OpenSearch adds S3 ingestion that makes document images searchable

Amazon OpenSearch Serverless now offers intelligent ingestion for logs and documents stored in Amazon S3. AWS’s October 22 release says users select a source bucket and a collection index, then the service detects data types, parses content, creates vector embeddings and indexes the results. The feature is available for OpenSearch Serverless NextGen collections, rather than every OpenSearch deployment. [AWS describes the ingestion workflow here](https://aws.amazon.com/about‑aws/whats‑new/2026/10/amazon‑opensearch‑intelligent‑ingestion‑logs‑documents‑s3). The addition is relevant to teams building search over document stores because it combines steps that otherwise need their own processing workflow. Its AI‑specific role extends beyond extracting text: the service can describe images inside documents and index those descriptions with the surrounding material.

For unstructured files, [AWS lists](https://aws.amazon.com/about‑aws/whats‑new/2026/10/amazon‑opensearch‑intelligent‑ingestion‑logs‑documents‑s3) PDF, Word, plain text and Markdown support. The service parses headings, reading order and tables. When a document includes charts, diagrams, screenshots or photographs, a vision‑language model produces text descriptions that can be searched alongside the document’s other content. That creates a way for a text search to find information represented visually. A generated description should still be understood as an interpretation of the image. It is not a guarantee that every chart label, relationship or detail will be captured correctly. Applications presenting results should preserve the connection to the source document so a reader can examine the underlying material.

An ingestion job handles a one‑time load, such as a migration or backfill. An ingestion stream keeps an index updated as new objects arrive in the bucket, according to [the release](https://aws.amazon.com/about‑aws/whats‑new/2026/10/amazon‑opensearch‑intelligent‑ingestion‑logs‑documents‑s3). Choosing between them depends on whether the source is a fixed collection or a continuing flow of documents. The distinction matters when reviewing completeness. Finishing a one‑time job does not, by itself, establish that later uploads have been added. A stream is intended for continuing arrivals, but an application still needs to check the documents and search results relevant to its own workflow. AWS’s announcement is a description of service behavior, not an OpenTools ingestion test.

Intelligent ingestion also supports JSON, NDJSON and JSON array formats for logs and structured data. AWS says it applies transformations and can optionally add embeddings without pipeline code. Source S3 buckets may be in the same AWS account as the collection or a different account. [These supported source and format choices are listed together](https://aws.amazon.com/about‑aws/whats‑new/2026/10/amazon‑opensearch‑intelligent‑ingestion‑logs‑documents‑s3). A team should decide which data needs semantic retrieval and which fields already answer its search questions directly. Embeddings, parsed document text and generated image descriptions represent different kinds of information. Their presence in one index does not make the evidence equally precise or turn every field into an observed fact.

[AWS says](https://aws.amazon.com/about‑aws/whats‑new/2026/10/amazon‑opensearch‑intelligent‑ingestion‑logs‑documents‑s3) the feature is available in seven AWS Regions for NextGen Serverless collections with S3 as the source. The announcement does not name those Regions in its body. Confirm the exact Region and collection support in the service documentation or console before planning a rollout. Capacity scales up to the collection’s configured limit, and AWS says charges apply to active processing. The release provides no workload‑specific price or comparison with a separately managed pipeline. The useful evaluation is whether the supported formats, source pattern and searchable output fit the documents a team actually needs to retrieve, followed by a review of the actual processing and retrieval results. *Photo: Tedder / [Wikimedia Commons](https://commons.wikimedia.org/wiki/File:Amazon_AWS_us‑west‑2_morrow_east_AZ_01.jpg), [CC BY‑SA 4.0](https://creativecommons.org/licenses/by‑sa/4.0/). AWS Oregon data‑center site, August 2023; context photograph, not evidence of where a customer request is processed. Source‑hosted image with responsive display cropping; image adaptations retain the same license.*