Upload knowledge documents to Azure SRE Agent

Your agent captures institutional knowledge during conversations by uploading runbooks, troubleshooting guides, and documentation directly to the knowledge base. When your agent discovers a fix or resolves an incident, it generates structured documents and indexes them for semantic search, so every future investigation benefits from past resolutions.

How your agent captures knowledge

Your agent can upload documents to Knowledge settings during conversations by using the Upload Knowledge Document tool. When your agent discovers a fix, creates a troubleshooting guide, or synthesizes investigation findings, it stores that knowledge directly and makes it searchable for every future conversation.

Create a runbook from the steps we just followed to fix this database
connection pool exhaustion issue and save it to Knowledge settings.

Your agent generates a structured runbook and uploads it in seconds. The document is indexed automatically and becomes searchable for future investigations.

Before and after

The following table compares knowledge capture workflows before and after using the Upload Knowledge Document tool.

What makes this approach different

Unlike manual uploads, your agent proactively creates knowledge. You don't need to remember to document what you learned. Your agent does it as part of the conversation.

Unlike chat history, uploaded documents are indexed for semantic search. When a similar issue occurs months later, your agent finds the relevant runbook automatically through intelligent retrieval, not by scrolling through old threads.

Unlike wiki connectors, uploaded documents don't require external services. The knowledge lives directly in your agent's Knowledge settings, available instantly without syncing delays.

How it works

The Upload Knowledge Document tool accepts three parameters.

When your agent uploads a document:

1. The agent validates the filename and content (maximum 16 MB per file).
2. The agent stores the document in your agent's Knowledge settings.
3. The agent indexes the content for semantic search.
4. The agent confirms the upload with a success message.

Supported file formats

Your agent handles files through three methods, each supporting different formats.

Chat attachments

Drag files into any chat to give your agent immediate context for analysis, including troubleshooting scripts, configuration files, network traces, and more.

Limits: 10 MB per file · 50 MB total per message · 5 files per message

Knowledge settings uploads

Upload files through Builder > Knowledge settings > Add file to persist documents for your agent to reference in future conversations. The system indexes uploaded files for semantic search.

Limits: 16 MB per file · 100 MB per upload

Agent-generated documents

When your agent creates documents during conversations (by using the Upload Knowledge Document tool), it generates .md or .txt files and saves them directly to Knowledge settings. This process happens when you ask your agent to "save this as a runbook."

Example: capture incident knowledge

During an incident investigation, ask your agent:

We just resolved the high CPU issue on web-app-prod. It was caused by a
memory leak in the connection pool. Create a troubleshooting guide from
what we learned and upload it to Knowledge settings.

Your agent generates a structured troubleshooting guide with:

• Scoping steps: How to identify the problem
• Quick mitigations: Immediate actions to reduce impact
• Root cause analysis: What to investigate
• Resolution steps: The fix that worked
• Prevention: How to avoid recurrence

The next time a similar CPU problem occurs, your agent automatically references this document during investigation, turning one engineer's experience into shared team knowledge.

Prerequisites

Before you use the Upload Knowledge Document tool, make sure you meet the following requirements.

Limits

The following limits apply to uploaded knowledge documents.

When to use something else

The following table describes scenarios where a different approach is more appropriate.

Get started

Related content