harshaAI / Browser prototype
WebLMS
A browser extension that turns visible DOM text and OCR into extractive answers, with IndexedDB caching and a local request monitor.
View repository01 / COMPONENTS & DATA FLOW
Inside the system.
- 01DOM + images
- 02Text + table extraction
- 03Question classification
- 04Extractive answer
Structured page extraction
A DOM TreeWalker skips scripts, styles, and hidden elements while collecting text. Separate paths gather viewport-adjacent content, document metadata, and table rows.
OCR off the page thread
A Comlink-exposed OCR worker wraps Tesseract.js and returns recognized text with OCR confidence. IndexedDB stores page contexts and model-related cache records.
Extractive Q&A
QAEngine classifies questions as summary, extraction, specific, or general, then scores and selects sentences from the page context. Timing markers cover processing and answer generation.
02 / IMPLEMENTATION DECISIONS
Why it works this way.
Local context is the core
The implementation does not require a FastAPI or Redis service. Its privacy monitor wraps fetch and XMLHttpRequest, permits extension/blob URLs and local hosts, and rejects other absolute hosts in the monitored context.
Extraction before generation
The implemented answer path uses sentence-selection heuristics. The model loader includes an ONNX integration placeholder; the current worker signals the main Q&A engine rather than running a downloaded model.
03 / SCOPE & TRADEOFFS
The implementation boundary.
A prototype of local page understanding. The current heuristics are not a generative LLM, and the request monitor is not a complete extension-wide security sandbox.
04 / READ THE IMPLEMENTATION