Amazon S3 Vectors now supports metadata pre-filtering for higher recall on filtered searches

Summary

• Amazon S3 Vectors introduces metadata pre-filtering to evaluate filter conditions before similarity search, improving recall on filtered queries.
• Users can filter on attributes such as tenant, category, status, or time, with support for prefix matching using $startsWith for paths, URLs, and hierarchical keys.
• Each vector supports up to 2 KB of filterable metadata, and a single query allows up to 100 filter constraints.
• The feature incurs no additional cost, requires no data re-ingestion, and involves no changes to existing query syntax.
• Pre-filtering applies to ENHANCED index mode, where S3 Vectors resolves filters first before searching matching vectors, whereas CLASSIC mode evaluates filters in tandem with the search.
• Existing indexes default to CLASSIC mode until updated to ENHANCED via the UpdateIndexMode action.
• On highly selective filters, pre-filtering returns up to 5x more matching vectors compared to CLASSIC indexes.
• Common use cases include legal services filtering by matter number or folder path, financial services filtering by issuer or document type, media filtering by content rating and region, and agentic applications filtering by owner or timestamp.
• Creating an index requires specifying the index name, vector bucket, dimension matching the embedding model, and distance metric.
• Vectors are written using the PutVectors API with up to 2 KB of metadata attached to each vector.
• Queries use the QueryVectors API with a compact JSON syntax supporting equality, $and, $or, $gt, and $startsWith operators.
• Prefix matching with $startsWith allows scoping searches to specific subtrees for document IDs or hierarchical keys.
• Enabling pre-filtering on existing indexes takes effect immediately without re-ingesting vectors or changing queries.
• Indexes created in buckets established on or after September 30, 2026, default to ENHANCED mode; older buckets default to CLASSIC until the bucket default is set.
• A single query supports up to 100 filter constraints, and exceeding this limit may require consolidating filters or splitting queries.
• Metadata pre-filtering is available at no additional cost in all commercial AWS Regions and AWS China Regions where Amazon S3 Vectors is available.
• Standard S3 Vectors pricing applies for storage, PUT requests, and queries.