Hybrid search combines lexical matching (such as BM25) with dense vector similarity so systems can honor exact terms and codes as well as natural-language meaning. It is useful when corpora mix jargon, identifiers, and free-text questions.
Whilst hybrid is sold as “best of both,” in practice fusion weights and candidate depths decide whether IDs or paraphrases win. For example, a part number query returns marketing copy because dense results dominate the merge. We often recommend Reciprocal Rank Fusion as a baseline, then tune on labeled queries that include both SKUs and plain-language asks.
Dense embeddings miss rare tokens, IDs, and regulatory phrases; BM25 misses paraphrase. Hybrid search runs both and merges rankings. Reciprocal Rank Fusion is a durable default because it needs little score calibration — it rewards documents that rank well on either list. Measure with labeled queries (recall@k, MRR) and keep a failure set where hybrid uniquely wins. Pair with metadata filters and a cross-encoder when precision at the top matters more than raw recall.
def rrf_fuse(rank_lists: list[list[str]], k: int = 60) -> list[str]:
scores: dict[str, float] = {}
for ranks in rank_lists:
for rank, doc_id in enumerate(ranks, start=1):
scores[doc_id] = scores.get(doc_id, 0.0) + 1.0 / (k + rank)
return [d for d, _ in sorted(scores.items(), key=lambda x: x[1], reverse=True)]
bm25_ids = bm25_search(query, n=50) # sparse
dense_ids = vector_search(query, n=50) # dense
hybrid = rrf_fuse([bm25_ids, dense_ids])[:20]