Pin a knowledge base and Aria or your agent searches it before answering, then cites the passages it used. You can pin knowledge bases in four places, and they add up:
| Where | Applies to |
|---|---|
| The chat (composer chip or Chat settings) | This chat only. |
| An agent (builder › Knowledge) | Every chat, routine and team thread the agent takes part in. |
| A project (project settings › Files) | Every chat in the project. See Projects. |
| A team (team settings) | Every member in the team’s threads. See Teams. |
Pin knowledge bases to a chat
Section titled “Pin knowledge bases to a chat”Click the knowledge bases chip in the composer (for example 3 knowledge bases), or Chat settings › Knowledge bases › Add. The picker lists your knowledge bases with a search box. For each one you see:
- its number of chunks, or Indexing · 40%, Empty, or that it needs a cloud provider’s embeddings to be searched;
- where a pin comes from when you didn’t pin it here: also pinned by Support bot, from project Q4 contracts (these stay checked);
- its privacy badge, Cloud OK or Device only — click it to change it.
The picker also has New knowledge base and Manage.
When to look things up
Section titled “When to look things up”Choose when the pinned knowledge bases are searched:
| Option | What happens |
|---|---|
| Always | Every message is searched, and the best passages are always given to the model. |
| When relevant (default) | Every message is searched, but passages less similar than the cut-off (0.35) are left out. Passages that contain most of the words you asked about count as relevant too. |
| When asked | Searched only when your message refers to your documents, for example “check my notes”, “according to our docs”, “in the knowledge base”, or @kb. |
Use Test search on a knowledge base to see the scores and where the cut-off falls for a question.
Advanced search settings
Section titled “Advanced search settings”Under Chat settings › Knowledge bases › Advanced search settings:
- Passages per knowledge base — how many passages each knowledge base may add (default 4, up to 12).
- Relevance cut-off — the similarity When relevant requires (default 0.35).
- Rewrite follow-up questions for search — on by default. A short follow-up such as “and the second one?” is turned into a full question, using the last few messages, before searching. If the rewrite takes too long, the message is searched as typed.
Steps that explain the result
Section titled “Steps that explain the result”Before the answer, a step shows what each knowledge base returned and what you can do about it:
- Searched Product handbook · 3 passages — Show lists the sources.
- Searched Support macros · nothing relevant — the best match was below the cut-off (for example “Best match was 0.29, below the 0.35 cut-off”). Actions:
- Use weaker matches for this reply — answers again with the passages under the cut-off. On an older reply it applies to your next message.
- Test this search — opens the knowledge base’s Test search with your question.
- Searched Notes · empty — the knowledge base has no sources yet, with Add sources.
- Skipped Legal templates · on this device only — the knowledge base is Device only and you’re answering with a cloud model (“You’re answering with … (cloud)”). Answer this one on-device answers this reply with a ready on-device model.
- An error, when a knowledge base couldn’t be searched.
Citations you can open
Section titled “Citations you can open”When an answer uses passages, they are numbered in the text as [1], [2]… and listed as chips under the reply with the real source titles, for example Pricing & plans 2026.pdf · p. 7 · from Product handbook.
Click a number (or focus it and press Enter) to open the passage card:
- the source title, the knowledge base, the chunk and page, and the match score;
- the passage itself, with the matching part highlighted;
- Open in knowledge base — opens the source in the knowledge base page;
- the web link, when the source is a web page;
- copy.
Press Esc to close it.
Knowledge for an agent
Section titled “Knowledge for an agent”Open an agent and go to the Knowledge tab. The knowledge bases you add here apply in every chat, routine and team the agent joins, including routines, the Run agent workflow node and when another agent asks it for help.
- Each pinned knowledge base is a card with its sources, chunks, last update, privacy badge and a remove button. Click its name to open it.
- An empty knowledge base shows a warning (“Empty. Agent will find nothing here”) with Add sources.
- Add knowledge base opens the same picker as in chat; Create one with files creates a new one.
- When to look things up — Always, When relevant or When asked, as above.
- Show sources in answers — numbered citations you can open. When off, answers still use the passages, without numbers or source chips.
- Let agent search on its own — adds a knowledge search tool, so the agent can search again with a better question or in a specific knowledge base. The first search before answering still happens. It needs a model that supports tools; the switch is disabled otherwise. Device-only knowledge bases are skipped when the agent runs on a cloud model.
- Try it — type a question a user might ask and see the passages this agent would get, with their scores. Passages below the cut-off are marked below the cut-off, not used. Open full test search opens the knowledge base’s Test search.
Teams and projects
Section titled “Teams and projects”- In a team thread, each member uses its own knowledge bases plus the team’s and the project’s.
- In a project, chats use the knowledge bases linked to the project in addition to the project’s own files. Project files are indexed by the same engine, in a hidden knowledge base that belongs to the project.
Notch and canvas chat
Section titled “Notch and canvas chat”When you ask Aria from the Assistant Notch or the canvas chat, she uses the knowledge bases pinned on Aria (agent builder › Aria › Knowledge). There she searches before answering; the search tool isn’t available.