Vector Search vs Keyword Search
Vector search finds results by meaning — matching what a query is about, even when the exact words differ. Keyword search finds results by matching the actual words in the query against the actual words in the stored content. Neither wins outright: vector search catches a differently-worded question about the same thing that keyword search would miss entirely, and keyword search catches an exact term, code, or acronym that vector search can blur past. Most serious systems that genuinely need both end up combining them rather than picking one.
What vector search does
Text gets converted into embeddings — numbers positioned so that text with similar meaning ends up near other text with similar meaning — and a search finds the stored embeddings closest to the query's. Ask "how many holidays do employees receive" and it can retrieve a section titled "Annual Leave Entitlement," even though not one word matches, because the two land near each other in meaning. A vector database is the usual system built to do this search quickly at scale.
What keyword search does
It matches the literal words in a query against the literal words in the stored content — the technique behind traditional search engines for decades, typically scored by some variant of BM25, a well-established formula that weighs how rare and how frequent each matching term is. It doesn't know two different words mean the same thing, but it also doesn't need embeddings, an index of vectors, or anything approximate — an exact term either appears in a document or it doesn't.
Side by side
| Vector search | Keyword search | |
|---|---|---|
| Matches by | Meaning | The literal words used |
| Good at | Different wording for the same idea | Exact terms, codes, acronyms, names |
| Bad at | An exact term that doesn't embed distinctively | A question worded differently than the source |
| Needs embeddings | Yes | No |
| Typical use | Semantic retrieval in a RAG pipeline | Product search, log search, anything with exact identifiers |
Which one your problem calls for
Vector search fits when questions get asked in many different ways but are really asking about the same underlying content — a support knowledge base where "how do I get a refund" and "cancel my order and get my money back" should both find the same policy page.
Keyword search fits when what matters is an exact match — a product SKU, an error code, a specific name — where meaning-based matching can miss the one thing that actually identifies the right result, because an exact code often doesn't embed distinctively from similar-looking ones.
Both together (hybrid search) fits when a system genuinely needs both — differently-worded questions and exact-term lookups in the same knowledge base. It's not a free upgrade: it means tuning two retrieval systems and how their results get combined or weighted against each other, and getting that balance wrong can produce worse results than either method alone. Reach for it once you have a real, specific gap vector search alone is missing — not by default.
In this guide
FAQ
Does vector search make keyword search obsolete?
No. They fail in different, complementary ways — vector search on wording variation the source never used, keyword search on exact identifiers that don't carry distinctive meaning. Neither one covers the other's real gap, which is exactly why hybrid search exists rather than one approach simply replacing the other.
Is hybrid search always the safer default?
No. It adds a real tuning problem — how much weight each method's results get — and getting that balance wrong can make results worse than a single well-chosen method alone. Start with whichever single approach fits the content, and add the other only once a real, specific gap shows up that it would fix.