How does the solution isolate one user's documents from another's?

It uses Amazon Cognito to authenticate users and derives the user's identity on the server from the validated JWT. Every retrieval applies an explicit filter on user_id, built from the verified identity, so users can only access their own documents.

What file size limits apply to direct ingestion?

Files up to 6 MB are sent inline in the API request. Larger files, up to 50 MB for text, are uploaded to Amazon S3 and ingested by reference.

How long does it take for a document to become searchable?

For plain text, about 2 to 3 seconds. For PDFs, text becomes queryable in 5 to 30 seconds (TEXT_INDEXED), and full multimodal indexing takes about 90 seconds. Times vary by size and load.

Large Language ModelsAI Business & IndustryAmazon AI BlogAug 31, 2026Sep 1, 2026, 04:00 JST

AWS launches multi-tenant agentic document chat on Bedrock

Send on LINE

3 Key Points

What happened
AWS published a reference architecture for building multi-tenant agentic document chat applications on Amazon Bedrock Managed Knowledge Base. The solution handles ingestion, retrieval, and isolation for each user's uploaded documents.
Why it matters
Enterprises can avoid building and securing the underlying retrieval infrastructure themselves. The service enforces per-user isolation on every retrieval hop, using verified identity from Amazon Cognito rather than client-supplied values.
What to watch
Direct ingestion through the custom connector makes documents retrievable within seconds. Files up to 6 MB are sent inline; larger files (up to 50 MB for text) are staged to Amazon S3. The repository lets you deploy the code in your own account.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, in your inbox every morning.

Get news like this every morning, free

Context & Analysis

The post addresses a common scaling pain point: building a document chat feature where each tenant's data is strictly isolated. AWS shifts the heavy lifting to Bedrock Managed Knowledge Base, which owns ingestion, vector indexing, ranker, planning, and final response generation. The application only needs to handle upload UI, authentication, and per-user logic.

A key design choice is direct ingestion via a custom connector instead of the S3 connector. This avoids scheduled syncs that could overwrite user uploads and allows assigning custom document IDs, enabling in-place updates. The architecture also handles large files by staging them to S3, keeping the upload endpoint responsive via SQS decoupling.

The indexing lifecycle distinguishes between TEXT_INDEXED and INDEXED. For PDFs, text becomes queryable first, while multimodal elements (images, tables) require full indexing. This matters for UX: showing a document as ready at TEXT_INDEXED gives faster feedback but may miss some content in PDFs. The performance figures are only reference, not guarantees.

FAQ

How does the solution isolate one user's documents from another's?: It uses Amazon Cognito to authenticate users and derives the user's identity on the server from the validated JWT. Every retrieval applies an explicit filter on user_id, built from the verified identity, so users can only access their own documents.
What file size limits apply to direct ingestion?: Files up to 6 MB are sent inline in the API request. Larger files, up to 50 MB for text, are uploaded to Amazon S3 and ingested by reference.
How long does it take for a document to become searchable?: For plain text, about 2 to 3 seconds. For PDFs, text becomes queryable in 5 to 30 seconds (TEXT_INDEXED), and full multimodal indexing takes about 90 seconds. Times vary by size and load.

Amazon AI BlogRead Original Article

Send on LINE

AI news that matters for your work, delivered every morning.

Pick your industry and the AI tools you use, and get news related to your work every day.

Get it free

Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.