Vector Search vs Keyword Search
Vector search ranks results by embedding similarity, matching meaning even without shared words; keyword search ranks by exact term overlap.
| Dimension | Vector (Semantic) Search | Keyword Search (BM25) |
|---|---|---|
| Paraphrased or synonym queries | Wins here "Forgot my login" matches a "reset password" document despite sharing zero words. | Misses documents that answer the question but use different wording than the query. |
| Exact identifiers (error codes, SKUs, names) | Embedding models have little reason to distinguish rare exact strings — they cluster meaning, not memorize codes. | Wins here Nails an exact error code or part number a paraphrase-oriented method treats as noise. |
| Explainability of a match | A cosine similarity score is a geometric abstraction, harder to justify to an end user. | Wins here A match is a visible shared term — easy to point to and explain. |
| Setup cost | Needs an embedding model, a vector store, and re-embedding whenever the embedding model changes. | Wins here An inverted index is cheap, well-understood infrastructure most databases already support. |
| Typos and misspellings | Somewhat robust by default, since embeddings often cluster similar spellings together — but not guaranteed. | Needs fuzzy matching bolted on explicitly; exact term matching alone is brittle to typos. |
| Rare or novel terminology | A term the embedding model rarely saw in training carries a weak, unreliable signal. | Wins here Exact match does not care whether a term was common in training data — a literal match is a literal match. |
| Production default | Alone, still misses queries built around exact identifiers. | Alone, still misses queries that share no keywords with the right document. |
When to choose Vector (Semantic) Search
Lean on vector search when queries are natural language, paraphrased, or conceptual, and semantic similarity is the actual signal that matters.
When to choose Keyword Search (BM25)
Lean on keyword search when queries contain exact codes, SKUs, names, or jargon, where literal string matching is what actually distinguishes the right document.
The verdict
Treat this as a false choice in production. Hybrid search — running both and fusing the results, typically with reciprocal rank fusion and often a reranker — consistently beats either method alone. Pick just one only for a quick prototype.