feat(experimental): integrate client-side WebGPU playground (ai4data_lab)

experimental/ai4data_lab/.gitignore

@@ -0,0 +1,31 @@
+# Logs
+logs
+*.log
+npm-debug.log*
+yarn-debug.log*
+yarn-error.log*
+pnpm-debug.log*
+lerna-debug.log*
+
+node_modules
+dist
+dist-ssr
+*.local
+
+# Editor directories and files
+.vscode/*
+!.vscode/extensions.json
+.idea
+.DS_Store
+*.suo
+*.ntvs*
+*.njsproj
+*.sln
+*.sw?
+
+# Exclude large binary model assets
+public/onnx/*.onnx
+public/wllama/*.wasm
+.venv
+venv
+

experimental/ai4data_lab/.oxlintrc.json

@@ -0,0 +1,8 @@
+{
+  "$schema": "./node_modules/oxlint/configuration_schema.json",
+  "plugins": ["react", "oxc"],
+  "rules": {
+    "react/rules-of-hooks": "error",
+    "react/only-export-components": ["warn", { "allowConstantExport": true }]
+  }
+}

experimental/ai4data_lab/README.md

@@ -0,0 +1,131 @@
+# ai4data Lab
+
+ai4data Lab is a fully client-side WebGPU playground that runs large language models, embeddings, and zero-shot entity extraction in the browser. Documents, prompts, and model state never leave the user's machine.
+
+## What runs in the browser
+
+- **Gemma 4 Mobile (WGSL)** for evidence-grounded document chat. Citation IDs travel from the parser through retrieval, prompt, and the evidence map.
+- **Bonsai 1.7B ONNX** as an alternate chat model loaded through Transformers.js.
+- **Xenova/all-MiniLM-L6-v2** dense embeddings with reciprocal rank fusion over lexical candidates.
+- **Anonym-IA/gliner_large-v2.1 ONNX** for zero-shot span extraction on arbitrary text and labels.
+- **LiquidAI/LFM2.5-2.6B-ONNX** research agent driven from a Web Worker, with tool permissions and a compiled artifact.
+
+## Requirements
+
+- Node.js 18 or newer (Vite 8 requires a recent Node).
+- A Chromium-based browser with WebGPU enabled (Chrome 113+, Edge, Brave, Arc). Safari and Firefox need nightly or experimental builds.
+- Roughly 3 GB of free disk for the cached model weights, plus enough GPU memory for the selected model.
+
+## Model assets (download separately)
+
+The repository excludes the large ONNX and WebAssembly assets that Vite serves directly. Download them into the matching paths before running the app:
+
+- `ai4data_lab/public/onnx/gliner_large-v2.1_q4.onnx` (about 900 MB) from `https://huggingface.co/Anonym-IA/gliner_large-v2.1/resolve/main/onnx/model_q4.onnx`
+- `ai4data_lab/public/onnx/gist-embedding-v0.onnx` and `ai4data_lab/public/onnx/splade-sparse.onnx` from the corresponding Hugging Face repos.
+- `ai4data_lab/public/wllama/wllama.wasm` from the `wllama` package's release artifacts, for example `https://unpkg.com/@wllama/wllama@3.5.1/dist/wllama.wasm`.
+
+The browser also streams chat and agent weights from Hugging Face at runtime, so an internet connection is still required on first run.
+
+## Quick start
+
+```bash
+cd ai4data_lab
+npm install
+npm run dev -- --host 0.0.0.0
+```
+
+Vite prints the local URL (default `http://localhost:5173/`). Use `--host 0.0.0.0` when running on a server so the page is reachable from another device.
+
+### Other commands
+
+```bash
+npm run build    # production bundle in dist/
+npm run preview  # serve the production bundle locally
+npm run lint     # oxlint over the source tree
+```
+
+## Project layout
+
+```
+ai4data_lab/
+  index.html              # Vite entry, sets <title> and loads main.jsx
+  vite.config.js          # React + Tailwind plugins, port 5174, COOP/COEP headers
+  package.json            # Scripts and dependencies
+  public/                 # Static assets served at /
+  src/
+    main.jsx              # React root
+    App.jsx               # All three tabs: Chat, Extract, Agent
+    App.css, index.css    # Tailwind and global styles
