16. Hybrid Search

Hybrid search runs dense (vector) and keyword (BM25) retrieval together and fuses the results — covering each method's blind spot.

Why both

	Strength	Blind spot
Dense (vectors)	meaning, paraphrases	exact terms: codes, names, acronyms
BM25 (keywords)	exact tokens	synonyms, rephrasing

A query like ERR_4012 on checkout needs the exact code (BM25) and the concept "checkout error" (dense). Either alone underperforms.

The pipeline

                ┌─▶ DENSE  ─▶ ranked list A ─┐
 query ─────────┤                             ├─▶ RRF ─▶ fused results
                └─▶ BM25   ─▶ ranked list B ─┘

The fusion step is Reciprocal Rank Fusion, which merges the two incompatible score scales by rank.

Code

from rank_bm25 import BM25Okapi
import numpy as np

# BM25 over tokenized chunks
bm25 = BM25Okapi([c["text"].split() for c in chunks])

def hybrid(query, query_vec, k=10):
    # dense ranking
    dense = sorted(chunks, key=lambda c: float(query_vec @ c["vector"]), reverse=True)
    dense_ids = [c["id"] for c in dense[:k]]
    # keyword ranking
    bm = np.argsort(bm25.get_scores(query.split()))[::-1][:k]
    bm25_ids = [chunks[i]["id"] for i in bm]
    # fuse
    return reciprocal_rank_fusion([dense_ids, bm25_ids])

Practical notes

• Weighting — some stores let you weight dense vs keyword; RRF's k also tunes influence. Start balanced.
• Almost always a win — for real apps with names/codes/jargon, hybrid beats pure vector search consistently.

Next: Reranking & Next Steps →