Services Work Pricing Blog About Book a free call →

S3 Vectors hybrid search: what changes for RAG

AWS added full-text and hybrid search to S3 Vectors on November 16, 2026. What changed, what to know first, and how to test it on your own queries.

On November 16, 2026, AWS announced hybrid search for Amazon S3 Vectors: one QueryVectors call can now combine vector similarity with BM25 full-text search. If your RAG system misses answers that hinge on exact terms such as error codes, SKUs or citations, this is worth a test. If your queries are purely conceptual and retrieval already works, you can ignore it.

What was announced, and what is different?

Before this, S3 Vectors ranked results by vector distance only. Now the same index can also hold searchable text. You write the text as a metadata field with PutVectors, and S3 Vectors indexes it with BM25. A query can then be one of three types, per the hybrid search documentation:

  • Vector-only: you send a query vector and no query text.
  • Full-text-only: you send query text and no vector. Results are ranked by BM25.
  • Hybrid: you send both. S3 Vectors fuses the two rankings with Reciprocal Rank Fusion (RRF) and returns one list.

AWS says it keeps subsecond latency for infrequent queries and about 100 ms or less for frequent ones. That is a vendor figure, so measure it on your own data.

Who should care, and who can ignore it?

Care if your corpus has many identifiers: product codes, error codes, legal citations, drug names, file names or acronyms. The documentation says pure vector search "can return less relevant results" for queries that hinge on specific terms, for example a chunk containing "E1234" can be missed for the question "what does E1234 mean?" (an illustrative case, based on the error-code example in AWS's own query documentation).

Care too if you already store agent memory or documents in S3 Vectors and run a second system only for keyword search.

You can ignore it if your queries are mostly open-ended questions and you are happy with recall.

What does it cost, and what are the limits?

As of November 2026, the announcement does not give a separate price for hybrid search and points to the S3 pricing page. That page lists S3 Vectors charges for storage, uploads and queries and does not mention full-text or hybrid search in that section. We would not assume a price: run a test and read the bill.

The documentation lists these constraints (see configuring full-text search):

  • Full-text and hybrid search work on ENHANCED indexes only. Indexes in vector buckets created before September 30, 2026 default to CLASSIC.
  • Changing an index to ENHANCED is, per the index mode page, done in place, with no re-ingestion and no additional charge.
  • The limits page allows up to 10 full-text searchable fields per index, 100 KB per field, up to 10 text fields searched per request, and 1,024 characters per query text entry.
  • Only string and list-of-string values are indexed for full-text search.
  • You cannot change an existing text configuration, such as its analyzer. You can add or remove fields.
  • Turning on full-text search for existing data starts a background backfill. The documentation says it can take minutes to hours, and queries on a field are rejected while it is BUILDING.
  • Only one UpdateMetadataConfiguration request runs at a time, so enable all fields in one request.

There are 10 built-in language analyzers, including English and French. The default analyzer works across most languages. If your content is French, test the French analyzer, because stemming and stopword removal change which queries match. The analyzer cannot be changed later, so choose it before you backfill.

How can a team try it safely this week?

  1. Build a small query set. Collect 50 to 100 real questions, label the chunks that should come back, and split them into term-driven and conceptual queries.
  2. Copy, do not migrate. Create a new ENHANCED index in a test bucket, or convert a copy, and enable full-text on the text field only. Put large search-only fields in nonFilterableMetadataKeys when you create the index.
  3. Compare three runs. Run vector-only, full-text-only and hybrid on the same set, with returnScores on. Count how often the right chunk is in the top K.
  4. Try your own fusion. The querying documentation notes you can run the two queries separately and combine them in your application if you want your own re-ranking.
  5. Check latency and the bill. Record p50 and p95, then compare request and storage charges with your current setup.

What would we check first?

We start with the evaluation, not the feature. A hybrid query can lift recall on term-heavy questions, but AWS itself recommends testing with representative data and queries, and says gains show most when the query mix is varied. So we would check:

  • Whether your misses are lexical (exact terms) or semantic. Hybrid mainly helps the first kind.
  • Whether the text you index is the chunk you want matched, since BM25 scores the field you configure.
  • Whether the analyzer matches your language.
  • Whether your filters stay within the documented limits once you move to ENHANCED: the index mode page asks you to confirm queries use 100 or fewer filter constraints.
  • Whether the built-in RRF fusion beats your own weighting on your data.

If retrieval quality is a bottleneck in your product, our services include scoping and building RAG systems with cost controls and observability, and a fractional CTO can own the decision.

General information only, based on vendor documentation as of November 2026; check current AWS documentation and pricing before you decide. This is not financial or professional advice.