Semantic search development
We build semantic search that understands what people mean, combining embeddings with keyword matching and your business rules. Customers find the right product and employees find the right document, even when their wording does not match yours.
What it is and when it fits
Semantic search turns text, and sometimes images, into embeddings so that queries match on meaning. On its own that misses exact terms like part numbers and brand names, so we build hybrid search: vector and keyword retrieval combined, reranked and adjusted with filters, stock, margin or recency. We tune it on your real search logs and measure relevance with labelled query sets.
This is a strong fit for large catalogues with inconsistent product data, multilingual content, or knowledge bases where people search in their own words. If your users mostly search by exact code or your collection is small and well tagged, tuning your current keyword search may be enough. We check your search logs before recommending either.
What we build
Product search
Catalogue search that handles synonyms, vague descriptions and typos, with ranking that respects stock, filters and merchandising rules.
Knowledge search
Search across documents, tickets and wikis that returns the relevant passage, filtered by the user's access rights.
Hybrid ranking pipelines
Vector and keyword retrieval fused and reranked, with scoring you can adjust without redeploying the model.
Multilingual search
Queries in one language matching content in another, useful for Dutch, English and German catalogues alike.
Relevance testing
Labelled query sets and offline metrics, plus click and conversion tracking to confirm changes improve real outcomes.
Semantic search, keyword search or hybrid search.
Semantic search matches on meaning: the query and your content are turned into embeddings and compared, so a search for a waterproof jacket also finds a product described as a rain coat. Keyword search matches the exact words. Hybrid search combines both, and it is what most product catalogues and knowledge bases need.
| Search type | Matches on | Strong at | Weak at |
|---|---|---|---|
| Keyword search | The exact words in the query | Part numbers, brand names and exact phrases | Synonyms and different wording |
| Semantic search | Meaning, through embeddings | Natural questions and varied wording | Exact codes and rare terms |
| Hybrid search | Both, combined and reranked | Collections that get both kinds of query | Needs tuning on real search logs |
Our guide to embeddings and semantic search covers choosing a model, pgvector or a vector database, and how to measure relevance.
How it works
- 01
Study the search logs
We look at what people search for, where they get zero or poor results and which queries matter most commercially.
- 02
Build a relevance baseline
We label a query set, measure your current search against it and set targets for the new ranking.
- 03
Build and tune
We choose embedding models, build the hybrid pipeline and tune ranking against the query set and your business rules.
- 04
Release and measure
New search runs alongside the old one on part of the traffic, and we compare conversion and engagement before switching fully.
Related work
Built with
All technologiesFurther reading
Embeddings explained: a practical guide for product teams
How embeddings power semantic search, how to choose a model and a vector store, and how to measure whether your results improved.
5 min read
How to evaluate RAG: retrieval metrics, faithfulness and golden sets
How to do RAG evaluation properly: separate retrieval from answers, check faithfulness claim by claim and build a golden set you can trust.
4 min read
Common questions
Often not. If you run Elasticsearch or Postgres, we can add vector search and reranking to it. We only suggest a separate vector database when volume, filtering or latency requirements call for one.
We test several on your data, including commercial models from OpenAI and open models from Hugging Face that can be self-hosted. The best model varies by language and domain, so the relevance set makes the decision.
Well-built hybrid search returns results fast enough for search-as-you-type on most catalogues. Reranking adds some latency, so we apply it selectively and cache where it makes sense.