Hybrid Search

Hybrid search combines results from multiple retrieval methods. A common setup pairs lexical retrieval with dense-vector retrieval to cover exact terms and paraphrases.

Complementary signals

Lexical methods such as BM25 help with identifiers, error codes, and explicit terminology. Dense embeddings can retrieve semantically related wording. Their raw scores need not share a scale.

Fusion and reranking

• Reciprocal rank fusion (RRF) combines ranks rather than raw scores.
• Score-based fusion requires attention to scales and weights.
• A reranker can score the merged candidates afterward.

Qdrant's Query API supports multiple prefetches and fusion. Fusion and pairwise relevance scoring are distinct operations.

Example: an error code

Suppose a user searches for “E104 upload failed.” We would retain exact error-code matches through the lexical channel and add related “cannot submit file” passages through the vector channel, then deduplicate and rank. This is an invented workflow, not a performance claim.

Assess the tradeoff

Our suggested evaluation compares lexical, vector, and hybrid retrieval on the same question set. Inspect evidence recall, top-result quality, latency, and storage cost. Extra complexity does not guarantee better results. Apply consistent document-version and access filtering across candidate channels.

Sources

• Qdrant: Hybrid Search — complementary signals and costs.
• Qdrant: Hybrid Queries — prefetches, RRF, and fusion.