Knowledge

Knowledge is whatever an assistant can draw on beyond the model’s own training: a file you attached, a source it can search, a website, a connected system, the live web. They behave differently in scope, in how long they last, and in who else can see them — and choosing the wrong one is the usual reason an assistant answers badly.

The kinds of knowledge

• Attachments — files loaded onto an assistant permanently, available in every conversation with it.
• Knowledge sources — material an assistant searches on demand rather than keeping loaded at all times.
• Web search — live access to the web, optionally restricted to specific sites.
• Integrations — connections to services your organization already uses, such as SharePoint or Google Workspace.
• Custom tools — connections your organization builds to its own systems.
• Legacy knowledge — the older Collections and Websites area, being retired in favor of attachments.

How they compare

Kind	Always loaded?	Where it lives	Typical use
Attachment	Yes	On the assistant	Material the assistant should always be able to draw on
Knowledge source	No — searched on demand	On the assistant	A larger body of reference material
Web search	N/A — live	The public web	Current information not in your own material
Integration	No — searched on demand	The connected service	Material that already lives in another system
Custom tool (MCP)	N/A — called when needed	Your organization’s own system	Actions or lookups beyond ordinary knowledge

Choosing between them

If you’re unsure, start with a file uploaded directly in the conversation. If you find yourself uploading the same document repeatedly, it belongs as an attachment. If the material is too large to attach comfortably, or it’s already indexed somewhere else your organization has connected, reach for a knowledge source or an integration instead.

Access control and knowledge

An assistant’s knowledge is only as available as the assistant itself. Knowledge doesn’t have its own separate audience — anyone who can use the assistant can draw on whatever it’s connected to, so review what’s attached before you give someone access to an assistant, not after.

Why an assistant may not find something

A few common reasons an assistant doesn’t surface something you expect it to:

• The material was uploaded as a temporary file in someone else’s conversation, not attached to the assistant.
• A knowledge source uses a different embedding model than the assistant’s other sources, so it can’t be searched alongside them. See Embedding models and compatibility.
• The source is real but its content genuinely doesn’t cover what you asked.