if \"Account\" in error_text and \"not found\" in error_text and \"Available accounts:\" in error_text:\n match = re.search(r'Available accounts:\\s*(\\w+)', error_text)\n if match:\n func_args['account'] = match.group(1)\n return True\n elif 'account' in func_args:\n del func_args['account']\n return True\n \n # Auto-fix eventId errors\n if \"Invalid event ID\" in error_text or (\"event ID\" in error_text.lower() and \"invalid\" in error_text.lower()):\n if 'eventId' in func_args:\n del func_args['eventId']\n return True\n \n # Auto-remove optional parameters that cause errors\n if \"validation error\" in error_text.lower() or \"invalid\" in error_text.lower():\n optional_params = ['account', 'eventId', 'timeZone', 'fields']\n for param in optional_params:\n if param in func_args:\n del func_args[param]\n return True\n \n return False\n\n\nasync def _execute_with_retry(session: ClientSession, func_name: str, func_args: Dict[str, Any], \n tools: List[Dict[str, Any]], max_retries: int = 3) -> str:\n \"\"\"Execute function call with automatic error handling and retry.\"\"\"\n retry_count = 0\n func_result = \"\"\n \n while retry_count <= max_retries:\n func_result = await execute_mcp_tool(session, func_name, func_args, tools)\n \n try:\n result_data = json.loads(func_result) if isinstance(func_result, str) else func_result\n if result_data.get('isError', False):\n error_text = \"\"\n if isinstance(result_data.get('content'), list):\n for item in result_data['content']:\n if item.get('type') == 'text':\n error_text = item.get('text', '')\n break\n \n if retry_count < max_retries and _handle_function_call_error(error_text, func_args):\n retry_count += 1\n continue\n break\n except Exception:\n break\n \n return func_result or \"\"\n\n\ndef init_vlm(tools: List[Dict[str, Any]]) -> VLM:\n \"\"\"Initialize VLM with tools.\"\"\"\n system_prompt = build_system_prompt(tools)\n print('[debug] system_prompt:', system_prompt)\n return VLM.from_(\"NexaAI/OmniNeural-4B\", config=ModelConfig(\n system_prompt=system_prompt,\n n_ctx=4096, n_threads=0, n_threads_batch=0, n_batch=0, \n n_ubatch=0, n_seq_max=0, n_gpu_layers=999\n ))\n\n\n\n@dataclass\nclass FunctionCallAgentResult:\n \"\"\"Result of function call agent execution.\"\"\"\n func_name: Optional[str]\n func_result: Optional[str]\n response_text: str\n\nasync def call_agent(\n vlm: VLM,\n session: ClientSession,\n tools: List[Dict[str, Any]],\n text: Optional[str] = None,\n image: Optional[str] = None,\n audio: Optional[str] = None\n) -> FunctionCallAgentResult:\n if not text and not image and not audio:\n raise ValueError(\"At least one of text, image, or audio must be provided\")\n \n contents = []\n image_paths = []\n audio_paths = []\n \n if image:\n image_path = os.path.abspath(image)\n if not os.path.exists(image_path):\n raise FileNotFoundError(f\"Image file not found: {image_path}\")\n image_paths.append(image_path)\n contents.append(VlmContent(type=\"image\", text=image_path))\n \n if audio:\n audio_path = os.path.abspath(audio)\n if not os.path.exists(audio_path):\n raise FileNotFoundError(f\"Audio file not found: {audio_path}\")\n audio_paths.append(audio_path)\n contents.append(VlmContent(type=\"audio\", text=audio_path))\n \n if text:\n contents.append(VlmContent(type=\"text\", text=text))\n\n conversation = [VlmChatMessage(role=\"user\", contents=contents)]\n \n # Generate initial response\n prompt = vlm.apply_chat_template(conversation)\n print('[debug] prompt:', prompt)\n print('[debug] generate_stream...')\n response_text = \"\"\n for token in vlm.generate_stream(prompt, config=GenerationConfig(\n max_tokens=2048, image_paths=image_paths or None, \n audio_paths=audio_paths or None, image_max_length=512\n )):\n print(token, end=\"\", flush=True)\n response_text += token\n print()\n print('[debug] response_text:', response_text)\n func_call = extract_function_call(response_text)\n if not func_call:\n print(f\"[error] Failed to extract function call from response\")\n return FunctionCallAgentResult(\n func_name=None,\n func_result=None,\n response_text=response_text\n )\n \n func_name, func_args = func_call\n if func_name and isinstance(func_name, str):\n print('[debug] calling function:', func_name)\n func_result = await _execute_with_retry(session, func_name, func_args, tools)\n print('[debug] func_result:', func_result)\n \n # Parse function result to extract success/error message\n result_message = \"\"\n try:\n result_data = json.loads(func_result) if isinstance(func_result, str) else func_result\n if result_data.get('isError', False):\n # Extract error message\n if isinstance(result_data.get('content'), list):\n for item in result_data['content']:\n if item.get('type') == 'text':\n result_message = item.get('text', '')\n break\n else:\n # Extract success message or summary\n if isinstance(result_data.get('content'), list):\n for item in result_data['content']:\n if item.get('type') == 'text':\n result_message = item.get('text', '')\n break\n except Exception:\n result_message = str(func_result)\n \n followup = conversation + [\n VlmChatMessage(role=\"assistant\", contents=[VlmContent(type=\"text\", text=response_text)]),\n VlmChatMessage(role=\"user\", contents=[VlmContent(type=\"text\", \n text=f\"Function execution completed. Result: {result_message}\\n\\n\"\n f\"Now respond to the user in natural language. You are in RESPONSE MODE, not function calling mode.\\n\"\n f\"- DO NOT output any JSON format\\n\"\n f\"- DO NOT use {{}} brackets\\n\"\n f\"- DO NOT call any function\\n\"\n f\"- Just speak naturally like a helpful assistant\\n\"\n f\"- Tell the user what happened with the calendar event in a friendly way\")])\n ]\n followup_response = \"\"\n for token in vlm.generate_stream(\n vlm.apply_chat_template(followup, enable_thinking=False),\n config=GenerationConfig(max_tokens=2048)\n ):\n followup_response += token\n return FunctionCallAgentResult(\n func_name=func_name,\n func_result=func_result,\n response_text=followup_response\n )\n \n return FunctionCallAgentResult(\n func_name=None,\n func_result=None,\n response_text=response_text\n )\n\n\nasync def call_agent_wrapper(\n text: Optional[str] = None,\n image: Optional[str] = None,\n audio: Optional[str] = None,\n credentials: str = \"gcp-oauth.keys.json\"\n) -> FunctionCallAgentResult:\n setup_logging()\n\n if not text and not image and not audio:\n raise ValueError(\"At least one of text, image, or audio must be provided\") \n \n server = create_calendar_server(credentials)\n async with stdio_client(server) as (read, write):\n async with ClientSession(read, write) as session:\n await session.initialize()\n tools = await get_mcp_tools(session)\n tools = [t for t in tools if t.get('function', {}).get('name', '') in \n ['create-event', 'get-current-time']]\n \n vlm = init_vlm(tools)\n result = await call_agent(vlm, session, tools, text, image, audio)\n return result\n \n \nasync def main():\n \"\"\"Command-line interface for the agent.\"\"\"\n setup_logging()\n \n parser = argparse.ArgumentParser()\n parser.add_argument(\"--credentials\", default=\"gcp-oauth.keys.json\")\n parser.add_argument(\"--text\", help=\"Text input\")\n parser.add_argument(\"--image\", help=\"Image file path\")\n parser.add_argument(\"--audio\", help=\"Audio file path\")\n args = parser.parse_args()\n \n if not args.text and not args.image and not args.audio:\n parser.print_help()\n return\n \n server = create_calendar_server(args.credentials)\n async with stdio_client(server) as (read, write):\n async with ClientSession(read, write) as session:\n await session.initialize()\n tools = await get_mcp_tools(session)\n tools = [t for t in tools if t.get('function', {}).get('name', '') in \n ['create-event', 'get-current-time']]\n \n vlm = init_vlm(tools)\n result = await call_agent(vlm, session, tools, args.text, args.image, args.audio)\n\n if result.response_text:\n print(result.response_text)\n \n\nif __name__ == \"__main__\":\n asyncio.run(main())\n" }, { "path": "cookbook/PC/function-calling/requirements.txt", "content": "mcp\nFlask[async]" }, { "path": "cookbook/PC/live-translate/README.md", "content": "# NexaAI Live Translator\n\nReal-time speech recognition and language translation demo using NexaAI's ASR and LLM models. Supports multi-language translation with low-latency sentence-level translation.\n\n## Features\n\n- 🎙️ **Real-time Speech Recognition**: Uses NexaAI's Parakeet ASR model for multi-language transcription\n- 🌐 **Instant Translation**: LLM provides fast, accurate translations\n- 📊 **Segment-based**: Automatic sentence/paragraph boundary detection for natural translations\n- 🖥️ **Modern Web UI**: Clean, responsive interface with live text display\n\n## Requirements\n- Windows/Linux with Qualcomm NPU device.