The pairing is the value: Crawl4AI puts the page into a Supabase vector store, and the retrieval strategies stacked on top — hybrid search, reranking, code-example extraction — are configuration switches rather than rewrites.
An MCP server built on Crawl4AI with Supabase as the vector store: pages are scraped, embedded and stored, and retrieval runs over what you crawled. Several strategies can be switched on — contextual embeddings, hybrid vector-plus-keyword search, code-example extraction, cross-encoder reranking, and a knowledge graph used for hallucination detection.
- A page crawled and stored for later retrieval
- Retrieval over your crawled content, vector and keyword search combined when hybrid search is on
- Contextual embeddings for richer matching
- Code examples extracted separately, for questions about how an API is used
- Cross-encoder reranking over the results
- A knowledge graph used to check generated code against repository code
A Supabase project for the vector store, and OpenAI for embeddings — support for other embedding models and for running locally with Ollama is described as planned, not present. The author calls the repository a testbed being folded into Archon, so expect it to move.
Build from source — clone the repository and build it, then point your client at the binary