Hybrid Search: BM25 and Dense Retrieval Combined › Summary

Hybrid search combines BM25 keyword retrieval with dense semantic retrieval to reach better recall than either system alone:

• BM25 uses saturated term frequency, inverse document frequency, and document length normalization to rank documents by keyword relevance. Its key hyperparameters $k_1$ and $b$ control saturation and length normalization.
• Dense retrieval encodes queries and documents as dense vectors and ranks by cosine similarity. This makes possible semantic generalization across paraphrases and synonyms.
• Reciprocal rank fusion (RRF) combines ranked lists by summing $1/(k + r)$ contributions per document, bypassing score scale incompatibility. The default $k = 60$ works broadly well without tuning.
• Weighted score combination normalizes scores to a common range and blends them with a weight $\alpha$, letting query-adaptive fusion but requiring careful normalization and hyperparameter tuning.
• Both systems excel at different query types: BM25 handles exact-match and rare-term queries; dense retrieval handles semantic and paraphrase queries. Their recall failures are complementary, which is the core reason hybrid search improves over either alone.
• In production, both indices should be queried in parallel to avoid latency doubling. Systems like Elasticsearch with vector extensions support hybrid search natively.

With hybrid retrieval improving recall, the next challenge is precision: so the top-ranked documents are truly the most relevant ones. The next chapter on Reranking addresses exactly this, using cross-encoder models that jointly encode query and document to produce more accurate relevance scores for the candidate set that hybrid retrieval gives.