\n- Python 3.9+\n\n## Installation\n\n### 1. Clone and Setup Environment\n\n```bash\ncd cookbook/PC/live-translate\npython -m venv .venv\n\n# Windows\n.venv\\Scripts\\activate\n# Linux\nsource .venv/bin/activate\n```\n\n### 2. Install Dependencies\n\n```bash\npip install -r requirements.txt\n```\n\n### 3. Download Models\n\nDownload the required models using NexaAI CLI:\n\n```bash\nnexa pull NexaAI/parakeet-tdt-0.6b-v3-npu\nnexa pull NexaAI/HY-MT1.5-1.8B-npu\n```\n\n## Usage\n\n### Quick Start\n\n```bash\npython app.py\n```\n\nThen open your browser to: **http://localhost:5000**\n\n\n### How to Use\n\n1. **Select Target Language**: Choose your desired translation language from the dropdown (e.g., Chinese, English, French, etc.)\n2. **Start Recording**: Click the \"Start Recording\" button\n3. **Speak**: Speak naturally into your microphone\n4. **Watch Real-time Translation**: \n - Left panel shows real-time transcription\n - Right panel shows real-time translation (auto-updated per sentence)\n5. **Stop Recording**: Click \"Stop Recording\" to end\n" }, { "path": "cookbook/PC/live-translate/app.py", "content": "import logging\nimport atexit\nimport threading\nimport numpy as np\nfrom queue import Queue, Empty\nfrom typing import Optional\nfrom flask import Flask, jsonify, render_template, request\nfrom flask_cors import CORS\nfrom flask_socketio import SocketIO, emit\nfrom nexaai import LLM, GenerationConfig, SamplerConfig, ASR, setup_logging, LlmChatMessage\nfrom nexaai.asr import ASRStreamConfig\n\nlogging.basicConfig(level=logging.INFO, format='[%(asctime)s] %(levelname)s: %(message)s', force=True)\nsetup_logging(level=logging.DEBUG)\nlogger = logging.getLogger(__name__)\n\napp = Flask(__name__, template_folder='.', static_folder='.')\nCORS(app)\n\nsocketio = SocketIO(app, cors_allowed_origins='*', async_mode='threading', logger=False, engineio_logger=False)\n\nasr_model: Optional[ASR] = None\nllm_model: Optional[LLM] = None\nstream_managers = {}\n\n\nclass TranslationStreamManager:\n def __init__(self, asr_model: ASR, llm_model: LLM, sid: str, target_language: str = 'zh'):\n self.asr = asr_model\n self.llm = llm_model\n self.sid = sid\n self.stream = None\n self.stream_context = None\n self.stream_active = False\n self.target_language = target_language\n self.audio_queue: Queue[bytes] = Queue(maxsize=50)\n self.audio_thread: Optional[threading.Thread] = None\n self.stop_event = threading.Event()\n self.translation_queue: Queue[Optional[str]] = Queue(maxsize=20)\n self.translation_thread: Optional[threading.Thread] = None\n self.translation_degraded = False\n self.last_committed_source = ''\n self.pending_text = ''\n self.stable_count = 0\n\n def start_stream(self, target_language: str = 'zh'):\n self.target_language = target_language\n self.stream_active = True\n logger.info('Starting ASR stream')\n config = ASRStreamConfig(\n sample_rate=16000,\n chunk_duration=4.0,\n overlap_duration=3.5,\n max_queue_size=10,\n buffer_size=1024,\n timestamps='segment',\n beam_size=4,\n )\n\n def on_transcription(text: str):\n try:\n if text and text.strip():\n self.on_new_segment(text)\n except Exception as e:\n logger.error(f'Error in transcription callback: {e}', exc_info=True)\n socketio.emit('error', {'message': f'Transcription error: {e}'}, to=self.sid)\n\n try:\n self.stream_context = self.asr.stream(language='en', config=config)\n self.stream = self.stream_context.__enter__()\n self.stream.start(on_transcription=on_transcription)\n logger.info('ASR stream started successfully')\n self.stop_event.clear()\n self.audio_thread = threading.Thread(\n target=self._audio_worker, name=f'audio-worker-{self.sid}', daemon=True\n )\n self.audio_thread.start()\n self.translation_thread = threading.Thread(\n target=self._translation_worker, name=f'translation-worker-{self.sid}', daemon=True\n )\n self.translation_thread.start()\n except Exception as e:\n logger.error(f'Error starting stream: {e}', exc_info=True)\n if self.stream_context is not None:\n try:\n self.stream_context.__exit__(None, None, None)\n except Exception:\n pass\n self.stream = None\n self.stream_active = False\n socketio.emit('error', {'message': f'Failed to start stream: {e}'}, to=self.sid)\n\n def push_audio(self, audio_bytes: bytes):\n if not self.stream_active:\n return\n if len(audio_bytes) % 2 == 1:\n audio_bytes = audio_bytes[:-1]\n if not audio_bytes:\n return\n try:\n self.audio_queue.put_nowait(audio_bytes)\n except Exception as e:\n logger.error(f'Error enqueuing audio: {e}')\n socketio.emit('error', {'message': f'Audio queue error: {e}'}, to=self.sid)\n\n def _audio_worker(self):\n try:\n while not self.stop_event.is_set():\n try:\n audio_bytes = self.audio_queue.get(timeout=0.5)\n except Empty:\n continue\n\n try:\n audio_array = np.frombuffer(audio_bytes, dtype=np.int16).astype(np.float32) / 32768.0\n if self.stream and self.stream_active:\n self.stream.push_audio(audio_array.tolist())\n except Exception as e:\n logger.error(f'Error pushing audio in worker: {e}')\n socketio.emit('error', {'message': f'Audio processing error: {e}'}, to=self.sid)\n finally:\n self.audio_queue.task_done()\n except Exception as e:\n logger.error(f'Audio worker crashed: {e}', exc_info=True)\n socketio.emit('error', {'message': f'Audio worker error: {e}'}, to=self.sid)\n\n def on_new_segment(self, segment_text: str):\n if not segment_text.strip():\n return\n try:\n logger.info(f'New segment: {segment_text}')\n if segment_text == self.pending_text:\n self.stable_count += 1\n else:\n self.pending_text = segment_text\n self.stable_count = 1\n socketio.emit('transcription', {'original': segment_text}, to=self.sid)\n is_sentence_end = segment_text.rstrip().endswith(('.', '?', '!'))\n if self.stable_count < 3 and not is_sentence_end:\n logger.debug(f'Waiting for stabilization (count={self.stable_count}, sentence_end={is_sentence_end})')\n return\n if segment_text == self.last_committed_source:\n logger.debug('Skipping translation duplicate (unchanged source)')\n return\n self.last_committed_source = segment_text\n logger.info('Committing stabilized segment for translation')\n self._enqueue_translation(segment_text)\n except Exception as e:\n logger.error(f'Error processing segment: {e}', exc_info=True)\n socketio.emit('error', {'message': f'Segment processing error: {e}'}, to=self.sid)\n\n def _translate_text(self, text: str, target_lang: str) -> Optional[str]:\n try:\n prompt = f\"\"\"\nTranslate the following segment into {target_lang}, without additional explanation.\\n\\n{text}\n\"\"\"\n prompt = self.llm.apply_chat_template([LlmChatMessage(role='user', content=prompt)], enable_thinking=False)\n logger.info(f'Translation prompt: {prompt}')\n result = self.llm.generate(\n prompt,\n GenerationConfig(max_tokens=256, sampler_config=SamplerConfig(temperature=0.3)),\n )\n self.llm.reset()\n logger.info(f'Translation result: {result.full_text}')\n return result.full_text\n except Exception as e:\n logger.error(f'Translation error: {e}', exc_info=True)\n return None\n\n def _enqueue_translation(self, text: str):\n if self.translation_degraded:\n logger.error(f'Translation degraded; dropping text: {text}')\n return\n try:\n self.translation_queue.put_nowait(text)\n except Exception as e:\n logger.error(f'Failed to enqueue translation: {e}')\n\n def _translation_worker(self):\n while not self.stop_event.is_set():\n try:\n text = self.translation_queue.get(timeout=0.5)\n except Empty:\n continue\n if text is None:\n self.translation_queue.task_done()\n break\n try:\n translated = self._translate_text(text, self.target_language)\n if translated:\n logger.info(f'Translated to: {translated}')\n socketio.emit(\n 'translation',\n {'translated': translated, 'original': text, 'language': self.target_language},\n to=self.sid,\n )\n else:\n socketio.emit('error', {'message': 'Translation failed'}, to=self.sid)\n except Exception as e:\n logger.error(f'Translation worker error: {e}', exc_info=True)\n finally:\n self.translation_queue.task_done()\n\n def stop_stream(self):\n self.stream_active = False\n\n self.stop_event.set()\n try:\n self.audio_queue.put_nowait(b'')\n except Exception:\n pass\n if self.audio_thread and self.audio_thread.is_alive():\n self.audio_thread.join(timeout=2.0)\n self.audio_thread = None\n\n try:\n self.translation_queue.put_nowait(None)\n except Exception:\n pass\n if self.translation_thread and self.translation_thread.is_alive():\n self.translation_thread.join(timeout=2.0)\n self.translation_thread = None\n\n if self.stream:\n try:\n self.stream.stop(graceful=True)\n self.stream = None\n except Exception as e:\n logger.error(f'Error stopping stream: {e}')\n\n if self.stream_context is not None:\n try:\n self.stream_context.