Search & Personalisation
Product Discovery Engine
Turns a natural-language shopping query into structured filters, with a trace of where each value came from.
The Problem
Shoppers type “black running shoes under 8000”, not filter objects. Search needs query understanding before ranking.
How It Works
- Match category by longest keyword, plus colour, brand and use-case dictionaries.
- Parse price phrases: under/over/between, with ₹, rs and “k” suffixes.
- Emit structured filters with an extractedFrom trace and validate them (min > max is an error).
Technically Interesting
Every extracted field records its source text, and a validation layer rejects contradictory output — the same layer you would put around an LLM extractor.
Why It Matters
It is the front door to the BM25 search engine: messy language in, validated structure out.
Scope & honesty. Rule-based extraction over a fixed vocabulary; no LLM.