How it works
- An index (
public/index.json) of the repo is precomputed by splitting files into chunks and attaching embeddings for semantic search. - In your browser, your question is embedded, similar chunks are retrieved with cosine similarity, and a small local model tries to generate an answer from those chunks.
- Everything runs client-side via WebGPU/CPU — no server, no API keys, and no paid services.
Note: This is a lightweight demo. Small in-browser models can miss context or give weak answers, especially for vague questions. For best results, ask specific “what/where/how” questions about this repo.