Semantic Search vs Vector Search vs Keyword Search

These three terms are often mixed up:

• Semantic search is a goal: find results that match the meaning of a query, even when they use different words.
• Vector search is a technique: turn items into vectors (embeddings) and find the vectors nearest to the query’s vector.
• Keyword search (also called lexical or full-text search) is another technique: find documents that contain the query’s words, ranked by a formula such as BM25.

Vector search over text embeddings is today’s most common way to do semantic search, which is why the two terms are used as synonyms. But they are not the same thing: semantic search can be built in other ways, and vector search is used for much more than text meaning.

How each works

Keyword search builds an inverted index: for every word, the list of documents that contain it. A query is split into words, matching documents are found instantly, and they are ranked by how often the words appear and how rare they are. It is exact and fast, and it explains itself: the matched words can be highlighted.

Vector search stores a vector for every item. A query is turned into a vector with the same model, and the system returns the items whose vectors are closest, by cosine similarity or another distance. Approximate nearest-neighbour indexes make this fast for millions of items. See vector databases and indexes.

Semantic search is whatever achieves meaning-based results: vector search over text embeddings, a language model that rewrites the query before a keyword search, a reranker model that re-scores candidates by meaning, or a combination.

Side by side

	Keyword search	Vector search	Semantic search
What it is	Technique: match words	Technique: nearest vectors	Goal: match meaning
“I want my money back” finds “refund policy”?	No, no shared words	Yes, if the embeddings are good	Yes, that is the point
Exact codes, names, error numbers	Excellent	Often weak	Depends on the method
Explainable	Yes: shows matched words	Hard: a similarity score only	Depends
Works beyond text	No	Yes: images, audio, products, users	Mostly about language
Needs a model	No	Yes, an embedding model	Yes, some model of meaning

Vector search is not only semantic

Because vector search works on any embedding, it also powers image search (“find photos like this one”), recommendations (“users similar to this user”), duplicate detection and anomaly detection. None of these is “semantic search” in the language sense. That is the clearest reason to keep the terms separate.

What production systems actually use

Most good search and RAG systems use hybrid search: run keyword and vector search together, merge the results (for example with reciprocal rank fusion), then let a reranker model order the top candidates by meaning. Keyword search catches exact terms, vector search catches paraphrases, and the reranker gives the final semantic ordering. The hybrid search and reranking lesson builds one step by step.

Common mistakes

1. Dropping keyword search when adding vectors, then failing on product codes and names.
2. Assuming any vector search is semantic. Quality depends entirely on the embedding model and on your domain.
3. Mixing embedding models in one index; their vectors cannot be compared.
4. No evaluation set. Measure how often the right document appears in the top results, for both methods, on real queries.

Interview angle

For “semantic search vs vector search”:

• Semantic search is the goal, vector search is one way to reach it, keyword search is the classic alternative.
• An example where each wins: paraphrases for vectors, exact codes for keywords.
• Why hybrid search plus a reranker is the usual production answer.
• That vector search also serves images, recommendations and deduplication.

Learn how embeddings work in what embeddings are, then practise with the LLM interview questions.

Key takeaway

Semantic search means finding by meaning. Vector search finds the nearest embeddings and is the most common way to do semantic search, but it also works for images, products and users. Keyword search finds exact words and is still essential. Most production systems combine keyword and vector search and finish with a reranker.