Offline-first MDN Web Docs RAG-MCP server ready for semantic search with hybrid vector and full‑text retrieval
Hybrid vector plus BM25 retrieval beats vector-only for technical docs.
AI developers building agents that need documentation access
LangChain · LlamaIndex · Sourcegraph Cody
While tinkering with RAG ideas I've thoroughly processed the entire MDN Web Docs original content, pre-ingested it into LanceDB, uploaded the 50k+ rows dataset (https://huggingface.co/datasets/deepsweet/mdn) to HuggingFace, and published a RAG-MCP server (https://github.com/deepsweet/mdn) ready for semantic search with hybrid vector (1024-d) and full‑text (BM25) retrieval.
A screenshot is worth a thousand words: https://raw.githubusercontent.com/deepsweet/mdn/main/example...
Replaces Windows Search with vector + full-text hybrid; EXIF geocoding and function-level code chunking actually work.
Pure Rust CPU-only code search with persistent index beats transformer-heavy alternatives.
Local RAG + MCP for Claude with zero external dependencies—elegant constraint execution.
Hybrid BM25 + vector search via MCP beats pure keyword or pure semantic for API docs.
Single binary with zero deps, but just another agent wrapper atop existing APIs.
Visual chunking comparison beats guessing — export production-ready code.
