Turbopuffer introduced native filtering for vector search—achieving >90% recall in 25ms, even under complex filter conditions, without sacrificing scalability. In traditional vector databases, filtering is a major bottleneck. You typically get two bad options: • Pre-filtering — filter by metadata first, then compute vector distances. Great recall (100%), but latency explodes (e.g. 10s). • Post-filtering — run ANN search, then discard results that don’t match the filter. Fast, but terrible recall (often 0%). Turbopuffer solves this with native filtering—tightly coupling metadata filters into the clustering-based vector index itself. Here’s how: 1.Cluster-Aware Filtering Instead of treating filters and ANN as separate concerns, Turbopuffer rewires its attribute index to understand vector index internals—specifically, SPFresh-inspired clustering. This allows skipping entire clusters that don’t contain any valid matches. 2.Efficient Addressing Each document is stored using a {cluster_id}:{local_id} scheme. Attribute indexes map directly to these addresses—so filter lookups immediately resolve to cluster-local candidates. 3.Two-Level Indexing •Row-level: Maps each attribute value to specific documents within clusters. •Cluster-level (Downsampled): Maps attribute values to clusters with matches, using compressed bitmaps to minimize roundtrips. This hierarchy allows fast, coarse filtering (which clusters to check) and then precise filtering only when necessary. 4.Optimized for Object Storage Cold queries from blob storage are fast because: •Only a small number of roundtrips are needed. •Index data is compact (thanks to bitmap compression). •Updates avoid full file rewrites via an LSM-based storage layer. The result? A filtered vector search system that’s faster than traditional pre-filtering, much more accurate than post-filtering, and scales without needing index rebuilds. Native filtering like this makes @turbopuffer particularly compelling for multi-tenant RAG, document retrieval with access control, and codebase-wide semantic search.