Building AI apps where each user has their own private data? HNSW indexes weren't designed for that; they're built for massive shared datasets, not thousands of smaller per-user collections. 💫 Introducing Flat Indexes for MongoDB Vector Search: Optimize multitenant AI apps with one parameter change, "indexingMethod": "flat", that delivers better recall, less memory overhead, and zero cross-tenant noise. Available now in Public Preview on Amazon Web Services (AWS), Google Cloud, and Microsoft Azure → https://lnkd.in/dJW9kfjs
per tenant vector search is a real pain. tried splitting hnsw by user in pgvector and the build cost ate me alive. flat might actually be the move
Veena C. + Chris Kabat ^ just saw this and thought about our conversation yesterday.
Gregory Badaire Florent Tragni
Trevor nailed it — HNSW at per-user granularity is a memory and recall nightmare. The flat index approach makes a lot of sense for AI apps being built in markets like Malaysia and SEA right now, where the pattern is hundreds of SME clients on a shared SaaS platform, each expecting their own AI that knows only their data. Cross-tenant bleed on vector search in those setups is a real commercial risk, not just a technical inconvenience. Flat indexing with proper partitioning solves that cleanly without rebuilding the entire infra.