URL Trigger
proxy.ts validates owner/repo and starts local processing.
Client-side • Free
Turn any GitHub repo into a chat you can query
Browser-native RAG. Embeddings, retrieval, and storage, all on-device via WebGPU. No server. API keys encrypted locally.
Evals⚡WebGPU InferenceEmbeddings computed on your GPU via WebGPU
🌲AST ChunkingCode split by syntax, not line count
🔍Hybrid SearchCombines vector and keyword search
🗜Binary Quantization32x smaller index. 750KB → 23KB for 500 chunks.
🔐Encrypted Key VaultAPI keys stored locally, locked by passkey
How It Works
GitHub Fetch
Repository tree and source blobs are pulled into browser memory.
AST Chunker
tree-sitter WASM segments code into semantically stable chunks.
Embedding Pipeline
transformers.js runs WebGPU embeddings for chunk vectors.
Binary Quantisation
Vectors are compressed for fast local similarity operations.
IndexedDB
IndexedDB stores vectors, metadata, and symbol structure locally.
Hybrid Search
Hamming distance and regex matching are fused for recall.
Matryoshka Reranker
Candidate chunks are reranked before prompt construction.
WebLLM Worker
Qwen2-0.5B generates grounded answers in a dedicated worker.
Chat UI + CoVe
Answer and verification loop are streamed to the interface.