AI search engine for product catalogs
One hybrid search (vector plus text) over products synced from four backend systems, serving three frontends. Built with RAG for an electronics client.
- Python
- FastAPI
- LangChain
- OpenAI embeddings
- MongoDB vector search
- Redis
What I built
A centralized AI search engine using Retrieval-Augmented Generation, shipped in 2024. It keeps products in sync from four backend systems, serves three frontends through one search API, handles complex product catalogs with many variations in size, model, memory and storage, and solved the client's main search problem.
How it works
- Product data is ingested and synchronized in real time from four backend systems into a centralized MongoDB-backed vector database.
- OpenAI’s embedding model generates the vector representations used for semantic search.
- Hybrid search combines MongoDB vector search with text search, with Redis feedback-based re-ranking.
- Modular FastAPI search APIs serve three frontend systems, with responses in the millisecond range.
- LangChain orchestrates the RAG workflows, and a Redis cache keeps responses fast.
Impact
Search was the client’s biggest product-discovery problem, and this solved it. The business saw strong growth after launch. I did not have a baseline or access to the sales figures, so I don’t quote a percentage.