__exit__(None, None, None)\n except Exception as e:\n logger.error(f'Error closing stream context: {e}')\n\n logger.info('ASR stream stopped')\n\n\ndef initialize_models():\n global asr_model, llm_model\n asr_model = ASR.from_(model='NexaAI/parakeet-tdt-0.6b-v3-npu')\n llm_model = LLM.from_(model='NexaAI/HY-MT1.5-1.8B-npu')\n logger.info('✓ LLM model loaded successfully')\n\n\n# ============================================================================\n# HTTP Routes\n# ============================================================================\n\n\n@app.route('/')\ndef index():\n return render_template('index.html')\n\n\n@app.route('/api/health', methods=['GET'])\ndef health():\n return jsonify(\n {\n 'status': 'ok',\n 'asr_loaded': asr_model is not None,\n 'llm_loaded': llm_model is not None,\n }\n )\n\n\n@app.route('/api/translate-segment', methods=['POST'])\ndef translate_segment():\n try:\n data = request.get_json()\n text = data.get('text', '').strip()\n target_lang = data.get('target_lang', 'zh')\n if not text:\n return jsonify({'error': 'Empty text'}), 400\n logger.info(f\"REST API: Translating '{text}' to {target_lang}\")\n if asr_model is None or llm_model is None:\n return jsonify({'error': 'Models not loaded'}), 500\n manager = TranslationStreamManager(asr_model, llm_model, 'rest-api', target_lang)\n translated = manager._translate_text(text, target_lang)\n if translated is None:\n return jsonify({'error': 'Translation failed'}), 500\n return jsonify(\n {\n 'original': text,\n 'translated': translated,\n 'target_lang': target_lang,\n }\n )\n\n except Exception as e:\n logger.error(f'Error in translate_segment: {e}', exc_info=True)\n return jsonify({'error': str(e)}), 500\n\n\n# ============================================================================\n# WebSocket Events\n# ============================================================================\n\n\n@socketio.on('connect')\ndef handle_connect():\n emit('connect', {'data': 'Connected to translation server'})\n\n\n@socketio.on('disconnect')\ndef handle_disconnect():\n sid = request.sid\n logger.info(f'Client disconnected: {sid}')\n\n if sid in stream_managers:\n stream_managers[sid].stop_stream()\n del stream_managers[sid]\n\n\n@socketio.on('start_stream')\ndef handle_start_stream(data):\n sid = request.sid\n if isinstance(data, dict):\n target_language = data.get('target_lang', 'zh')\n else:\n target_language = 'zh'\n try:\n logger.info(f'[{sid}] start_stream event received, target_language: {target_language}')\n logger.debug(f'[{sid}] asr_model loaded: {asr_model is not None}, llm_model loaded: {llm_model is not None}')\n if asr_model and llm_model:\n logger.info(f'[{sid}] Creating TranslationStreamManager...')\n manager = TranslationStreamManager(asr_model, llm_model, sid, target_language)\n logger.info(f'[{sid}] Starting ASR stream...')\n manager.start_stream(target_language)\n stream_managers[sid] = manager\n logger.info(f'[{sid}] Stream registered in stream_managers, total streams: {len(stream_managers)}')\n emit(\n 'stream_started',\n {\n 'status': 'ok',\n 'target_language': target_language,\n },\n )\n logger.info(f'[{sid}] stream_started event emitted')\n else:\n logger.error(\n f'[{sid}] Models not loaded! asr_model={asr_model is not None}, llm_model={llm_model is not None}'\n )\n emit('error', {'message': 'Models not loaded'})\n except Exception as e:\n logger.error(f'[{sid}] Error starting stream: {e}', exc_info=True)\n emit('error', {'message': f'Failed to start stream: {e}'})\n\n\n@socketio.on('audio_chunk')\ndef handle_audio_chunk(data):\n sid = request.sid\n\n if sid not in stream_managers:\n logger.warning(\n f'[{sid}] Received audio_chunk but stream not started. Active streams: {list(stream_managers.keys())}'\n )\n emit('error', {'message': 'Stream not started. Click \"Start Recording\" first.'})\n return\n\n try:\n if isinstance(data, (bytes, bytearray)):\n audio_bytes = bytes(data)\n else:\n raise ValueError(f'Unsupported audio payload type: {type(data)}')\n\n if not audio_bytes:\n raise ValueError('Empty audio payload')\n\n if len(audio_bytes) % 2 == 1:\n audio_bytes = audio_bytes[:-1]\n\n logger.debug(f'[{sid}] Received audio chunk: {len(audio_bytes)} bytes')\n stream_managers[sid].push_audio(audio_bytes)\n\n except Exception as e:\n logger.error(f'[{sid}] Error processing audio chunk: {e}', exc_info=True)\n emit('error', {'message': f'Audio processing error: {e}'})\n\n\n@socketio.on('stop_stream')\ndef handle_stop_stream():\n sid = request.sid\n\n if sid and sid in stream_managers:\n stream_managers[sid].stop_stream()\n del stream_managers[sid]\n emit('stream_stopped', {'status': 'ok'})\n logger.info(f'[{sid}] Stream stopped')\n\n\ndef cleanup_models():\n global asr_model, llm_model\n logger.info('Cleaning up models...')\n\n if asr_model:\n try:\n del asr_model\n asr_model = None\n except Exception as e:\n logger.error(f'Error cleaning up ASR model: {e}')\n\n if llm_model:\n try:\n del llm_model\n llm_model = None\n except Exception as e:\n logger.error(f'Error cleaning up LLM model: {e}')\n\n logger.info('Models cleaned up')\n\n\nif __name__ == '__main__':\n logger.info('=' * 80)\n logger.info('Starting NexaAI Live Translator...')\n logger.info('=' * 80)\n atexit.register(cleanup_models)\n initialize_models()\n\n logger.info('✓ Starting Flask+SocketIO server on http://127.0.0.1:5000')\n logger.info('=' * 80)\n socketio.run(\n app,\n host='127.0.0.1',\n port=5000,\n )\n" }, { "path": "cookbook/PC/live-translate/index.html", "content": "\n\n\n\n \n \n NexaAI Live Translator\n \n \n\n\n\n
\n
\n

🎙️ NexaAI Live Translator

\n

Real-time speech recognition and translation

\n
\n\n
\n \n
\n
\n Original\n
\n
\n
\n
\n\n \n
\n
\n Translation\n \n
\n
\n
\n
\n
\n\n \n
\n \n \n \n
\n\n
\n
\n\n \n\n\n" }, { "path": "cookbook/PC/live-translate/requirements.txt", "content": "nexaai>=v1.0.44-rc1\nFlask>=2.3.0\nflask-cors>=4.0.0\nflask-socketio>=5.3.0\npython-socketio>=5.9.0\nnumpy>=1.24.0\njsonify>=0.5\n" }, { "path": "cookbook/README.md", "content": "# Nexa SDK Cookbook\n\nThis repository contains comprehensive demos and examples for Nexa SDK across different platforms: **PC (Python)**, **Android**, and **Linux**.\n\n---\n\n## 🖥️ PC Demos (Python Bindings)\n\nAll PC demos are located in the [`PC/`](./PC) directory and run on **Windows (x64/ARM64 Snapdragon)**, **macOS**, and **Linux**.\n\n### 🤖 Agent-Granite\n\nLightweight on-device AI assistant with function calling (web search) using Granite-4-Micro model. Includes Gradio interface for local agentic workflows.\n\n- [Python-Binding-Example](./PC/Agent-Granite/Python-Binding-Example)\n- [Serve-Example](./PC/Agent-Granite/Serve-Example)\n\n### 🔧 Function-Calling\n\nFunction calling capabilities with NexaAI VLM model, integrated with Google Calendar via MCP protocol. Supports multi-modal input (text, image, audio) with Web UI and CLI interfaces.\n\n- [Demo](./PC/function-calling)\n\n### 📚 RAG-LLM\n\nEnd-to-end Retrieval-Augmented Generation pipeline with embeddings, reranking, and generation models. Query your own documents (PDFs, Word, text) locally on device.\n\n- [Python-Binding-Example](./PC/RAG-LLM/Python-Binding-Example)\n- [Serve-Example](./PC/RAG-LLM/Serve-Example)\n\n### 🖼️ Multimodal-Qwen3VL\n\nVision-Language Model (VLM) with structured JSON outputs, multi-image input, and custom system prompt control.\n\n- [Python-Binding-Example](./PC/Multimodal-Qwen3VL/Python-Binding-Example)\n\n### 🔍 RAG-VLM\n\nLightweight RAG system with Qwen3VL multimodal model powered by Nexa Serve. Supports PDFs, Word docs, text files, and images. Includes CLI and Gradio UI.\n\n- [Demo](./PC/RAG-VLM)\n\n### 🌐 Web-Agent-Qwen3VL\n\nLocal multimodal LLM-driven browser automation using Qwen3-VL. Enables autonomous web navigation, searches, and complex web tasks.\n\n- [Demo](./PC/Web-Agent-Qwen3VL)\n\n---\n\n## 📱 Android Demos\n\nFor Android SDK demos and examples, please refer to [`android/`](./android).\n\n---\n\n## 🐧 Linux Demos\n\nLinux-specific demos coming soon in the [`linux/`](./linux) directory." }, { "path": "cookbook/README_zh.md", "content": "# Nexa SDK 使用指南\n\n本仓库包含 Nexa SDK 在不同平台的完整演示和示例：**PC (Python)**、**Android** 和 **Linux**。\n\n---\n\n## 📁 目录结构\n\n- **`PC/`** - 适用于 Windows、macOS 和 Linux 的 Python 绑定演示\n- **`android/`** - Android 平台专属演示（即将推出，Android SDK 请参考 [bindings/android](../bindings/android)）\n- **`linux/`** - Linux 专属演示（即将推出）\n\n---\n\n## 🖥️ PC 演示（Python 绑定）\n\n所有 PC 演示位于 [`PC/`](./PC) 目录，支持 **Windows (x64/ARM64 Snapdragon)**、**macOS** 和 **Linux**。\n\n### 🤖 Agent-Granite\n\n轻量级本地 AI 助手，支持函数调用（网页搜索），基于 Granite-4-Micro 模型。包含 Gradio 界面，可实现本地 Agent 流程。\n\n- [Python绑定示例](./PC/Agent-Granite/Python-Binding-Example)\n- [服务示例](./PC/Agent-Granite/Serve-Example)\n\n### 🔧 函数调用\n\n使用 NexaAI VLM 模型实现函数调用能力，并通过 MCP 协议集成 Google 日历。支持多模态输入（文本、图片、音频），包含 Web UI 和 CLI 界面。\n\n- [演示](./PC/function-calling)\n\n### 📚 RAG-LLM\n\n端到端的检索增强生成（RAG）流程，集成嵌入、重排序与生成模型。可在本地查询 PDF、Word、文本等文档。