Semantic Search
Search that matches on meaning rather than exact keywords, usually by comparing the embedding of a query with the embeddings of documents.
Definition
Traditional (lexical) search finds documents that contain the words in the query, weighted by how rare and how frequent those words are. Semantic search instead converts the query and every document into embeddings and returns the documents whose vectors are closest. Because the vectors encode meaning, synonyms, paraphrases and even other languages can match.
The trade-off is precision on exact strings. A semantic index may rank "order 10234" below a paragraph about ordering in general, because the number carries little meaning to the model. It is also only as good as the embedding model and the way documents were chunked.
Example: a staff member searches the intranet for "can I expense a taxi from the airport?". Semantic search surfaces the travel policy section titled "Ground transport reimbursement", which shares no words with the query.
Why it matters for your business
- Fewer failed searches: users do not have to guess the vocabulary your documents use.
- Works across languages and across departments that name the same thing differently.
- It is the retrieval step of every RAG assistant, so its quality caps the assistant's quality.
- Needs an evaluation set and ongoing tuning; "install a vector database" is not the whole job.
How Teksmobile uses it
Our semantic search service covers embedding selection, indexing and relevance testing for intranets, product catalogues and support content. For most clients we recommend hybrid search, combining meaning with keywords, and we build it on the same vector search stack.
Related terms
Part of the Teksmobile AI Glossary. For longer reads, see our guides.
Frequently asked questions
Does semantic search replace keyword search?
Rarely. Keyword search is still better for exact identifiers, product codes, names and rare terms. Most production systems combine both as hybrid search.
Does semantic search need an LLM?
No. It needs an embedding model and a vector index. An LLM is only involved if you want generated answers on top of the results, which is RAG.
How do we know it is working?
Build a small set of real queries with known good results and measure recall and ranking before and after each change. Without that, tuning is guesswork.