Building a Hybrid Search Engine from Scratch: Combining BM25, Vector Search, and RRF
When building search systems, developers traditionally face a dilemma: Should we rely on exact keyword matching (like traditional databases or Lucene/BM25) or semantic context understanding (like mode
Responses
Markdown supported · ⌘↵ to post
RRF is such an underrated trick, glad to see it get a clear writeup. I run a hybrid setup in production on a fairly large 3D-model catalog and the thing that surprised me is how little the fusion math matters compared to everything around it. RRF basically just works out of the box, the k constant barely moves the needle. Where I actually spend my time is the embedding model choice, and the fact that BM25 keeps saving me on exact part names and model IDs that the vector side blurs together, your CVE example is the same failure mode. One thing worth adding for anyone going to production, watch your latency budget: running both retrievers and then fusing is fine at lab scale but you feel it once the corpus grows, so precompute and cache aggresively. Nice clean explanation of a topic people usually overcomplicate.