\n\n- [Python绑定示例](./PC/RAG-LLM/Python-Binding-Example)\n- [服务示例](./PC/RAG-LLM/Serve-Example)\n\n### 🖼️ 多模态-Qwen3VL\n\n视觉语言模型（VLM），支持结构化 JSON 输出、多图片输入及自定义系统提示。\n\n- [Python绑定示例](./PC/Multimodal-Qwen3VL/Python-Binding-Example)\n\n### 🔍 RAG-VLM\n\n采用 Qwen3VL 多模态模型的轻量级 RAG 系统，由 Nexa Serve 驱动。支持 PDF、Word、文本、图片。带有 CLI 和 Gradio UI。\n\n- [演示](./PC/RAG-VLM)\n\n### 🌐 Web-Agent-Qwen3VL\n\n本地多模态 LLM 驱动的网页自动化，基于 Qwen3-VL。支持自主网页导航、搜索及复杂网络任务。\n\n- [演示](./PC/Web-Agent-Qwen3VL)\n\n---\n\n## 📱 Android 演示\n\nAndroid SDK 演示和示例请参考 [`bindings/android`](../bindings/android)。\n\n---\n\n## 🐧 Linux 演示\n\nLinux 专属演示即将在 [`linux/`](./linux) 目录推出。\n\n---\n\n## 🔒 隐私优先\n\n**所有演示均在本地运行** — 数据不离开你的设备。\n\n" }, { "path": "cookbook/ag2/README.md", "content": "# AG2 + Nexa SDK Examples\r\n\r\nPractical examples of integrating Nexa SDK with AG2 (formerly AutoGen).\r\n\r\n\r\n## Prerequisites\r\n- Python 3.10+\r\n- [Nexa SDK installed](../../README.md#installation)\r\n\r\n## Quick Start\r\n\r\nSimple agent creation and multi-agent conversation with Nexa SDK.\r\n1. Run nexa serve in separate terminal.\r\n```bash\r\nnexa pull NexaAI/Qwen3-4B-GGUF\r\nnexa serve\r\n```\r\n2. Install dependencies.\r\n```\r\npip install -r requirements.txt\r\n```\r\n3. Run the example script.\r\n\r\n```bash\r\npython example.py\r\n```\r\n" }, { "path": "cookbook/ag2/example.py", "content": "#!/usr/bin/env python3\r\n\"\"\"\r\nBasic agent example with Nexa SDK and AG2.\r\n\r\nDemonstrates:\r\n- Configuring AG2 with Nexa SDK's OpenAI-compatible endpoint\r\n- Creating a single conversable agent\r\n- Running a two-agent conversation\r\n\"\"\"\r\n\r\nfrom autogen import ConversableAgent, LLMConfig\r\n\r\n\r\ndef create_llm_config():\r\n \"\"\"Create LLM config pointing to Nexa's OpenAI-compatible endpoint.\"\"\"\r\n return LLMConfig({\r\n \"model\": \"NexaAI/Qwen3-4B-GGUF\",\r\n \"base_url\": \"http://localhost:18181/v1\",\r\n \"api_type\": \"openai\",\r\n \"api_key\": \"not-needed\",\r\n })\r\n\r\n\r\ndef single_agent_example():\r\n \"\"\"Example of a single conversable agent.\"\"\"\r\n print(\"=\" * 50)\r\n print(\"Single Agent Example\")\r\n print(\"=\" * 50)\r\n\r\n llm_config = create_llm_config()\r\n\r\n agent = ConversableAgent(\r\n name=\"helpful_agent\",\r\n system_message=\"You are a helpful AI assistant that gives concise answers.\",\r\n llm_config=llm_config,\r\n )\r\n\r\n print(\"\\nRunning agent...\")\r\n response = agent.run(\r\n message=\"Write a Python function to calculate the factorial of a number.\",\r\n max_turns=3,\r\n )\r\n response.process()\r\n\r\n for msg in response.messages:\r\n print(f\"\\n[{msg.get('role', 'unknown')}]: {msg.get('content', '')}\")\r\n\r\n\r\ndef two_agent_example():\r\n \"\"\"Example of a two-agent conversation.\"\"\"\r\n print(\"\\n\" + \"=\" * 50)\r\n print(\"Two-Agent Conversation Example\")\r\n print(\"=\" * 50)\r\n\r\n llm_config = create_llm_config()\r\n\r\n # Create a coding assistant\r\n coder = ConversableAgent(\r\n name=\"coder\",\r\n system_message=(\r\n \"You are a Python developer. Write clean, well-documented code. \"\r\n \"Reply TERMINATE when the task is complete.\"\r\n ),\r\n llm_config=llm_config,\r\n )\r\n\r\n # Create a code reviewer\r\n reviewer = ConversableAgent(\r\n name=\"reviewer\",\r\n system_message=(\r\n \"You are a code reviewer. Review the code for correctness, style, \"\r\n \"and best practices. Reply TERMINATE when the review is done.\"\r\n ),\r\n llm_config=llm_config,\r\n )\r\n\r\n print(\"\\nStarting two-agent conversation...\")\r\n result = coder.initiate_chat(\r\n recipient=reviewer,\r\n message=\"Write a Python function to check if a string is a palindrome.\",\r\n max_turns=3,\r\n )\r\n\r\n print(f\"\\nConversation summary:\\n{result.summary}\")\r\n\r\n\r\ndef main():\r\n \"\"\"Run all agent examples.\"\"\"\r\n print(\"\\n\" + \"=\" * 50)\r\n print(\"AG2 + Nexa SDK Basic Agent Examples\")\r\n print(\"=\" * 50 + \"\\n\")\r\n\r\n single_agent_example()\r\n two_agent_example()\r\n\r\n print(\"\\n\" + \"=\" * 50)\r\n print(\"All examples completed!\")\r\n print(\"=\" * 50)\r\n\r\n\r\nif __name__ == \"__main__\":\r\n main()\r\n" }, { "path": "cookbook/ag2/requirements.txt", "content": "ag2[openai]>=0.8.0\r\n" }, { "path": "cookbook/android/RAG-LLM/.gitignore", "content": "# Gradle files\n.gradle/\nbuild/\narm64-v8a/\n.DS_Store\n# Local configuration file (sdk path, etc)\nlocal.properties\n\n# Log/OS Files\n*.log\n\n# Android Studio generated files and folders\ncaptures/\n.externalNativeBuild/\n.cxx/\n*.aab\n*.apk\noutput-metadata.json\n\n# IntelliJ\n*.iml\n.idea/\nmisc.xml\ndeploymentTargetDropDown.xml\nrender.experimental.xml\n\n# Keystore files\n*.jks\n*.keystore\n\n# Google Services (e.g. APIs or Firebase)\ngoogle-services.json\n\n# Android Profiling\n*.hprof\napp/libs/app-release.aar\napp/libs/app-debug.aar\napp/release/\nrepo/" }, { "path": "cookbook/android/RAG-LLM/README.md", "content": "# Nexa Android SDK Demo App \n\n## Overview\n\nThe Nexa AI Android SDK enables on-device AI inference for Android applications with NPU acceleration. Run Large Language Models (LLMs), Vision-Language Models (VLMs), Embeddings, Speech Recognition (ASR), Reranking, and Computer Vision models on Android devices with support for NPU, GPU, and CPU inference. Follow [Android SDK Doc](https://docs.nexa.ai/en/nexa-sdk-android/overview) for more details. This folder contains the demo app for the Android SDK.\n\n## Device Compatibility\n\n### Supported Hardware\n\n- **NPU**: Qualcomm Snapdragon 8 Gen 4 (optimized)\n- **GPU**: Qualcomm Adreno GPU\n- **CPU**: ARM64-v8a\n- **RAM**: 4GB+ recommended\n- **Storage**: 100MB - 4GB (varies by model)\n\n### Minimum Requirements\n\n- Android API Level 27+ (Android 8.1 Oreo)\n- **Architecture**: ARM64-v8a\n- **Android SDK Version**: 27+\n\n## Build and Run\n\n1. Clone the repository\n\n```bash\ngit clone https://github.com/NexaAI/nexa-sdk/\n```\n\n2. Open this folder `bindings/android` in Android Studio\n\n3. Follow the instructions in [Android SDK Doc](https://docs.nexa.ai/en/nexa-sdk-android/overview), download model (For example, [`OmniNeural-4B`](https://huggingface.co/NexaAI/OmniNeural-4B)) and put it in App's assets folder (For example, `/data/data/com.nexa.demo/files/models/omni-neural-4b `).\n\n4. Build and run the app" }, { "path": "cookbook/android/RAG-LLM/app/.gitignore", "content": "/build\n/src/main/assets/model_list_backup.json\n/src/main/assets/model_list_local.json" }, { "path": "cookbook/android/RAG-LLM/app/build.gradle.kts", "content": "// Copyright 2024-2026 Nexa AI, Inc.\n//\n// Licensed under the Apache License, Version 2.0 (the \"License\");\n// you may not use this file except in compliance with the License.\n// You may obtain a copy of the License at\n//\n// http://www.apache.org/licenses/LICENSE-2.0\n//\n// Unless required by applicable law or agreed to in writing, software\n// distributed under the License is distributed on an \"AS IS\" BASIS,\n// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n// See the License for the specific language governing permissions and\n// limitations under the License.\n\nplugins {\n alias(libs.plugins.android.application)\n alias(libs.plugins.kotlin.android)\n alias(libs.plugins.kotlin.compose)\n id(\"org.jetbrains.kotlin.plugin.serialization\") version \"1.9.23\"\n}\n\nandroid {\n namespace = \"com.nexa.demo\"\n compileSdk = 36\n\n signingConfigs {\n create(\"release\") {\n storeFile = file(\"test\")\n storePassword = \"123456\"\n keyAlias = \"test\"\n keyPassword = \"123456\"\n }\n }\n\n defaultConfig {\n applicationId = \"com.nexa.demo\"\n minSdk = 27\n targetSdk = 36\n versionCode = 1\n versionName = \"1.0\"\n\n testInstrumentationRunner = \"androidx.test.runner.AndroidJUnitRunner\"\n }\n\n buildTypes {\n release {\n isMinifyEnabled = false\n applicationIdSuffix = \".rag\"\n proguardFiles(\n getDefaultProguardFile(\"proguard-android-optimize.txt\"),\n \"proguard-rules.pro\"\n )\n signingConfig = signingConfigs.getByName(\"release\")\n }\n\n debug {\n isMinifyEnabled = false\n applicationIdSuffix = \".rag\"\n proguardFiles(\n getDefaultProguardFile(\"proguard-android-optimize.txt\"),\n \"proguard-rules.pro\"\n )\n signingConfig = signingConfigs.getByName(\"release\")\n }\n }\n compileOptions {\n sourceCompatibility = JavaVersion.VERSION_11\n targetCompatibility = JavaVersion.VERSION_11\n }\n kotlinOptions {\n jvmTarget = \"11\"\n }\n// sourceSets {\n// getByName(\"main\") {\n// jniLibs.srcDirs(\"src/main/jniLibs\")\n// }\n// }\n packagingOptions {\n jniLibs.useLegacyPackaging = true\n }\n\n buildFeatures {\n viewBinding = true\n dataBinding = true\n compose = true\n buildConfig = true\n }\n}\n\nval bridgePathExist = gradle.extra[\"bridgePathExist\"] as Boolean\nprint(\"bridgePathExist: $bridgePathExist\\n\")\n\ndependencies {\n\n // ===== NEXA CLOUD SDK =====\n implementation(\"ai.nexa:core:0.0.22\")\n // ===== NEXA CLOUD SDK END =====\n implementation(project(\":transform\"))\n