Skip to main content

​
Custom knowledge

Custom knowledge is reference material you give an agent directly — policy documents, how-to guides, FAQs, recordings — organized into collections the agent can search and cite in its answers. It complements integrations: integrations reach live systems; custom knowledge covers the documents that don’t live anywhere the agent can already search. Knowledge is per agent: each agent has its own collections, on its Capabilities → Knowledge tab.

​
Collections

A collection is a named group of items with a shared purpose (“HR policies”, “IT runbooks”). Create one with a name and optional description. Two switches control what the agent actually sees:
  1. The collection’s Active switch — collections start inactive (draft), so you can assemble and review content before any of it goes live.
  2. Each item’s Active toggle — new items also arrive inactive.
An item is searchable only when both are on; the page shows a “Live to agent” / “Draft — hidden” pill so there’s never doubt about what’s visible.

​
Adding content

Two ways to add items:
  • Upload files — drag and drop PDF, DOCX, TXT, Markdown, and audio or video files (MP3, M4A, WAV, MP4, MOV, WebM), up to 200 MB each. Embedded images and audio are extracted automatically, so a slide deck’s figures or a recording’s speech become searchable too. Uploads process in the background — you don’t have to keep the tab open — through extraction, chunking, and embedding, with a live status on each item (and clear failure states if a file has no extractable text).
  • Write a document — author Markdown directly in the console. Authored items can be reopened and edited any time.
The corpus table shows every item with its type, size, status, and toggles; bulk-select to activate, deactivate, or delete many at once. A search bar above the collections searches the corpus itself — useful for checking what an upload actually contains.

​
Citable vs. background

Each item is either:
  • Citable — the agent can quote it and shows it as a source (a citation chip / source card in the answer, including timestamped links into audio and video).
  • Background — the agent can use it to inform an answer but never quotes or cites it. Useful for internal context you want reflected but not surfaced.
Set a per-collection default for new items, and override per item.

​
How the agent uses it

When a question touches your material, the agent searches the active corpus (combining keyword and semantic matching), reads the most relevant passages, and — for citable items — attaches numbered citations the user can open. If a snippet isn’t enough it can read the full document, and it can pull up embedded media (an image from a PDF, a moment in a video) when that’s the real answer.

​
Indexing integration content

Custom knowledge is for material you upload. To make an integration’s own content searchable — SharePoint sites, Confluence spaces, a knowledge base — use the integration’s Index tab instead: indexed content is searched together with the live system and every result is permission-checked against the source as the asking user, so nobody sees a document they couldn’t open themselves. Each indexed resource shows its sync status, schedule, and history, with sync-now, pause, and disable controls.