A developer has upgraded a $5/month retrieval-augmented generation (RAG) system to surpass much pricier solutions. Initially relying only on vector search, the system now combines hybrid search using Reciprocal Rank Fusion (RRF) with both semantic and keyword matching, vastly improving accuracy for specific queries like document IDs or technical terms. Key upgrades include cross-encoder reranking for relevance, smarter document chunking with 15% overlap, and an interactive dashboard.

Integration with the Model Context Protocol (MCP) allows AI agents direct access. Built entirely on Cloudflare’s edge, the stack uses Workers AI, Vectorize, and D1 SQL, eliminating external dependencies. Compared to Pinecone, which costs 10-30 times more, this setup offers native hybrid search and reranking at a fraction of the price.

Real-world tests show a +9.3 percentage point gain in accuracy. Despite increased latency from reranking, users can toggle it off for speed. The project is open-source, deployable in minutes.