implementation(\":okdownload-core@aar\")\n implementation(\":okdownload-sqlite@aar\")\n implementation(\":okdownload-okhttp@aar\")\n implementation(\":okdownload-ktx@aar\")\n implementation(kotlin(\"reflect\"))\n implementation(libs.glide)\n implementation(libs.gson)\n implementation(libs.markwon.core)\n implementation(libs.markwon.strikethrough)\n implementation(libs.markwon.tables)\n implementation(libs.markwon.linkify)\n implementation(libs.recyclerview)\n implementation(libs.toaster)\n implementation(libs.material)\n implementation(libs.imm.bar)\n implementation(libs.imm.bar.ktx)\n implementation(libs.auto.size)\n implementation(libs.okhttp)\n implementation(libs.kotlinx.serialization.json)\n implementation(libs.androidx.core.ktx)\n implementation(libs.androidx.lifecycle.runtime.ktx)\n implementation(libs.androidx.activity.compose)\n implementation(platform(libs.androidx.compose.bom))\n implementation(libs.androidx.ui)\n implementation(libs.androidx.ui.graphics)\n implementation(libs.androidx.ui.tooling.preview)\n implementation(libs.androidx.material3)\n implementation(libs.androidx.appcompat)\n implementation(libs.androidx.activity)\n implementation(libs.androidx.constraintlayout)\n testImplementation(libs.junit)\n androidTestImplementation(libs.androidx.junit)\n androidTestImplementation(libs.androidx.espresso.core)\n androidTestImplementation(platform(libs.androidx.compose.bom))\n androidTestImplementation(libs.androidx.ui.test.junit4)\n debugImplementation(libs.androidx.ui.tooling)\n debugImplementation(libs.androidx.ui.test.manifest)\n}" }, { "path": "cookbook/android/RAG-LLM/app/proguard-rules.pro", "content": "# Copyright 2024-2026 Nexa AI, Inc.\n#\n# Licensed under the Apache License, Version 2.0 (the \"License\");\n# you may not use this file except in compliance with the License.\n# You may obtain a copy of the License at\n#\n# http://www.apache.org/licenses/LICENSE-2.0\n#\n# Unless required by applicable law or agreed to in writing, software\n# distributed under the License is distributed on an \"AS IS\" BASIS,\n# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n# See the License for the specific language governing permissions and\n# limitations under the License.\n\n# Add project specific ProGuard rules here.\n# You can control the set of applied configuration files using the\n# proguardFiles setting in build.gradle.\n#\n# For more details, see\n# http://developer.android.com/guide/developing/tools/proguard.html\n\n# If your project uses WebView with JS, uncomment the following\n# and specify the fully qualified class name to the JavaScript interface\n# class:\n#-keepclassmembers class fqcn.of.javascript.interface.for.webview {\n# public *;\n#}\n\n# Uncomment this to preserve the line number information for\n# debugging stack traces.\n#-keepattributes SourceFile,LineNumberTable\n\n# If you keep the line number information, uncomment this to\n# hide the original source file name.\n#-renamesourcefileattribute SourceFile" }, { "path": "cookbook/android/RAG-LLM/app/src/androidTest/java/com/nexa/demo/ExampleInstrumentedTest.kt", "content": "// Copyright 2024-2026 Nexa AI, Inc.\n//\n// Licensed under the Apache License, Version 2.0 (the \"License\");\n// you may not use this file except in compliance with the License.\n// You may obtain a copy of the License at\n//\n// http://www.apache.org/licenses/LICENSE-2.0\n//\n// Unless required by applicable law or agreed to in writing, software\n// distributed under the License is distributed on an \"AS IS\" BASIS,\n// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n// See the License for the specific language governing permissions and\n// limitations under the License.\n\npackage com.nexa.demo\n\nimport androidx.test.platform.app.InstrumentationRegistry\nimport androidx.test.ext.junit.runners.AndroidJUnit4\n\nimport org.junit.Test\nimport org.junit.runner.RunWith\n\nimport org.junit.Assert.*\n\n/**\n * Instrumented test, which will execute on an Android device.\n *\n * See [testing documentation](http://d.android.com/tools/testing).\n */\n@RunWith(AndroidJUnit4::class)\nclass ExampleInstrumentedTest {\n @Test\n fun useAppContext() {\n // Context of the app under test.\n val appContext = InstrumentationRegistry.getInstrumentation().targetContext\n assertEquals(\"com.nexa.demo\", appContext.packageName)\n }\n}" }, { "path": "cookbook/android/RAG-LLM/app/src/main/AndroidManifest.xml", "content": "\n\n\n\n \n \n \n \n \n \n\n \n \n \n\n \n \n\n \n \n \n\n \n \n \n\n \n \n \n \n\n" }, { "path": "cookbook/android/RAG-LLM/app/src/main/assets/model_list.json", "content": "[\n {\n \"id\": \"Llama3.2-3B-NPU-Turbo-NPU-mobile\",\n \"displayName\": \"Llama3.2-3B\",\n \"modelName\": \"files-1-2.nexa\",\n \"versionCode\": 1,\n \"pluginIds\": 273,\n \"dependencies\": [\"Llama-3.2-3B-Instruct-Q4_0\"],\n \"baseUrl\": \"https://nexa-model-hub-bucket.s3.us-west-1.amazonaws.com/public/nexa_sdk/huggingface-models/Llama3.2-3B-NPU-Turbo-NPU-mobile/\",\n \"modelUrl\": \"files-1-2.nexa\"\n },\n {\n \"id\": \"Llama-3.2-3B-Instruct-Q4_0\",\n \"displayName\": \"Llama3.2-3B Q40 GGUF\",\n \"modelName\": \"Llama-3.2-3B-Instruct-Q4_0.gguf\",\n \"type\": \"chat\",\n \"versionCode\": 1,\n \"pluginIds\": 17,\n \"show\": false,\n \"dependencies\": [\"embeddinggemma-300m-npu-mobile\"],\n \"modelUrl\": \"https://nexa-model-hub-bucket.s3.us-west-1.amazonaws.com/public/nexa_sdk/huggingface-models/Llama-3.2-3B-Instruct-GGUF/Llama-3.2-3B-Instruct-Q4_0.gguf\"\n },\n {\n \"id\": \"Qwen3-4B-Instruct-2507-npu\",\n \"displayName\": \"Qwen3-4B\",\n \"modelName\": \"files-1-2.nexa\",\n \"versionCode\": 1,\n \"pluginIds\": 273,\n \"dependencies\": [\"Qwen3-4B-Q4_0\"],\n \"baseUrl\": \"https://nexa-model-hub-bucket.s3.us-west-1.amazonaws.com/public/nexa_sdk/huggingface-models/Qwen3-4B-Instruct-2507-npu-mobile/\",\n \"modelUrl\": \"files-1-2.nexa\"\n },\n {\n \"id\": \"Qwen3-4B-Q4_0\",\n \"displayName\": \"Qwen3-4B Q40 GGUF\",\n \"modelName\": \"Qwen3-4B-Q4_0.gguf\",\n \"type\": \"chat\",\n \"versionCode\": 1,\n \"pluginIds\": 17,\n \"show\": false,\n \"dependencies\": [\"embeddinggemma-300m-npu-mobile\"],\n \"modelUrl\": \"https://nexa-model-hub-bucket.s3.us-west-1.amazonaws.com/public/nexa_sdk/huggingface-models/Qwen3-4B-GGUF/Qwen3-4B-Q4_0.gguf\"\n },\n {\n \"id\": \"Granite-4-Micro-NPU-mobile\",\n \"displayName\": \"Granite4-Micro(3B)\",\n \"modelName\": \"files-1-2.nexa\",\n \"versionCode\": 1,\n \"pluginIds\": 273,\n \"dependencies\": [\"granite-4.0-micro-Q4_0\"],\n \"baseUrl\": \"https://nexa-model-hub-bucket.s3.us-west-1.amazonaws.com/public/nexa_sdk/huggingface-models/Granite-4-Micro-NPU-mobile/\",\n \"modelUrl\": \"files-1-2.nexa\"\n },\n {\n \"id\": \"granite-4.0-micro-Q4_0\",\n \"displayName\": \"Granite4-Micro(3B) Q40 GGUF\",\n \"modelName\": \"granite-4.0-micro-Q4_0.gguf\",\n \"type\": \"chat\",\n \"versionCode\": 1,\n \"show\": false,\n \"dependencies\": [\"embeddinggemma-300m-npu-mobile\"],\n \"modelUrl\": \"https://nexa-model-hub-bucket.s3.us-west-1.amazonaws.com/public/nexa_sdk/huggingface-models/granite-4.0-micro-GGUF/granite-4.0-micro-Q4_0.gguf\"\n },\n {\n \"id\": \"Granite-4.0-h-350M-NPU-mobile\",\n \"displayName\": \"Granite4-350M\",\n \"modelName\": \"files-1-2.nexa\",\n \"versionCode\": 1,\n \"pluginIds\": 273,\n \"dependencies\": [\"granite-4.0-350m-Q4_0\"],\n \"baseUrl\": \"https://nexa-model-hub-bucket.s3.us-west-1.amazonaws.com/public/nexa_sdk/huggingface-models/Granite-4.0-h-350M-NPU-mobile/\",\n \"modelUrl\": \"files-1-2.nexa\"\n },\n {\n \"id\": \"granite-4.0-350m-Q4_0\",\n \"displayName\": \"Granite4-350M Q40 GGUF\",\n \"modelName\": \"granite-4.0-350m-Q4_0.gguf\",\n \"type\": \"chat\",\n \"versionCode\": 1,\n \"show\": false,\n \"dependencies\": [\"embeddinggemma-300m-npu-mobile\"],\n \"modelUrl\": \"https://nexa-model-hub-bucket.s3.us-west-1.amazonaws.com/public/nexa_sdk/huggingface-models/granite-4.0-350m-GGUF/granite-4.0-350m-Q4_0.gguf\"\n },\n {\n \"id\": \"LFM2-1.2B-npu-mobile\",\n \"displayName\": \"LFM2-1.2B\",\n \"modelName\": \"files-1-2.nexa\",\n \"versionCode\": 1,\n \"pluginIds\": 273,\n \"dependencies\": [\"LFM2-1.2B-Q4_0\"],\n \"baseUrl\": \"https://nexa-model-hub-bucket.s3.us-west-1.amazonaws.com/public/nexa_sdk/huggingface-models/LFM2-1.2B-npu-mobile/\",\n \"modelUrl\": \"files-1-2.nexa\"\n },\n {\n \"id\": \"LFM2-1.2B-Q4_0\",\n \"displayName\": \"LFM2-1.2B Q40 GGUF\",\n \"modelName\": \"LFM2-1.2B-Q4_0.gguf\",\n \"type\": \"chat\",\n \"versionCode\": 1,\n \"pluginIds\": 17,\n \"show\": false,\n \"dependencies\": [\"embeddinggemma-300m-npu-mobile\"],\n \"modelUrl\": \"https://nexa-model-hub-bucket.s3.us-west-1.amazonaws.com/public/nexa_sdk/huggingface-models/LFM2-1.2B-GGUF/LFM2-1.2B-Q4_0.gguf\"\n },\n {\n \"id\": \"embeddinggemma-300m-npu-mobile\",\n \"displayName\": \"embed-gemma\",\n \"modelName\": \"files-1-2.nexa\",\n \"versionCode\": 1,\n \"show\": false,\n \"baseUrl\": \"https://nexa-model-hub-bucket.s3.us-west-1.amazonaws.com/public/nexa_sdk/huggingface-models/embeddinggemma-300m-npu-mobile/\",\n \"modelUrl\": \"files-1-2.nexa\"\n }\n]" }, { "path": "cookbook/android/RAG-LLM/app/src/main/java/com/nexa/demo/ChatAdapter.kt", "content": "// Copyright 2024-2026 Nexa AI, Inc.