SageMaker Catalog: Governed Unstructured Data Becomes Queryable by Conversation

What happened

AWS published the consumer side of an unstructured-data workflow on 23 September 2026.

In this approach, users search SageMaker Catalog for assets using metadata previously extracted from PDFs, images, emails and other unstructured content.

The user subscribes to an S3-backed asset and waits for approval from the data producer. Once access is granted, the content can be queried through either:

1. An Amazon Bedrock chat-agent application.
2. Amazon Bedrock model inference inside a notebook.

The earlier producer workflow uses Amazon Textract and a foundation model to extract information and enrich the asset’s metadata before it is published in the catalogue.

Why it matters

Many document-questioning demonstrations concentrate on the conversational interface while giving less attention to ownership, approval and access controls.

This design places discovery and approval before model interaction.

That brings unstructured content closer to the governed data-product model already used for structured analytical assets.

Important risks remain:

• document-level permissions;
• sensitive-information handling;
• source attribution;
• answer accuracy;
• access revocation;
• prompt injection;
• incomplete or misleading retrieval.

Catalogue approval does not automatically guarantee that every generated answer is correct or properly grounded.

Who it affects

Data-governance teams, AI engineers, knowledge-management teams and organisations building assistants over documents, images and other unstructured assets.

What to do next

Start with a small document collection that has clear owners and known answers.

Evaluate:

• metadata quality;
• access enforcement;
• document-level filtering;
• citation fidelity;
• retrieval accuracy;
• prompt-injection exposure;
• model and storage costs;
• access-revocation behaviour.

Generated answers should retain traceable links to the documents used as evidence.

Signal Take

The significant development is not simply conversational access to documents. It is conversational access after discovery, ownership and approval have been made explicit.

Scope and limitations

This analysis interprets the linked AWS material. Recommendations are editorial judgement, not tested implementation instructions. No independent product benchmark or deployment test was performed.