AI Search

Natural language search powered by vector embeddings + LLM re-ranking. Works for both properties and listings.

How It Works

"Pet-friendly villas in Goa for 6 adults this weekend under 20k"
  │
  ├─ 1. Embed query → text-embedding-3-small → float[1536]
  ├─ 2. Vector similarity → pgvector cosine → top 50 candidates
  ├─ 3. Geo anchor → Google Places API → Goa coordinates
  ├─ 4. LLM re-rank → GPT ranks by relevance + extracts filters
  │
  └─ Result:
     listingIds: [ranked by relevance]
     checkInDate: "2026-04-05"
     checkOutDate: "2026-04-07"
     adultCount: 6
     city: "Goa"
     propertyType: "Villa"
     budgetPerNight: 20000
     petFriendly: true

Two Search Types

	Property Search	Listing Search
Table	property_embedding	listing_embedding
Returns	Property IDs	Listing IDs + extracted filters
Combines with filters	No	Yes (merges with rawParams)
Endpoint	/api/v1/search/properties/search/ids	/api/v1/crs/listings/ai-search

Caching

Same query on the same day returns cached results (no token burn):

Cache key: "2026-04-03:pet-friendly villas in goa this weekend"
TTL: 10 minutes
Max entries: 5000

Date is included in the key so "this weekend" resolves correctly across days.