+    agentConfig.js        # Tools and prompts for the research agent
+    modelConfig.js        # Model metadata for the UI
+    worker.js             # Shared worker entry helper
+    hooks/
+      useAgentWorker.js   # Web Worker bridge for the LFM agent
+    workers/
+      agent.worker.js     # LFM runtime and tool execution
+    lib/
+      gemma-4-e2b.js      # Generated Gemma 4 WGSL runtime
+      bonsai27b.js        # Generated Bonsai runtime
+    assets/               # Logos and static images
+  WEBGPU_MASTERY_GUIDE.md # Companion notes on WebGPU usage
+```
+
+## How a document chat turn works
+
+1. The user picks a PDF, JSON, text, or Markdown file in the sidebar. `App.jsx:handleFileUpload` parses the file into citation-bearing section blocks.
+2. When the chat model is ready, MiniLM embeddings are computed for the parsed blocks. The UI shows a stem-search fallback when embeddings fail to load.
+3. `App.jsx:hybridRetrieve` combines lexical and dense scores with reciprocal rank fusion, `App.jsx:rerankCandidates` keeps the top five sections, and the prompt builder constrains Gemma to the surviving evidence.
+4. The Gemma runtime streams tokens back into `gemmaMessages`. `AssistantMessage` and `EvidenceMapModal` render the answer with numbered citations and an evidence map.
+5. Clearing chat preserves the document, embeddings, and model instance.
+
+See `docs/gemma4-document-chat-architecture.md` in the repository root for the full developer walkthrough.
+
+## Configuration
+
+Set defaults through environment variables before `npm run dev` when needed:
+
+- `VITE_DEFAULT_MODEL`: which chat model is preselected in the dropdown.
+- `VITE_GLINER_VARIANT`: which GLiNER ONNX variant is preselected.
+
+Vite already loads `.env` files at the project root, so create one if you need overrides. Most users do not need any configuration.
+
+## Privacy
+
+- The browser downloads model weights directly from Hugging Face or from the local `/onnx/` static asset.
+- Documents are parsed in memory; the server never receives their contents.
+- The LFM research agent runs inside a Web Worker and only contacts external services the user has enabled (for example, Wikipedia search) through the agent's tool list.
+
+## CLI bridge (experimental)
+
+The page can stream prompts to a CLI over a local WebSocket relay. The browser is the only WebGPU host; the CLI is a thin client.
+
+```bash
+cd ai4data_lab
+python -m venv .venv && source .venv/bin/activate
+pip install -r requirements.txt
+python scripts/bridge_server.py
+```
+
+Open the page with the bridge flag and load a model:
+
+```
+http://localhost:5174/?bridge=1
+```
+
+The page reports `ready` to the relay when the model is loaded. The CLI tool in `app/cli/ai4data_lab.py` connects to `ws://127.0.0.1:8765` and streams prompts. See `app/cli/ai4data_lab.py` for the command-line interface.
+
+### Bridge limitations
+
+- One CLI client and one page at a time.
+- The bridge is opt-in. Open the page with `?bridge=1`; otherwise the relay sees no `ready` frame.
+- `stop` cancels the current request and any in-flight generation because the UI composer and the CLI share the same stop flag.
+
+## Troubleshooting
+
+- **`navigator.gpu` missing**: enable WebGPU in your browser flags. The chat model cannot start without it.
+- **First run is slow**: weights are streamed on demand and cached in CacheStorage. Subsequent runs reuse the cached files.
+- **GLiNER extraction fails on long text**: the implementation builds all spans up to width 12. Shorten the input or split it into smaller batches.
+
+## License
+
+This project is part of the AI-DQSS repository and follows the same license as the parent project.