Search That Understands Intent
Search and discovery rebuilt around what customers actually type, blending keyword and semantic matching, with relevance you can measure rather than argue about.
- To first release
- 2 weeks
- Relevance baseline
- Measured
- OpenSearch
- pgvector
- Claude
- TypeScript
- Next.js
The problem
Site search is usually the highest-intent surface on a product and the least invested in. The symptoms are familiar: no results for plural forms or synonyms, the best-selling item buried on page three, and a team unable to say whether last month of tuning made anything better because there was never a baseline to compare against.
How we would sequence it
- 01Mine your own search logs for the queries that fail, ranked by how much traffic they carry
- 02Blend lexical and semantic retrieval rather than replacing one with the other
- 03Build a relevance judgement set from real queries, so improvements are measurable
- 04Handle synonyms, misspellings and category intent explicitly rather than hoping the model infers them
- 05Ship behind a flag and compare against the current engine on live traffic
This solution describes our method for a class of problem. It is not a record of a completed client engagement, and the figures above are scope estimates rather than measured results.
Have this problem?
Scope this with us