Knowledge & RAG
Knowledge Collections let users ground AI answers in their own documents. A collection is a named, user-owned set of files that exists independently of any single chat and can be referenced from any thread. Files are automatically indexed for retrieval, so models answer from your content instead of guessing.
System Design
File ↑ Comment (model-visible) Mode Tokens
architecture-backend.md — RAG 3k
architecture-frontend.md — RAG 4k
sync-jobs.md Background Sync Jobs RAG 1k
A collection groups related files; each is indexed for retrieval and referenced by name in chat.
How it works
Section titled “How it works”- Create a collection with a name and description, and add files to it.
- Files are auto-extracted and embedded into a vector database when added.
- In a chat, reference a collection (via a
#trigger or the toolbar) to make its contents retrievable. - At response time, the model retrieves the most relevant passages from the collection to ground its answer. Presets can also inject a document’s full content when full-context grounding is preferred over retrieval.
Sharing and permissions
Section titled “Sharing and permissions”Collections are shareable with fine-grained access. Each grant carries a permission level:
- Read — use the collection in chats.
- Edit — add or remove files.
- Manage — control sharing and settings.
You can share with individual users, groups, or teams, and team and group membership is resolved live from ControlTower. An organization-wide grant is capped to read access.
Helpful touches
Section titled “Helpful touches”- Auto-generated file descriptions give models a short, useful summary of each file.
- Cascade delete removes a collection’s files, stored blobs, and index entries together, so nothing is left behind.
Related
Section titled “Related”- Chat App — reference collections in any conversation.
- Cowork Agent — give the agent grounded context.
- ControlTower — the teams and groups you share with.