A knowledge system with hybrid retrieval
I designed and built a retrieval-augmented (RAG) knowledge system on ChromaDB over resolved incidents and internal documentation. Retrieval is hybrid: dense vector similarity and BM25 keyword matching, fused with reciprocal rank fusion, so a query finds both the passage that means the same thing and the one that contains the exact error code.
Search this page the same way
BM25 keywords
Similarity trigrams
Fused RRF, k = 60
Both rankers run in your browser over the sections of this page. BM25 is the real algorithm. The similarity ranker compares character trigrams and stands in for dense embeddings, because no model is shipped to your browser; it is what catches the typo.