Semantic search

How it works

Classic keyword search, the kind built into Elasticsearch or a website search box, scores documents by how often the query's words appear in them. Semantic search converts the query and every document passage into embeddings, then returns the passages whose vectors sit closest. A search for "laptop won't turn on" can find an article titled "Troubleshooting power issues" even though no words match. Sentence embedding models such as Sentence-BERT (Reimers and Gurevych, 2019 (opens in a new tab)) made this fast enough to run over large collections.

Results can then pass through a reranker: a second model that reads the query and each candidate together and scores relevance more precisely. Rerankers are slower, so they usually see only the top 20 to 100 results.

How businesses use it

• Help center and intranet search, where staff and customers describe problems in their own words
• The retrieval step of RAG assistants
• Finding similar past cases: support tickets, claims, contracts, project files
• Matching job descriptions to résumés, or products to customer requests

It works best on prose, such as articles, emails, notes and call transcripts. Our RAG knowledge systems work uses it as one half of the retrieval layer.

Where it goes wrong

Semantic search is poor at exact matches. Search for invoice INV-20817 or part 4410-B and it returns things that look similar, which is the wrong answer. Queries of 1 or 2 words also give it very little meaning to work with.

It also returns something every time, even when nothing relevant exists. An assistant built on it needs a relevance threshold, below which it says it found nothing instead of answering from a weak match.

The fix for exact terms is hybrid search, which runs keyword and semantic search together and merges the results. Before buying a new search product, test candidate setups on 50 real queries from your own search logs and read what comes back.