AI real estate search: natural-language queries across 1.4M listings

The situation

• A Sydney real estate startup wowed investors with a Cursor-built demo of natural-language property search, 'show me Federation-style 3-bed homes under $1.4M in inner-west suburbs with a north-facing yard'.
• The demo used hand-curated results.
• Behind the scenes, search was Postgres ILIKE on listing titles.

What we did

• Natural-language filter extraction with strict tool-call schema
• pgvector embedding search across 1.4M listings with HNSW index
• Hybrid ranker blending structured + semantic + lexical scores
• Inngest re-index pipeline running every 6 hours
• Eval harness with 200 labeled queries running on every change
• Edge search API with sub-400ms latency globally

The result

• The public launch hit 90,000 unique searchers in week 1.
• The investor follow-on round was secured 6 weeks after launch. The natural-language feature now drives 38% of search traffic, with a 22% higher inquiry-per-search rate than traditional filter search.