# What is Semantic Search?

Also called Vector Search, Dense Retrieval.

Semantic search retrieves results by meaning rather than by literal word overlap. Queries and documents are converted into embeddings, and the system returns the items whose vectors sit closest to the query vector. This lets a search for ending my plan surface a document titled cancellation policy, even when the two share no terms.

Keyword search scores documents on term frequency and rarity, so a document is found when it contains the words typed. Semantic search replaces that test with vector proximity. Both the query and the stored content pass through the same embedding model, and ranking follows a distance measure such as cosine similarity.

The advantage appears with vocabulary mismatch, which is common when users describe a problem in their own words while documentation uses formal terminology. Semantic search also handles questions phrased as full sentences, and it degrades gracefully across paraphrase, synonym, and loose topical relation. It generalizes to other languages when the embedding model was trained on them.

The weaknesses are the mirror image. Semantic search can miss an exact identifier, a version number, or a rare proper noun, because such tokens contribute little to a vector that encodes overall meaning. It can also return something topically near but factually wrong, since proximity is not relevance and never verifies a claim.

For this reason production systems commonly combine semantic and keyword retrieval, then rerank the merged list. Evaluation uses standard information retrieval measures over a labeled question set, and teams that skip this step usually cannot tell whether a change helped or hurt. A small set of representative questions with known correct passages is enough to start.

## Key points

- Matches meaning through vector proximity, not word overlap.
- Handles synonyms, paraphrase, and natural language questions.
- Misses exact identifiers, codes, and rare proper nouns.
- Often combined with keyword search and reranking.

## In practice

A user searches for the app keeps kicking me out. No document contains that phrase. The embedding places the query near a troubleshooting article titled session expires unexpectedly, which ranks first. A keyword engine would have returned nothing useful, while the semantic engine matched the described symptom to the documented cause.

## Related terms

- [Embedding](/en/glossary/embedding)
- [Hybrid Search](/en/glossary/hybrid-search)
- [Vector Database](/en/glossary/vector-database)
- [Reranking](/en/glossary/reranking)
- [Cosine Similarity](/en/glossary/cosine-similarity)

[Back to the AI Glossary](/en/glossary)