Knowledge sources

Knowledge & Data is the single place to manage everything your agents can read: live MCP connections, uploaded documents, and synced data connectors, in one unified list.

Open Organization Settings > Knowledge & Data. Three tabs partition the list: All Sources, Live (MCP), and Indexed (RAG), each with a live count. Live sources are queried at request time; Indexed sources are chunked and embedded for retrieval. Every row shows Name, Type, Mode (Live, Indexed, or Both), Status, Count, Last Updated, and Actions, with name search and a status filter on top. Unifying both kinds of source in one list means agents miss nothing, and admins confirm what's connected without hunting across separate MCP and RAG screens.

Knowledge & Data page with live MCP sources and indexed RAG documents in one table
↳One table for everything agents can read: live MCP connections beside indexed documents and connectors.

Adding sources

  1. 1Upload Files: drag and drop or pick files. Uploads are chunked and indexed; each file is capped at 50 MB.
  2. 2Connect Source: pick a source type. MCP servers connect live; SharePoint and OneDrive connect as syncing connectors with Azure app credentials (tenant ID, client ID, client secret), a site URL or target user, and an optional folder path.
  3. 3Other connector types listed (Google Drive, Confluence, Notion, Git Repository, Amazon S3, Web Crawl) are not yet wired; the dialog says so and offers Upload Files Instead.

Sync behavior and failure

Connectors sync in the background. While a sync runs, the page polls so status and document counts update live, and the row's button reads Syncing. A failed sync surfaces the real failure reason on the row, not a generic error. A completed sync may report skipped files; sub-folders are not recursed, so a sync that completes with 0 documents usually means a wrong folder path or file ownership. Use Sync now to re-run a connector on demand.

The Connect Source modal over the Data admin page: MCP server, SharePoint, and OneDrive source types with the credential fields scoped per source
↳One modal, every source type: the credentials stay scoped to the connection, and nothing reaches agents until it lands in this governed list.

Per-row actions

  • →MCP rows: Test (connect and list discovered tools), Edit, Delete.
  • →Document rows: Delete. Ingestion status runs pending, processing, completed, or failed, with chunk counts per document.
  • →Connector rows: Sync now and Disconnect.

The Type column on document rows records provenance: File Upload, Swarm Generated, Agent Generated, or Generated. Connector-synced files roll up under their connector instead of appearing as standalone rows.

Storage limits

Org admins see a File Upload & Storage Limits panel at the top of the page: max upload size for non-media files (MB), a separate max for audio/video (MB), and a total storage cap (GB, blank for no limit). Org values are bounded by a platform hard maximum shown next to the field.

Managing sources requires the owner, org admin, or platform admin role; other members see the list read only. Connector credentials are stored encrypted and are never returned by the API after you save them.