
A natural-language product search that deliberately keeps the LLM out of retrieval. The model parses queries like “gaming laptop under 80k with good reviews” into a strict, server-validated JSON constraint schema (constrained JSON-mode, temperature 0), while Typesense does the actual retrieval via hybrid search — 80% keyword + 20% semantic vectors in one query. The LLM is a translator, not the engine: fast, injection-proof, and it falls back to plain search instantly if the model fails.
- Next.js
- Typesense
- Groq
- LLM



