Venue Search Recall

Venue search recall measures the percentage of relevant venues returned in the top-10 results for a given intent query. We define 'relevant' using a manually curated evaluation set of 200 query-venue pairs across 35 intent patterns (date night, quick lunch, sunset drinks, coworking, etc.). Each query has 3-8 ground-truth relevant venues annotated by local experts who know the Bali venue landscape.

The 87% recall was achieved after tuning the pgvector HNSW index parameters: ef_construction=128, m=16, ef_search=64. Before tuning (with default IVFFlat), recall sat at 72%. The improvement came from two changes: switching to HNSW for better approximate nearest neighbor performance, and enriching venue embeddings with intent-tagged descriptions rather than raw venue names alone.

We track recall separately from precision because for venue discovery, missing a great venue is worse than showing an okay one. Users can scroll past irrelevant results, but they can't find venues that never appeared. Our target is 90% recall by Q3 2026, primarily through adding multi-vector representations per venue (one per intent category the venue matches).