Switch it on in a thread
- In a thread, open MCP Servers under the message box.
- Select Knowledge.
- Ask your question. Saying “using our knowledge” or “from our documentation” makes the model reach for the tool rather than answer from its own general knowledge.
What the Knowledge tool does
The model chooses which tool to call, and may call several in one reply — listing first, then
asking. Each call appears in the reply as a step you can expand to see exactly what the tool
returned. Each
ask_assistant call is also saved as a one-question conversation in the console’s
Conversations, so it shows up in your analytics like any other answer.
How a grounded reply differs from an ordinary one
Chat is a general chat workspace, so not every reply comes from your knowledge.
When an answer must be correct to your documentation, check that the reply used the Knowledge
tool — the expandable step shows it — and open the sources. For answers where grounding is
guaranteed on every turn, use the console’s Playground, the website widget or
the HTTP API.
When your content does not cover a question,
ask_assistant returns an abstention rather than a
guess. A reply that says your knowledge does not cover it is the tool working as designed, and a gap
worth filling; Coverage gaps lists them.
What it will not do
- Change anything. It reads with the access of a Member: it cannot add, edit or re-sync a source, change an assistant, or touch a setting.
- Reach another organisation’s knowledge. Your Chat serves one organisation, and the key it uses belongs to that organisation; MITHUNAI derives the organisation from the key, never from anything Chat sends.
- Read as you personally. Everyone in your Chat reads the same knowledge, through one organisation key. Your own role in the console does not widen or narrow it.
- Search the web. Web search is switched off unless your deployment has set it up. See What is switched off.
Files you attach
A file you attach to a message — text, Markdown, CSV, PDF or an image — joins that turn’s context so the model can read it. It is not added to your knowledge and is not searchable by the Knowledge tool. To make content answerable everywhere, upload it as a source in the console.Limits
ask_assistantwaits for the whole answer. It does not stream, so a long answer appears all at once. A call that takes longer than 60 seconds is abandoned.- Answers count. Each
ask_assistantcall uses one of your organisation’s monthly answers; at the limit it is refused, and the reply says so. - The chat model writes the final reply. It can summarise or reword what the tool returned. The expandable step is the record of what your knowledge actually said.
Related
- MCP server — the same tools, from your own AI client.
- Configure an assistant — what
ask_assistantasks. - How answers work