Modem swaps pgvector for turbopuffer, cutting p90 search time 76%

Modem moved its embeddings out of a shared Postgres pgvector table and into turbopuffer, cutting p90 search time by 76% and taking most of the load off its database. The shared multi-tenant table had let the smallest customers wait up to ten seconds, and adding an HNSW index made performance materially worse because vector search ignored the organization_id and time filters. By August 2026 the median search took about 0.8 seconds and the slowest 1% took as long as 28 seconds.