Rethinking Search: Understanding User Intent at the Core
From Openmart engineering team
Quick background on Openmart 1.0:
We’ve indexed all local businesses on the map, and users can search for any type of business through a simple map-based interface. People often enter all kinds of unstructured queries when searching on our platform.
We did some iterations and eventually figured that we should build a light-weight Intent‑Detection Agents.
We use lightweight AI agents to figure out what a customer really wants before we run a single database query. Every search phrase is first classified
Common Issue
Users mix brands, categories, and places in the same search bar. Without intent detection, we waste time (and API calls) fetching the wrong data.
How our agent pattern works
Ultra‑fast intent classifier (o4‑mini)
Labels each query as name, type, location, or a combination—usually in under 10 ms.
Task‑specific fetchers
Brand matcher hits our canonical business‑entity index.
Category search fans out best category relevancy
Geo filter narrows results to the user’s city or map viewport.
Result
Whether someone types a brand (“Target”), a service (“plumber”), or a location‑tinted phrase (“pizza near Brooklyn Bridge”), our intent agents route the query to the right backend instantly—cutting latency, trimming costs, and surfacing the most relevant results the first time.
Impact:
Customers are way more happier and we didn’t see a lot of complaints in search after implementing this small hack.
reach out to kathryn@openmart.com for a discussion