← back to the archiveCover illustration for “Semantic search: search that understands what you meant”
TERM8mo ago·Published · Updated ·by Andy Padia

Semantic search: search that understands what you meant

In short: Semantic search returns results by the meaning and intent of a query instead of its exact words, and works best blended with keyword ranking.

Semantic search is search that matches the meaning and intent of a query instead of its exact words. It recognises synonyms, paraphrases and related concepts, so "cheap flights to the hills" can find "budget airfare to Shimla". It is usually built with embeddings and vector search, using sentence-embedding models of the kind Sentence-BERT introduced in 2019, often combined with keyword ranking and a reranking model.

Semantic search in plain words

Think of a good shop assistant who understands "something for my dad who likes gardening" and does not insist that you name the exact product.

How it works

The system converts queries and documents into embeddings, lists of numbers that stand for meaning, then ranks documents by similarity.

The Sentence-BERT paper, submitted in August 2019, shows why purpose-built sentence embeddings matter. Finding the most similar pair among 10,000 sentences took about 65 hours with BERT and about 5 seconds with its approach.

Better systems add query understanding (spelling, intent, filters), hybrid scoring with keyword signals, and a reranker that reads the top candidates closely before showing them.

Example

As an illustration, an HR portal: "can I work from Goa for a month?" returns the remote-work and travel policy pages, even though neither contains the word "Goa".

Where you'll meet it

  • Web search: Google said in October 2019 that BERT would help it better understand one in 10 English searches in the US
  • Site search and documentation search
  • Enterprise knowledge search and RAG
  • E-commerce product discovery

What it changes for you

For product teams, semantic search reduces "no results" dead ends and supports natural-language questions. For developers, it needs evaluation sets of real queries to tune. My view: for content teams, it rewards pages that answer questions clearly, which overlaps with SEO and AI search optimisation.

Keyword search finds exact terms and is precise for names, codes and quotes. Semantic search finds meaning and handles vague or reworded queries. Most good systems blend both.

Common mistake

My view: replacing keyword search entirely. Users still search for exact product codes, error messages and people's names, and pure semantic search often fumbles those. Blend, then measure on real queries.

Vector search · Hybrid search · Embeddings · Reranking · BM25

From the archive: Why your RAG pipeline returns irrelevant results

Blend semantic and keyword search, then measure on real queries before you drop either.

#field-guide#search#retrieval#embeddings

Sources

  1. arxiv.org/abs/1908.10084
  2. blog.google/products-and-platforms/products/search/search-language-understanding-bert
← older drop
Context engineering: deciding what the model sees
newer drop →
KV cache: the memory that keeps LLM generation fast

related drops

explore all 557 drops →
← back to the archiveday 175