\n//\n// Licensed under the Apache License, Version 2.0 (the \"License\");\n// you may not use this file except in compliance with the License.\n// You may obtain a copy of the License at\n//\n// http://www.apache.org/licenses/LICENSE-2.0\n//\n// Unless required by applicable law or agreed to in writing, software\n// distributed under the License is distributed on an \"AS IS\" BASIS,\n// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n// See the License for the specific language governing permissions and\n// limitations under the License.\n\npackage com.nexa.demo\n\nimport android.content.Intent\nimport android.net.Uri\nimport android.text.TextUtils\nimport android.text.method.LinkMovementMethod\nimport android.view.LayoutInflater\nimport android.view.View\nimport android.view.ViewGroup\nimport android.widget.Button\nimport android.widget.ImageView\nimport android.widget.LinearLayout\nimport android.widget.TextView\nimport androidx.recyclerview.widget.RecyclerView\nimport com.nexa.demo.activity.FileContentActivity\nimport com.nexa.demo.bean.EmbedResultBean\nimport io.noties.markwon.Markwon\nimport io.noties.markwon.ext.strikethrough.StrikethroughPlugin\nimport io.noties.markwon.ext.tables.TablePlugin\nimport io.noties.markwon.linkify.LinkifyPlugin\nimport java.io.File\n\n\ndata class Message(\n val content: String,\n val type: MessageType,\n val images: List = emptyList(),\n val audio: List = emptyList(),\n val embedResultBean: EmbedResultBean? = null, // Deprecated, keeping for compatibility\n val retrievedChunks: List = emptyList(), // Store multiple retrieved chunks\n val formattedPrompt: String? = null // Store the full LLM prompt\n)\n\nenum class MessageType(val value: Int) {\n USER(0),\n ASSISTANT(1),\n PROFILE(2),\n IMAGES(3);\n\n companion object {\n fun from(value: Int): MessageType =\n entries.firstOrNull { it.value == value } ?: PROFILE\n }\n}\n\ninterface OnCitationsClick {\n fun onClick(position: Int, retrievedChunks: List)\n}\n\nclass ChatAdapter(private val messages: List, private val onCitationsClick: OnCitationsClick) :\n RecyclerView.Adapter() {\n\n override fun getItemViewType(position: Int): Int {\n val message = messages[position]\n return message.type.value\n }\n\n override fun onCreateViewHolder(parent: ViewGroup, viewType: Int): RecyclerView.ViewHolder {\n val inflater = LayoutInflater.from(parent.context)\n val type = MessageType.from(viewType)\n return if (type == MessageType.USER) {\n UserViewHolder(inflater.inflate(R.layout.item_user_message, parent, false))\n } else if (type == MessageType.ASSISTANT) {\n AiViewHolder(inflater.inflate(R.layout.item_ai_message, parent, false))\n } else if (type == MessageType.IMAGES) {\n ImagesViewHolder(inflater.inflate(R.layout.item_image_message, parent, false))\n } else {\n ProfileViewHolder(inflater.inflate(R.layout.item_profile_message, parent, false))\n }\n }\n\n override fun onBindViewHolder(holder: RecyclerView.ViewHolder, position: Int) {\n val message = messages[position]\n if (holder is UserViewHolder) holder.bind(message)\n if (holder is AiViewHolder) holder.bind(message)\n if (holder is ImagesViewHolder) holder.bind(message)\n if (holder is ProfileViewHolder) holder.bind(position, message, onCitationsClick)\n }\n\n override fun getItemCount() = messages.size\n\n class UserViewHolder(itemView: View) : RecyclerView.ViewHolder(itemView) {\n private val tvMessage: TextView = itemView.findViewById(R.id.tv_message)\n fun bind(message: Message) {\n tvMessage.text = message.content\n }\n }\n\n class AiViewHolder(itemView: View) : RecyclerView.ViewHolder(itemView) {\n private val tvMessage: TextView = itemView.findViewById(R.id.tv_message)\n private val markwon: Markwon = Markwon.builder(itemView.context)\n .usePlugin(StrikethroughPlugin.create())\n .usePlugin(TablePlugin.create(itemView.context))\n .usePlugin(LinkifyPlugin.create())\n .build()\n\n fun bind(message: Message) {\n markwon.setMarkdown(tvMessage, message.content)\n tvMessage.movementMethod = LinkMovementMethod.getInstance()\n }\n }\n\n class ProfileViewHolder(itemView: View) : RecyclerView.ViewHolder(itemView) {\n private val tvMessage: TextView = itemView.findViewById(R.id.tv_message)\n private val btnCitations: Button = itemView.findViewById(R.id.btn_citations)\n private val btnViewPrompt: Button = itemView.findViewById(R.id.btn_view_prompt)\n \n fun bind(position: Int, message: Message, onCitationsClick: OnCitationsClick) {\n tvMessage.text = message.content\n \n // Handle multiple citations\n if (message.retrievedChunks.isNotEmpty()) {\n btnCitations.text = \"${message.retrievedChunks.size} Citations\"\n btnCitations.visibility = View.VISIBLE\n btnCitations.setOnClickListener {\n onCitationsClick.onClick(position, message.retrievedChunks)\n }\n } else {\n btnCitations.visibility = View.GONE\n }\n \n // Handle view prompt button\n if (!TextUtils.isEmpty(message.formattedPrompt)) {\n btnViewPrompt.setOnClickListener {\n it.context.startActivity(\n Intent(\n it.context,\n FileContentActivity::class.java\n ).apply {\n this.putExtra(\n FileContentActivity.KEY_PROMPT_CONTENT,\n message.formattedPrompt\n )\n })\n }\n btnViewPrompt.visibility = View.VISIBLE\n } else {\n btnViewPrompt.visibility = View.GONE\n }\n }\n \n private fun dpToPx(dp: Int, context: android.content.Context): Int {\n return (dp * context.resources.displayMetrics.density).toInt()\n }\n }\n\n class ImagesViewHolder(itemView: View) : RecyclerView.ViewHolder(itemView) {\n private val imageContainer: LinearLayout = itemView.findViewById(R.id.image_container)\n fun bind(message: Message) {\n val savedImageFiles = message.images\n imageContainer.removeAllViews()\n val context = itemView.context\n\n for (file in savedImageFiles) {\n val itemView = LayoutInflater.from(context)\n .inflate(R.layout.item_image_item_message, imageContainer, false)\n val ivImage = itemView.findViewById(R.id.iv_image)\n ivImage.setImageURI(Uri.fromFile(file))\n imageContainer.addView(itemView)\n }\n }\n }\n}" }, { "path": "cookbook/android/RAG-LLM/app/src/main/java/com/nexa/demo/FileConfig.kt", "content": "// Copyright 2024-2026 Nexa AI, Inc.\n//\n// Licensed under the Apache License, Version 2.0 (the \"License\");\n// you may not use this file except in compliance with the License.\n// You may obtain a copy of the License at\n//\n// http://www.apache.org/licenses/LICENSE-2.0\n//\n// Unless required by applicable law or agreed to in writing, software\n// distributed under the License is distributed on an \"AS IS\" BASIS,\n// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n// See the License for the specific language governing permissions and\n// limitations under the License.\n\npackage com.nexa.demo\n\nimport android.content.Context\nimport java.io.File\n\nclass FileConfig {\n companion object {\n val modelsDir = fun(context: Context): File {\n return File(context.filesDir, \"models\").apply { if (!exists()) mkdirs() }\n }\n }\n}" }, { "path": "cookbook/android/RAG-LLM/app/src/main/java/com/nexa/demo/GenerationConfigSample.kt", "content": "// Copyright 2024-2026 Nexa AI, Inc.\n//\n// Licensed under the Apache License, Version 2.0 (the \"License\");\n// you may not use this file except in compliance with the License.\n// You may obtain a copy of the License at\n//\n// http://www.apache.org/licenses/LICENSE-2.0\n//\n// Unless required by applicable law or agreed to in writing, software\n// distributed under the License is distributed on an \"AS IS\" BASIS,\n// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n// See the License for the specific language governing permissions and\n// limitations under the License.\n\npackage com.nexa.demo\n\nimport com.nexa.sdk.bean.GenerationConfig\nimport com.nexa.sdk.bean.SamplerConfig\n\n// Configuration sample for generation with defaults compatible with bridge\n// maxTokens: 0 = no limit, generates until model's natural stopping point\ndata class GenerationConfigSample(\n var maxTokens: Int = 2048,\n var stopWords: List? = null,\n var stopCount: Int = 0,\n var nPast: Int = 0,\n var imagePaths: List? = null,\n var imageCount: Int = 0,\n var audioPaths: List? = null,\n var audioCount: Int = 0,\n var systemPrompt: String = \"\"\n) {\n // Convert to GenerationConfig with minimal sampler setup for bridge compatibility\n // Sampler config uses bridge defaults (no custom parameters applied)\n fun toGenerationConfig(grammarString:String? = null): GenerationConfig {\n return GenerationConfig(\n maxTokens = this.maxTokens,\n stopWords = this.stopWords?.toTypedArray(),\n stopCount = this.stopCount,\n nPast = this.nPast,\n //samplerConfig = SamplerConfig(\n // grammarString = grammarString\n // All other sampler parameters use bridge defaults\n // No temperature, topK, topP, penalties applied\n //),\n imagePaths = this.imagePaths?.toTypedArray(),\n imageCount = this.imageCount,\n audioPaths = this.audioPaths?.toTypedArray(),\n audioCount = this.audioCount\n )\n }\n}" }, { "path": "cookbook/android/RAG-LLM/app/src/main/java/com/nexa/demo/MainActivity.kt", "content": "// Copyright 2024-2026 Nexa AI, Inc.