Knowledge library: answer from your own material

Upload documents, write Q&A that must always get the same answer, choose which teams can retrieve each one, and pick the assistants that use them.

What the knowledge library holds#

Open Knowledge in the console. The page holds two kinds of material:

• Documents: Markdown, plain text or PDF that you upload. When answering, an assistant finds the sections relevant to the question and quotes from them.
• Q&A: questions and answers you write yourself. An assistant that uses them sees them on every reply, so use them for questions that must always get the same answer.

The library belongs to the whole workspace. Nothing in it is used automatically: an assistant only draws on a document or Q&A once you pick it in that assistant's settings (see Use knowledge in an assistant).

Who can do what#

Members who open Knowledge see "Changing this is available to workspace owners and admins. You can read what is here; ask one of them to make a change." at the top, and the forms for adding documents and Q&A don't appear.

Add a document#

1. Open Knowledge and scroll to Add a document.
2. Enter a Document title, up to 200 characters.
3. Paste text into Content, or select Choose file and pick a .md, .markdown, .txt or .pdf file. If the title is still empty, the file name fills it.
4. Under Who it is for, choose Everyone in this workspace or Only the teams I name (see Audience).
5. Select Add to the knowledge base.

Pasted text that contains a Markdown heading (a line starting with #) is treated as Markdown and split into sections at its headings. Text without one is treated as plain text.

The document then appears in the Documents table. Its Status moves from Queued to Indexing, and ends at Ready or Failed. Assistants only use documents that are Ready.

When indexing fails, the reason appears under the title, with a Retry button beside it. A common cause is a PDF with no extractable text, such as a scan or an image-only PDF. Export it as a PDF with selectable text, or paste the text instead.

Each row shows:

• Title, with who uploaded it ("Uploaded by …")
• Type: Markdown, Plain text or PDF
• Status
• Who can retrieve it: Everyone, Restricted or Only you
• Size
• Used by: how many assistants use it. Select it to list those assistants and who attached the document to each.

Formats and limits#

At a document limit you'll see "This is full: an assistant holds up to 50 documents, a workspace 500, and your own files 50. Remove one first." "Your own files" are the files you add to your own private assistants; see Files only you can see.

Q&A#

Q&A is for questions that must always get the same answer, such as a return window, opening hours or a pricing rule.

1. On the Knowledge page, find Write a Q&A in the Q&A section.
2. Fill in Question, phrased the way someone would actually ask it.
3. Fill in Answer, written the way the assistant should say it.
4. Choose Who it is for and select Add Q&A.

Each entry shows who wrote it, who last edited it, and Used by N assistants. Select that chip to see which assistants. Use Search questions and answers to find an entry.

Q&A and documents are used differently. A document goes through retrieval, and only the sections relevant to the question are used. A Q&A entry an assistant uses goes in front of the model on every reply, with no relevance contest. If one assistant uses more than 50 entries, each reply carries the 50 most relevant to the question.

Audience: who can retrieve it#

Every document and Q&A entry has an audience, which decides whose questions the assistant can answer from it:

• Everyone in this workspace: the default. Any member's question can draw on it.
• Only the teams I name: only questions from members of the named teams can. When anyone else asks the same assistant, it answers as though the material isn't there.

Workspace owners and admins can always retrieve everything, so they can audit what assistants read. If you choose Only the teams I name and name no team, only owners and admins can retrieve it, and the form warns you before you save.

Create teams first under Team & access. If the workspace has no teams yet, the form says so.

To change the audience of something already in the library:

1. Select Everyone or Restricted in the Who can retrieve it column of the documents table, or under the Q&A entry.
2. In the Who can retrieve "…"? dialog, change Who it is for and pick teams if needed.
3. Select Save.

Only owners and admins can open this dialog. Other members just see the label.

Use knowledge in an assistant#

Material in the library takes effect once you pick it in an assistant's settings. The Data sources section of an assistant's page has a Knowledge base part and a Q&A part:

• Documents: pick one under Use a document from the knowledge base and select Use it. To stop, select Stop using on its row. You can also add a document right there with Upload: it goes into the library and this assistant uses it.
• Q&A: pick one under Pick a Q&A from the library and select Use it. To stop, select Stop using.

For the rest of an assistant's settings, see Assistants.

Retrieval test

Below the assistant's Knowledge base part is Retrieval test. Enter what a user would ask in A question to test and select Test. It lists the sections and Q&A the assistant actually retrieves, in the order it reads them. This is the same retrieval a real reply runs.

Whose view has two options:

• As me: retrieves as you. Owners and admins can retrieve every restricted item, so you may see more than a colleague would.
• As someone Agent Lab cannot identify: as a person who can't be matched to a member of the workspace.

How assistants use knowledge#

On every reply, an assistant:

1. Retrieves the sections most relevant to the question from the documents it uses.
2. Adds the Q&A it uses (the 50 most relevant, if it uses more than 50).
3. Filters both by audience, keeping only what the person asking can retrieve.

Things to know:

• Only documents with status Ready are retrieved. Queued and Indexing documents are not.
• When the person asking is someone Agent Lab cannot identify, meaning they can't be matched to a member of this workspace, the assistant uses no documents and no Q&A. It answers from its instructions and skills alone. The retrieval test's As someone Agent Lab cannot identify option shows this.
• If reading the knowledge fails, the reply still goes out, without the knowledge.

Files only you can see#

Files you add to your own private assistant are visible only to you. They appear in the Knowledge documents table marked Only you. They don't count toward the workspace's 500 documents; they count toward your own 50. You can delete or retry them yourself without admin rights.

Sharing that private assistant shares its files with whoever you share it with.

Update and remove#

• Update a document: a document's content can't be edited after upload. To replace it, upload the new version, switch each assistant that uses the old one to the new one, then delete the old one. The Used by column shows which assistants to switch.
• Edit a Q&A entry: select Edit under it, change it, and select Save. Every assistant that uses it gets the new answer right away.
• Delete a document: select Delete on its row. The dialog lists the assistants that cite it; select Delete document to confirm. Those assistants can no longer cite it, and their existing conversations are unaffected. This can't be undone.
• Delete a Q&A entry: select Delete. The dialog says how many assistants use it; select Delete to confirm.

Uploads, attaching and detaching documents, and audience changes are recorded in the audit log under Team & access.

Next#