Vector Search vs. Keyword Search – Definition and Differences
Compare vector and keyword search: how each ranks, where each fails, why hybrid retrieval wins, and how to choose for a given corpus.
The two approaches fail in opposite directions, which is why production systems rarely choose one. Understanding each failure mode is what makes combining them effective rather than merely fashionable.
1. What Is Keyword Search?
Lexical retrieval ranks documents by term overlap with the query, weighted by how informative each term is.
- Mechanism: an inverted index from terms to documents, scored by term frequency against document frequency.
- Strength: exact matching, including rare identifiers.
- Tuning: analysers, stemming, synonyms, and field weights.
2. What Is Vector Search?
Semantic retrieval embeds query and documents into a shared space and ranks by distance.
- Mechanism: an embedding model plus an approximate nearest-neighbour index.
- Strength: paraphrase and conceptual similarity.
- Tuning: model choice, chunking, and re-ranking.
3. Key Differences
| Dimension | Keyword | Vector |
|---|---|---|
| Matches | Terms and their variants | Meaning |
| Rare identifiers | Excellent | Poor |
| Paraphrase | Poor without synonyms | Excellent |
| Negation | Literal and reliable | Weak — opposites sit close together |
| New vocabulary | Works immediately | Needs the term in training data |
| Explainability | Matched terms are visible | Distances are opaque |
| Index cost | Low | Higher: embedding plus vector index |
| Update cost | Reindex the document | Re-embed the document |
4. Where Each Fails
Keyword search fails when the user's vocabulary differs from the author's: a symptom described in plain language against documentation written in product terms.
Vector search fails on exact tokens — part numbers, error codes, version strings — and on negation, where "supports X" and "does not support X" are neighbours.
5. Why Hybrid Wins
Hybrid retrieval runs both and fuses the rankings, usually by combining ranks rather than raw scores, since the two scales are not comparable.
The benefit is clearest on mixed queries — natural language plus an identifier — which are extremely common in technical and commercial search. Vectors carry the prose; the lexical index anchors the token.
6. Re-Ranking Sits Above Both
Whichever retrievers are used, a re-ranking model that scores query and passage together is usually the cheapest large quality improvement. Retrieve broadly, re-rank precisely, and keep only the top few.
7. Choosing for a Corpus
- Identifier-heavy (catalogues, code, logs) — keyword first, vectors added for descriptive queries.
- Prose-heavy (documentation, articles, tickets) — vectors first, keyword added for exact tokens.
- Mixed — hybrid from the start; it is the common case.
- Small corpus — an exact vector scan is simple and adequate; no ANN index is required.
8. Summary
Keyword search matches terms, vector search matches meaning, and each is weakest exactly where the other is strongest.
Hybrid retrieval with re-ranking is the default that survives contact with real queries — not because combining is sophisticated, but because real queries contain both prose and identifiers.