Hybrid search
Retrieval that runs keyword matching and vector similarity together, then merges the results, because each catches what the other misses.
In depth
What it really means
Semantic search understands meaning but handles exact strings badly. It will happily match a passage about pricing when you asked about a specific SKU. Keyword search nails exact strings and misses paraphrases entirely. Production systems run both.
The practical consequence is that keywords still matter, just not in the old way. Your exact product names, version numbers, error codes and pricing figures need to appear verbatim in the text, because that is what the keyword half retrieves on.
How it works
- The query runs against a keyword index, usually BM25.
- The same query runs against a vector index. See vector embedding.
- The two result sets are merged and scored, often by reciprocal rank fusion.
- The merged set is reranked and the survivors passed to the model.
Pros & cons
Pros
- Exact product names and figures stay findable.
- Paraphrased and conceptual queries still match.
- It rewards content that is both specific and well-covered conceptually.
Cons
- Two indexes means more places your content can be missing from.
- Merging is tuned per system, so behavior varies and is unpredictable.
- Writing for both at once takes more care than writing for either.
Common mistakes
- Dropping exact terms because ‘keywords are dead’. The keyword half is still running.
- Referring to your product only by a nickname or pronoun after the first mention.
- Leaving version numbers, error codes and price figures out of the text.
- Writing purely conceptually with no concrete, matchable strings.
Best practices
- Use exact product, feature and version names in full, more than once.
- Write specific figures as text, not only inside images.
- Cover the concept properly for the semantic half.
- Include the phrasings buyers actually type, including abbreviations and full forms.
- Repeat the subject noun rather than relying on pronouns.
FAQs
What is hybrid search?
Running two retrieval systems at once and combining the answers. One matches exact strings, the other matches meaning, and production systems use both because neither is sufficient alone.
Do keywords still matter with AI search?
Yes. The keyword half of hybrid retrieval is still running. Exact product names, version numbers and figures need to appear verbatim in your text.
Why not just use semantic search?
It handles exact strings poorly. Ask about a specific SKU or error code and pure semantic retrieval returns something topically close but wrong.
How do I write for hybrid search?
Cover the concept thoroughly for the semantic half and include exact names, numbers and phrasings verbatim for the keyword half.
Keep reading
Related on LymLyt
Beyond LymLyt
Further reading
Want this working on your site?
We build the content behind the term, ranked in search and cited by AI.
Book a 30-min call →