Hybrid search combines semantic (vector) search with traditional keyword-based search — covering both "similar meaning" and "exact term match" needs, which pure semantic search alone can miss.
Why Semantic Search Alone Isn't Always Enough
Query: "error code E4521"
Semantic search alone might return: documents broadly about
"error codes" or "troubleshooting," without necessarily
surfacing the one document that mentions this EXACT code —
specific identifiers, codes, and proper nouns don't always
carry strong distinguishing signal in embedding space the way
they do in exact keyword matching.
Keyword search: reliably finds documents containing the
literal string "E4521".Product codes, specific error codes, proper nouns, and exact technical terms are cases where keyword matching can outperform pure semantic similarity — hybrid search aims to get the benefit of both.
How Hybrid Search Typically Combines Results
semantic_results = vector_search(query_embedding, top_k=20)
keyword_results = keyword_search(query_text, top_k=20)
combined_results = merge_and_rerank(
semantic_results,
keyword_results,
weights={"semantic": 0.6, "keyword": 0.4} # illustrative —
# tunable per use case
)
final_results = combined_results[:5]The exact merging strategy (simple weighted scoring, reciprocal rank fusion, or a dedicated reranking model — see Reranking) varies by implementation and tool.
When Hybrid Search Is Worth the Added Complexity
| Signal | Favor Hybrid Search |
|---|---|
| Content includes specific codes, IDs, or exact technical terms users search for literally | Yes |
| Content is mostly natural-language prose where meaning-based matching is what users need | Pure semantic search may be sufficient |
| You've measured semantic-only search missing relevant exact-match results | Strong signal to add hybrid search |
Practical Use Case
Technical documentation search (with error codes, API method names, specific version numbers) is a classic case where hybrid search meaningfully outperforms pure semantic search — general conversational or narrative content search often does fine with semantic search alone.
Common Mistakes
- Adopting hybrid search complexity by default without first measuring whether pure semantic search is actually missing relevant results for real queries
- Weighting semantic and keyword scores arbitrarily without testing/tuning against real query patterns
Interview Relevance
"When would pure semantic search underperform, and how does hybrid search address it?" — exact-match needs (codes, IDs, specific terminology) that embedding similarity doesn't strongly distinguish is the core scenario to identify.
Practice Question
A technical support search feature is missing results for queries containing specific product SKU numbers. Explain why, and how hybrid search would help.
