Hybrid Search

Neither semantic search nor keyword search is universally superior. Semantic search excels when the user paraphrases or queries conceptually, but struggles with rare proper nouns, product codes, serial numbers, and technical identifiers where exact spelling is critical. Keyword search handles these exact-match cases well but fails on synonymy and paraphrase. Hybrid search addresses this by running both retrieval pipelines in parallel and merging their ranked result lists, so each modality compensates for the other's weaknesses.

The sparse retrieval component is typically BM25 — the probabilistic ranking function that scores documents based on term frequency, inverse document frequency, and document length normalization. Modern sparse retrieval has also been augmented by learned sparse models (SPLADE, BM25+), which expand queries and documents with semantically related tokens while maintaining sparse representations compatible with inverted index infrastructure. The dense component uses standard embedding-based ANN search. Score fusion is most commonly achieved via Reciprocal Rank Fusion (RRF), which combines ranked lists from multiple retrievers without requiring score normalization, or via learned linear interpolation of normalized scores.

In enterprise deployments, hybrid search is the recommended baseline for most knowledge retrieval applications because real-world query distributions are heterogeneous: the same system may receive conceptual questions ("what is our return policy?"), exact-match lookups ("order #INV-29847"), and technical queries ("CVE-2024-12345 mitigation steps"). Tuning the balance between dense and sparse components — the alpha weight in interpolation approaches — should be driven by offline evaluation against a representative held-out query set sampled from actual production traffic.