Hotel Entity Extractor (Qwen3-1.7B, LoRA merged)
Fine-tuned to extract structured hotel-search params from natural-language queries
and return strict JSON. Trained filters-free of codes: filters is emitted as
human-readable PHRASES (e.g. "swimming pool, pet friendly"); a downstream matcher
resolves phrases -> production codes.
- Input: a hotel-search query +
today date (DDMMYYYY date math).
- Output: JSON with destination, locality, hotelName, checkin/checkoutDate,
adult/room/child/childAges/infantCount, sortCriteria, min/maxPrice, filters,
deepSearch, isNearMe, resetAction.
- Prompt/schema contract: bundled as
contract.py in this repo. The serving
layer MUST use this exact prompt (it verifies byte-parity at startup).
Serve with vLLM (see the entity-extraction-serving repo). Raw model output is
exposed by the API; date validation + filter phrase->code resolution happen in
the calling service.
