Amazon S3 Vectors Adds Metadata Pre-Filtering for Higher Recall in Filtered Searches
Amazon S3 Vectors now supports metadata pre-filtering, significantly improving recall for filtered similarity searches. This new capability evaluates metadata filters, including a new `$startsWith` prefix match for hierarchical keys, before performing the vector search, ensuring more relevant results for scoped queries. It is crucial for applications like RAG and agentic systems that require both relevant and correctly scoped results. Existing indexes can enable pre-filtering by updating their index mode to ENHANCED, with no additional cost or data re-ingestion.
Features (1) ›
- Enhanced Metadata Pre-Filtering for S3 Vectors
Amazon S3 Vectors now offers metadata pre-filtering, which evaluates attribute filters before performing a similarity search to deliver higher recall on filtered queries. This enhancement supports prefix matching with `$startsWith` and allows each vector to carry up to 2 KB of filterable metadata with up to 100 constraints per query. Existing indexes can be updated to `ENHANCED` mode to activate this capability without re-ingestion.
https://aws.amazon.com/blogs/aws/amazon-s3-vectors-now-supports-metadata-pre-filtering-for-higher-recall-on-filtered-searches/
Related releases
- Amazon Redshift Adds Support for Apache Iceberg Materialized Views AWS What's New ·
- AWS details Iceberg materialized views in Amazon Redshift for data lake interoperability AWS Big Data Blog ·
- Amazon Bedrock AgentCore Gateway supports private TLS certificates for VPC endpoints AWS What's New ·
- Amazon Redshift now supports cross-Region S3 data lake queries AWS What's New ·
- AWS Glue now offers system-managed write protection for Apache Iceberg materialized views AWS What's New ·
- AWS launches system-managed write protection for Apache Iceberg materialized views AWS What's New ·