Building Low-Latency Search with Hybrid Indexes

Developer search fails when you over-commit to either lexical matching or embeddings alone. Hybrid retrieval wins in most practical corpora.

Retrieval Plan

• Lexical BM25 for precision anchors.
• Vector ANN for semantic recall.
• Reciprocal rank fusion for stable blending.

RRF(d) = Σ 1 / (k + rank_i(d))

Operations Notes

Keep vector index rebuilds decoupled from lexical refresh cadence to avoid synchronized regressions.

Search quality is mostly ranking engineering, not model hype.