The Problem

The client operated a massive, high-velocity laptop catalog where traditional keyword search was failing. Queries like "something lightweight for heavy video editing under 1.5kg" returned irrelevant results: the search engine had no concept of intent, only tokens. Low engagement and missed conversions were the direct result.

What I Built

An end-to-end RAG pipeline that moves beyond keywords to understand the deep context of every query. The system extracts structured intent from natural language, retrieves semantically relevant products, and streams a conversational response, all under 200ms.

Technical Highlights

Intent Extraction Used Gemini Flash to map raw natural language into structured filters, including price range, GPU class, weight threshold, and use case, before ever touching the database. Semantic Retrieval Implemented Sentence Transformers (e5-base-v2) generating dense embeddings stored in pgvector. Queries are embedded at runtime and matched via cosine similarity, not keyword overlap. High-Speed Serving Built the backend on FastAPI with Redis caching for hot catalog entries. Streaming responses via SSE keep perceived latency low even on complex multi-filter queries. Production Hardening Deployed as a Dockerized microservice with structured logging, request-level auth, and connection pooling tuned for high-concurrency traffic.

Stack

FastAPI · PostgreSQL + pgvector · Redis · GPT-4o · Gemini Flash · Sentence Transformers · Docker · Python

Result

• Sub-200ms end-to-end latency for real-time streaming responses
• 99.5% uptime in production
• Transformed product search into a conversational advisory experience: users describe needs, the engine recommends with reasoning