\n//\n// Licensed under the Apache License, Version 2.0 (the \"License\");\n// you may not use this file except in compliance with the License.\n// You may obtain a copy of the License at\n//\n// http://www.apache.org/licenses/LICENSE-2.0\n//\n// Unless required by applicable law or agreed to in writing, software\n// distributed under the License is distributed on an \"AS IS\" BASIS,\n// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n// See the License for the specific language governing permissions and\n// limitations under the License.\n\npackage com.nexa.demo\n\nimport android.Manifest\nimport android.app.Activity\nimport android.content.Context\nimport android.content.DialogInterface\nimport android.content.Intent\nimport android.content.SharedPreferences\nimport android.content.pm.PackageManager\nimport android.graphics.Bitmap\nimport android.graphics.BitmapFactory\nimport android.graphics.Color\nimport android.net.Uri\nimport android.os.Bundle\nimport android.os.Environment\nimport android.provider.MediaStore\nimport android.system.Os\nimport android.text.Editable\nimport android.text.TextUtils\nimport android.text.TextWatcher\nimport android.util.Log\nimport android.view.KeyEvent\nimport android.view.LayoutInflater\nimport android.view.View\nimport android.view.inputmethod.InputMethodManager\nimport android.widget.AdapterView\nimport android.widget.Button\nimport android.widget.EditText\nimport android.widget.HorizontalScrollView\nimport android.widget.ImageButton\nimport android.widget.ImageView\nimport android.widget.LinearLayout\nimport android.widget.PopupWindow\nimport android.widget.ProgressBar\nimport android.widget.SeekBar\nimport android.widget.BaseAdapter\nimport android.widget.SimpleAdapter\nimport android.widget.Spinner\nimport android.widget.TextView\nimport android.widget.Toast\nimport android.window.OnBackInvokedDispatcher\nimport androidx.activity.ComponentActivity\nimport androidx.activity.OnBackPressedCallback\nimport androidx.activity.result.ActivityResultLauncher\nimport androidx.activity.result.contract.ActivityResultContracts\nimport androidx.appcompat.app.AlertDialog\nimport androidx.compose.ui.unit.TextUnit\nimport androidx.core.app.ActivityCompat\nimport androidx.core.content.ContextCompat\nimport androidx.core.content.FileProvider\nimport androidx.core.view.ViewCompat\nimport androidx.core.view.WindowInsetsCompat\nimport androidx.fragment.app.Fragment\nimport androidx.fragment.app.FragmentActivity\nimport androidx.recyclerview.widget.RecyclerView\nimport com.google.android.material.bottomsheet.BottomSheetDialog\nimport com.gyf.immersionbar.ktx.immersionBar\nimport com.hjq.toast.Toaster\nimport com.liulishuo.okdownload.DownloadContext\nimport com.liulishuo.okdownload.DownloadTask\nimport com.liulishuo.okdownload.OkDownload\nimport com.liulishuo.okdownload.core.cause.EndCause\nimport com.liulishuo.okdownload.core.connection.DownloadOkHttp3Connection\nimport com.liulishuo.okdownload.kotlin.listener.createDownloadContextListener\nimport com.liulishuo.okdownload.kotlin.listener.createListener1\nimport com.nexa.demo.activity.FolderActivity\nimport com.nexa.demo.adapter.ChunkAdapter\nimport com.nexa.demo.bean.DownloadState\nimport com.nexa.demo.bean.EmbedResultBean\nimport com.nexa.demo.bean.ModelData\nimport com.nexa.demo.bean.DownloadableFileWithFallback\nimport com.nexa.demo.bean.downloadableFiles\nimport com.nexa.demo.bean.downloadableFilesWithFallback\nimport com.nexa.demo.bean.downloadableFilesWithNpuList\nimport com.nexa.demo.bean.getNexaManifest\nimport com.nexa.demo.bean.getNonExistModelFile\nimport com.nexa.demo.bean.getSupportPluginIds\nimport com.nexa.demo.bean.allModelFilesExist\nimport com.nexa.demo.bean.isDownloaded\nimport com.nexa.demo.bean.isNpuModel\nimport com.nexa.demo.bean.withFallbackUrls\nimport com.nexa.demo.utils.ModelFileListingUtil\nimport com.nexa.demo.bean.mmprojTokenFile\nimport com.nexa.demo.bean.modelDir\nimport com.nexa.demo.bean.modelFile\nimport com.nexa.demo.bean.tokenFile\nimport com.nexa.demo.databinding.ActivityMainBinding\nimport com.nexa.demo.databinding.DialogConfigBinding\nimport com.nexa.demo.databinding.DialogSelectPluginIdBinding\nimport com.nexa.demo.databinding.DialogTopkConfigBinding\nimport com.nexa.demo.fragments.IndexFragment\nimport com.nexa.demo.listeners.CustomDialogInterface\nimport com.nexa.demo.utils.DensityUtil\nimport com.nexa.demo.utils.ExecShell\nimport com.nexa.demo.utils.GenerateEmbedStringsUtil\nimport com.nexa.demo.utils.ImgUtil\nimport com.nexa.demo.utils.PermissionUtil\nimport com.nexa.demo.utils.SharePreferenceKeys\nimport com.nexa.demo.utils.WavRecorder\nimport com.nexa.demo.utils.inflate\nimport com.nexa.sdk.AsrWrapper\nimport com.nexa.sdk.CvWrapper\nimport com.nexa.sdk.EmbedderWrapper\nimport com.nexa.sdk.LlmWrapper\nimport com.nexa.sdk.NexaSdk\nimport com.nexa.sdk.RerankerWrapper\nimport com.nexa.sdk.VlmWrapper\nimport com.nexa.sdk.bean.AsrCreateInput\nimport com.nexa.sdk.bean.AsrTranscribeInput\nimport com.nexa.sdk.bean.CVCapability\nimport com.nexa.sdk.bean.CVCreateInput\nimport com.nexa.sdk.bean.CVModelConfig\nimport com.nexa.sdk.bean.ChatMessage\nimport com.nexa.sdk.bean.DeviceIdValue\nimport com.nexa.sdk.bean.EmbedderCreateInput\nimport com.nexa.sdk.bean.EmbeddingConfig\nimport com.nexa.sdk.bean.LlmCreateInput\nimport com.nexa.sdk.bean.LlmStreamResult\nimport com.nexa.sdk.bean.ModelConfig\nimport com.nexa.sdk.bean.RerankConfig\nimport com.nexa.sdk.bean.RerankerCreateInput\nimport com.nexa.sdk.bean.VlmChatMessage\nimport com.nexa.sdk.bean.VlmContent\nimport com.nexa.sdk.bean.VlmCreateInput\nimport kotlinx.coroutines.CoroutineScope\nimport kotlinx.coroutines.Dispatchers\nimport kotlinx.coroutines.async\nimport kotlinx.coroutines.awaitAll\nimport kotlinx.coroutines.launch\nimport kotlinx.serialization.json.Json\nimport okhttp3.OkHttpClient\nimport okhttp3.Request\nimport java.io.File\nimport java.io.FileNotFoundException\nimport java.io.FileOutputStream\nimport java.security.SecureRandom\nimport java.security.cert.CertificateException\nimport java.security.cert.X509Certificate\nimport javax.net.ssl.SSLContext\nimport javax.net.ssl.SSLSession\nimport javax.net.ssl.SSLSocketFactory\nimport javax.net.ssl.TrustManager\nimport javax.net.ssl.X509TrustManager\nimport androidx.core.content.edit\nimport com.nexa.demo.bean.DownloadableFile\nimport kotlinx.serialization.StringFormat\n\nclass MainActivity : FragmentActivity() {\n\n private val binding: ActivityMainBinding by inflate()\n private var downloadContext: DownloadContext? = null\n private var downloadState = DownloadState.IDLE\n private var downloadingModelData: ModelData? = null\n private lateinit var spDownloaded: SharedPreferences\n private lateinit var llDownloading: LinearLayout\n private lateinit var tvDownloadProgress: TextView\n private lateinit var pbDownloading: ProgressBar\n // Inline download progress views\n private lateinit var llControlButtonsRow: LinearLayout\n private lateinit var llDownloadProgress: LinearLayout\n private lateinit var pbDownloadProgress: ProgressBar\n private lateinit var tvDownloadPercent: TextView\n private lateinit var tvDownloadFailed: TextView\n private lateinit var btnCancelDownloadInline: Button\n private lateinit var btnRetryDownloadInline: Button\n private var downloadFailed = false\n private lateinit var spModelList: Spinner\n private lateinit var btnDownload: Button\n private lateinit var btnLoadModel: Button\n private lateinit var btnUnloadModel: Button\n private lateinit var btnStop: Button\n private lateinit var etInput: EditText\n private lateinit var btnSend: Button\n private lateinit var btnClearHistory: Button\n private lateinit var btnAddImage: Button\n private lateinit var btnAudioRecord: Button\n\n private lateinit var recyclerView: RecyclerView\n private lateinit var adapter: ChatAdapter\n private lateinit var bottomPanel: LinearLayout\n private lateinit var btnAudioDone: Button\n private lateinit var btnAudioCancel: Button\n\n private lateinit var scrollImages: HorizontalScrollView\n private lateinit var topScrollContainer: LinearLayout\n private lateinit var llLoading: LinearLayout\n private lateinit var vTip: View\n\n private lateinit var llmWrapper: LlmWrapper\n private