Hybrid Search

Lexical search scores shared terms and can preserve rare identifiers, while dense search compares learned representations that may connect different wording. A hybrid system runs both or incorporates both signals into a retrieval engine. Scores can be normalized and weighted, or rankings can be fused with a method such as reciprocal rank fusion. A fusion rule is necessary because raw scores from different retrievers usually have different scales. Hybrid retrieval is distinct from reranking: it combines candidate generation signals, while a reranker may then rescore the smaller candidate set more expensively.

The practitioner creates relevance judgments covering semantic queries and exact-match cases, then measures the hybrid against each component separately. Candidate counts, fusion weights and filters are tuned on development data and checked on held-out queries. Deduplication should preserve source identity. Useful artifacts include index mappings, query definitions, a documented fusion rule and rank-level error analysis. Operational measurement records latency and resource use for both retrievers, since combining them can increase work even when the returned result count stays small.

Combining weak retrievers does not guarantee a strong ranking. Score normalization, candidate truncation and filter placement can suppress relevant items before fusion. Hybrid gains also depend on query distribution and corpus vocabulary. A final similarity score is not a probability of relevance or evidence of truth. The combination should be reported with its concrete method and measured against component baselines, rather than presenting hybrid search as automatically superior for every collection.