AWS announced the addition of metadata prefiltering to Amazon S3 Vectors, allowing conditions such as tenant, category, status, or date to be applied before similarity search is performed, rather than checking the filter in parallel with result selection. The company says this change may provide up to five times as many matching vectors in queries with selective filters.
The feature targets applications that do not search an entire index but need to restrict results to a specific scope, such as a single customer's documents in a multi-tenant system, a particular user's files, or content permitted in a specific region and under a specific license. The use cases AWS mentions include retrieval-augmented generation (RAG) knowledge bases, agentic applications, document search, and financial research and legal services platforms.
What Changes in Practice?
In ENHANCED mode, the service resolves the filter first and then performs vector search only within the matching vector set. CLASSIC mode, which is used by default in existing indexes, performs search and filtering concurrently, which may reduce the number of matching results when the required set represents a small portion of the index.
AWS illustrates the difference with an example of an index containing 8 million support tickets, 400 of which belong to a single customer. In enhanced mode, the search is performed within that customer's tickets instead of selecting candidates from the full index and then filtering the results. In another example, updating an index changed a query restricted to one tenant from just two results out of the ten requested to a complete result set from that tenant's documents.
Capabilities and Limitations
Each vector can carry up to 2 KB of filterable metadata, while a single query supports up to 100 filter conditions. The query language includes equality, Boolean logic through $and and $or, numeric comparisons, set membership, and existence checks. AWS also added the $startsWith operator to restrict searches to paths, URLs, or hierarchical keys, such as searching within a subfolder or a document identifier prefix.
Existing indexes do not require vectors to be re-ingested, and queries do not need to be changed. However, users must invoke UpdateIndexMode to convert an index to ENHANCED. New indexes in vector buckets created on September 30, 2026, or later use enhanced mode, while older buckets and their indexes remain on CLASSIC until the user changes the setting.
Why Does This News Matter?
The practical value lies not only in increasing the number of results, but also in reducing the likelihood that vector search will obscure the correct documents because of the narrow scope required. This is important for multi-tenant systems and agents that must adhere to a defined document set, because retrieval quality here depends on combining semantic relevance with scope correctness.
The feature is available at no additional cost in all AWS commercial Regions where Amazon S3 Vectors is available, as well as in AWS Regions in China, while storage fees and PUT request and query charges continue to be calculated according to the usual Amazon S3 Vectors pricing. The limits on the number of conditions and metadata size remain factors to consider; AWS suggests combining large filters or splitting them into smaller queries when exceeding the 100-condition limit.