Glossary
Semantic search
Semantic search finds documents by meaning rather than by shared words. The query and every document passage are converted into embeddings, and the search returns the passages whose embeddings lie nearest the query, even when the two have no words in common.
A match with no shared words
An engineer searches for "server won't boot". The fix sits in a runbook titled "Power-on self-test failure recovery". A keyword index sees no overlap and returns nothing. Semantic search converts the query and each runbook passage into embeddings, finds the runbook close by, and returns it. Multilingual models extend the same effect across languages, so a question in German can match a procedure written in English. The stored points are searched in a vector database.
Semantic and keyword search compared
| Keyword search | Semantic search | |
|---|---|---|
| Matches on | Shared words | Closeness in meaning |
| Paraphrase and synonyms | Missed | Found |
| Exact codes, names, versions | Found reliably | Often blurred into near neighbours |
| Meaning of a score | Stable within one index | Varies by embedding model |
The two fail in opposite places, which is why production systems usually run both and merge the lists, the approach called hybrid search.
What does result quality depend on?
The embedding model sets the ceiling, and text preparation sets how close to it results get: a passage cut mid-argument by poor document chunking embeds as a muddled idea and matches nothing well. Because scores shift between models, systems keep results by rank position rather than by a fixed score cutoff that breaks at the next model change. Closeness in meaning also says nothing about authority: a draft and a final version of the same policy sit side by side, and only metadata such as status and date separates them.
Search results versus generated answers
On its own, semantic search returns documents for a person to read and judge. Inside retrieval-augmented generation the same results become the evidence for a written answer, so a retrieval miss turns into a confident wrong answer instead of an empty results page.