lateinit var vlmWrapper: VlmWrapper\n var embedderWrapper: EmbedderWrapper? = null\n private lateinit var rerankerWrapper: RerankerWrapper\n private lateinit var cvWrapper: CvWrapper\n private lateinit var asrWrapper: AsrWrapper\n private val modelScope = CoroutineScope(Dispatchers.IO)\n\n private val chatList = arrayListOf()\n private lateinit var llmSystemPrompt: ChatMessage\n private val vlmChatList = arrayListOf()\n private lateinit var vlmSystemPrompty: VlmChatMessage\n private lateinit var modelList: List\n private var selectModelId = \"\"\n private var spinnerText = \"\"\n\n // ADD: Track which model type is loaded\n private var isLoadLlmModel = false\n private var isLoadVlmModel = false\n private var isLoadEmbedderModel = false\n private var isLoadRerankerModel = false\n private var isLoadCVModel = false\n private var isLoadAsrModel = false\n\n private var enableThinking = false\n\n private var wavRecorder: WavRecorder? = null\n private var audioFile: File? = null\n\n private var lastFormattedPrompt: String? = null // Store the last formatted prompt\n private var retrievedChunksList = mutableListOf() // Store retrieved chunks\n\n private val savedImageFiles = mutableListOf()\n private val messages = arrayListOf()\n\n private val embedMsgList = arrayListOf()\n private val embedResultList = arrayListOf()\n private lateinit var selectFolderResult: ActivityResultLauncher\n\n private lateinit var chunkAdapter: ChunkAdapter\n\n override fun onCreate(savedInstanceState: Bundle?) {\n super.onCreate(savedInstanceState)\n immersionBar {\n transparentStatusBar()\n statusBarDarkFont(false)\n }\n requestPermissions(arrayOf(Manifest.permission.RECORD_AUDIO), 1002)\n okdownload()\n initData()\n initView()\n setListeners()\n }\n\n private val embedStringPrefix =\n \"You are a careful assistant. Use ONLY the provided context to answer.\"\n\n private fun createEmbedChatMessage(embedString: String): ChatMessage {\n val embedString =\n \"$embedStringPrefix\\n\\n\\\\n$embedString\\\\n\"\n return ChatMessage(role = \"system\", embedString)\n }\n\n private fun addEmbedChatMessage(embedString: String, msgList: ArrayList) {\n createEmbedChatMessage(embedString).let { msg ->\n var hasAdded = false\n embedMsgList.forEach {\n if (it.content == msg.content) {\n hasAdded = true\n return@forEach\n }\n }\n if (!hasAdded) {\n embedMsgList.add(msg)\n }\n }\n embedMsgList.forEach {\n if (!msgList.contains(it)) {\n msgList.add(it)\n }\n }\n }\n\n // Format file list with chunk counts for display\n private fun formatFilesInSearch(): String {\n if (embedResultList.isEmpty()) {\n return \"Files: 0 | Chunks: 0\"\n }\n\n // Group by file path and count chunks\n val fileChunkCounts = embedResultList.groupBy { it.path }\n .mapValues { it.value.size }\n .toList()\n .sortedBy { File(it.first).name }\n\n val fileCount = fileChunkCounts.size\n val totalChunks = embedResultList.size\n\n return \"Files: $fileCount | Chunks: $totalChunks\"\n }\n\n private fun resetLoadState() {\n isLoadLlmModel = false\n isLoadVlmModel = false\n isLoadEmbedderModel = false\n isLoadRerankerModel = false\n isLoadCVModel = false\n isLoadAsrModel = false\n }\n\n private fun initView() {\n chunkAdapter = ChunkAdapter()\n binding.rvCitation.adapter = chunkAdapter\n adapter = ChatAdapter(messages, object : OnCitationsClick {\n override fun onClick(position: Int, retrievedChunks: List) {\n chunkAdapter.updateData(retrievedChunks)\n binding.llCitations.visibility = View.VISIBLE\n }\n })\n binding.rvChat.adapter = adapter\n\n // Set status bar padding for immersive status bar\n val headerView = findViewById(R.id.header_view)\n headerView?.let { view ->\n ViewCompat.setOnApplyWindowInsetsListener(view) { v, insets ->\n val statusBarHeight = insets.getInsets(WindowInsetsCompat.Type.statusBars()).top\n v.setPadding(0, statusBarHeight, 0, 0)\n insets\n }\n }\n\n // When keyboard is visible: only IME padding so input sits flush with keyboard.\n // When keyboard is hidden: only nav bar padding so buttons are not covered.\n ViewCompat.setOnApplyWindowInsetsListener(binding.root) { v, insets ->\n val navBottom = insets.getInsets(WindowInsetsCompat.Type.navigationBars()).bottom\n val imeBottom = insets.getInsets(WindowInsetsCompat.Type.ime()).bottom\n val bottomPadding = if (imeBottom > 0) imeBottom else navBottom\n v.setPadding(v.paddingLeft, v.paddingTop, v.paddingRight, bottomPadding)\n insets\n }\n\n llDownloading = findViewById(R.id.ll_downloading)\n tvDownloadProgress = findViewById(R.id.tv_download_progress)\n pbDownloading = findViewById(R.id.pb_downloading)\n // Initialize inline download progress views\n llControlButtonsRow = findViewById(R.id.ll_control_buttons_row)\n llDownloadProgress = findViewById(R.id.ll_download_progress)\n pbDownloadProgress = findViewById(R.id.pb_download_progress)\n tvDownloadPercent = findViewById(R.id.tv_download_percent)\n tvDownloadFailed = findViewById(R.id.tv_download_failed)\n btnCancelDownloadInline = findViewById(R.id.btn_cancel_download_inline)\n btnRetryDownloadInline = findViewById(R.id.btn_retry_download_inline)\n spModelList = findViewById(R.id.sp_model_list)\n spModelList.dropDownVerticalOffset = DensityUtil.dpToPx(this, 40f)\n spModelList.post {\n // Set dropdown width to match spinner width for symmetric padding\n spModelList.dropDownWidth = spModelList.width\n }\n val spinnerData = modelList.filter { it.show }.map {\n val map = mutableMapOf()\n map[\"modelId\"] = it.id\n map[\"displayName\"] = it.displayName\n map\n }\n // Set custom view for spinner selected item\n spModelList.adapter = object : BaseAdapter() {\n override fun getCount(): Int = spinnerData.size\n\n override fun getItem(position: Int): Any = spinnerData[position]\n\n override fun getItemId(position: Int): Long = position.toLong()\n\n override fun getView(position: Int, convertView: View?, parent: android.view.ViewGroup?): View {\n val view = convertView ?: LayoutInflater.from(this@MainActivity)\n .inflate(R.layout.spinner_custom, parent, false)\n val textView = view.findViewById(android.R.id.text1)\n textView.text = spinnerData[position][\"displayName\"]\n return view\n }\n\n override fun getDropDownView(position: Int, convertView: View?, parent: android.view.ViewGroup?): View {\n val view = convertView ?: LayoutInflater.from(this@MainActivity)\n .inflate(R.layout.item_spinner_dropdown, parent, false)\n val textView = view.findViewById(android.R.id.text1)\n val checkIcon = view.findViewById(R.id.iv_check)\n textView.text = spinnerData[position][\"displayName\"]\n // Show check icon for selected item\n val isSelected = position == spModelList.selectedItemPosition\n checkIcon.visibility = if (isSelected) View.VISIBLE else View.GONE\n return view\n }\n }\n spModelList.onItemSelectedListener = object : AdapterView.OnItemSelectedListener {\n override fun onItemSelected(\n parent: AdapterView<*>?, view: View?, position: Int, id: Long\n ) {\n selectModelId = spinnerData[position].get(\"modelId\") ?: \"\"\n spinnerText = spinnerData[position].get(\"displayName\") ?: \"\"\n Log.d(TAG, \"spinnerText:$spinnerText\")\n // Use actual file existence check (not just SharedPreferences)\n val selectedModel = modelList.first { it.id == selectModelId }\n val filesExist = isModelDownloaded(selectedModel) == null\n // Check if this model is currently downloading\n val isDownloading = downloadState == DownloadState.DOWNLOADING && \n downloadingModelData?.displayName == spinnerText\n if (isDownloading && !filesExist) {\n // Show progress row for current downloading model\n llControlButtonsRow.visibility = View.GONE\n llDownloadProgress.visibility = View.VISIBLE\n } else {\n changeOperationUI(\n if (filesExist) {\n OperationState.DOWNLOADED\n } else {\n OperationState.DEFAULT\n }\n )\n }\n messages.clear()\n adapter.notifyDataSetChanged()\n binding.rvChat.scrollTo(0, 0)\n }\n\n override fun onNothingSelected(parent: AdapterView<*>?) {\n selectModelId = \"\"\n }\n }\n btnDownload = findViewById(R.id.btn_download)\n btnLoadModel = findViewById(R.id.btn_load_model)\n btnUnloadModel = findViewById(R.id.btn_unload_model)\n btnStop = findViewById(R.id.btn_stop)\n etInput = findViewById(R.id.et_input)\n btnAddImage = findViewById(R.id.btn_add_image)\n btnAudioRecord = findViewById(R.id.btn_voice)\n\n bottomPanel = findViewById(R.id.bottom_panel)\n btnAudioCancel = findViewById(R.id.btn_audio_cancel)\n btnAudioDone = findViewById(R.id.btn_audio_done)\n\n btnSend = findViewById(R.id.btn_send)\n btnClearHistory = findViewById(R.id.btn_clear_history)\n scrollImages = findViewById(R.id.scroll_images)\n topScrollContainer = findViewById(R.id.ll_images_container)\n llLoading = findViewById(R.id.ll_loading)\n vTip = findViewById(R.id.v_tip)\n\n btnAudioCancel.setOnClickListener {\n stopRecord(true)\n }\n\n btnAudioDone.setOnClickListener {\n stopRecord(false)\n }\n\n findViewById