Semantic search for e-commerce

The challenge

Keyword search breaks down exactly when shoppers describe what they need instead of naming the product. The goal was a search that understands intent and respects hard constraints (price, category, attributes) without making results up.

What we built

• Five-stage content pipeline: URL discovery, HTML-to-structured-JSON parsing, LLM summaries, vector indexing in Pinecone, and search.
• Hybrid search: vector similarity plus LLM query enhancement and structured filter extraction from natural language, combined with database filters. Fully async, with fallbacks on errors.
• Human-in-the-loop taxonomies: the LLM proposes a hierarchical classification of the catalog and a person approves it before it goes live.
• Model-agnostic: OpenRouter as the gateway, so switching providers is a config change. Tracing in LangSmith.
• Admin dashboard in React and TypeScript: accounts, pipeline metrics, interactive search and taxonomy approval.
• WooCommerce plugin: semantic search over the store’s catalog, re-indexing in the background whenever a product changes.

Quality

A FastAPI API with auto-generated OpenAPI docs and standard error responses (RFC 7807), around 126 pytest test files, 50+ endpoints covered by API tests, and Sentry monitoring.