K-Food RAG: AI-Powered Culinary Discovery

Context

This project explored how a food discovery product could move past static filters and keyword search. The goal was to let users ask culinary questions naturally and still receive grounded, useful results.

Problem

Food search breaks down quickly when the query is descriptive instead of categorical. People search with intent, mood, ingredients, or cultural context, and traditional filters rarely capture that well.

Approach

I designed the product around retrieval first. The system combines embeddings, vector search, structured food records, and language-model output so results stay both flexible and grounded.

What Shipped

• A web interface for natural-language food discovery and guided exploration.
• Backend services for embeddings, vector retrieval, relational lookup, and response composition.
• Queue-based processing for synchronization and retry-safe data operations.
• Monitoring-oriented services to surface sync quality and operational issues across the retrieval stack.

Results

• The product could answer broader food questions than a standard filter UI while staying anchored to real data.
• The underlying architecture became more reliable than a typical AI prototype by accounting for sync, retries, and observability early.
• The system proved out a practical RAG pattern for consumer-style discovery rather than a pure demo experience.

Technical Notes

The build combined Next.js, NestJS, PostgreSQL, Qdrant, Redis, and Google GenAI. The key technical challenge was keeping the AI layer useful without letting it drift away from the structured data model.