Summary
A knowledge base you choose, not one the model inherited.
The strategies here — hybrid search, reranking, separate indexing of code examples — are the ones that decide whether RAG helps or quietly hurts, and each is a switch rather than a baked-in choice. The author describes the repository as a testbed on its way into Archon, so expect the interface to move.
What it is
An MCP server pairing the Crawl4AI crawler with a Supabase vector store, so pages you scrape become a knowledge base the assistant can search.
What you get
- Pages crawled and stored, then retrieved over — answers come from documentation you chose
- Hybrid search combining vector similarity with keyword matching
- Contextual embeddings for richer semantic matching, switchable in configuration
- Cross-encoder reranking, for when the right passage keeps coming back third
- Code examples extracted and indexed separately, so retrieval can return a snippet rather than prose about one
- A knowledge graph mode used to check generated code against a real repository, catching invented APIs
Requirements
Supabase for the vector store and OpenAI for embeddings. Each retrieval strategy is a configuration switch, off by default.
Setup effort
Build from source — clone the repository and build it, then point your client at the binary