Rebuilding Search Relevance with RAG
A retrieval-augmented search system lifted conversion 28% and cut “no results” pages by 90%.
The challenge
Atlas’s legacy keyword search returned “no results” for 14% of queries — anything phrased differently from the product title failed. Their data science team had a notebook proving semantic search would help, but no path to production, no evals, and no way to measure relevance after launch.
The solution
We deployed a retrieval-augmented search service using pgvector for embeddings, a canary rollout gated by offline relevance evals, and an online A/B framework. We added LLM guardrails for query rewriting and a feedback loop that turned click data into labeled eval examples weekly.
Tech stack
Search went from our biggest complaint to our biggest growth lever. The team built the evals before the model — that’s why it worked.
Related projects
Northwind Capital · Financial Services
99.99% Uptime for a High-Frequency Trading Platform
We rebuilt the deployment and observability backbone of a trading platform processing 4M orders/day, hitting five-nines availability.
Read moreSentinel Mutual · Insurance
Strangling a 20-Year-Old Insurance Mainframe
Incremental strangler-fig migration cut a monolith into 14 services with zero downtime and a 5x deploy rate.
Read moreWant results like these?
Tell us about your system. We'll tell you what's possible.