# Knowledge base

> Giving Triagic your internal docs to read during investigations, by uploading files or connecting a docs integration.

Source: https://triagic.com/docs/admin/knowledge

**Knowledge** is where an organization gives Triagic the reference material a new
support engineer would get on day one: runbooks, product notes, escalation
rules, glossaries, exported spreadsheets. The agent searches it while
investigating a ticket and in the console, and names the document it used.

Until an organization has either uploaded a document or connected a docs
integration, admins see a reminder at the top of every portal page.

## Two ways to set it up [#two-ways-to-set-it-up]

| Route | What Triagic reads | Where |
| --- | --- | --- |
| **Upload documents** | The text of each file, extracted once at upload. | This page. |
| **Connect a docs integration** | Live pages through the integration's MCP server. Notion and Atlassian (Confluence) count. | [Integrations](/docs/admin/integrations). |

Both can be used together. A connected integration answers "what does our wiki say
right now"; uploads are for material that lives in files, or that you don't want the
agent to reach over an API.

## Uploading [#uploading]

Admins pick a file and, optionally, write a short description. The description
is shown to the agent alongside the text, so a line like *"Escalation matrix for
payments incidents, updated quarterly"* helps it choose the right document.

| Type | Extensions | Notes |
| --- | --- | --- |
| Text | .md, .txt, .csv, .json | Read as-is. |
| Word | .docx | Text only; images inside the document are dropped. |
| PDF | .pdf | Text layer only. A scanned PDF with no text layer yields nothing. |
| Spreadsheet | .xlsx | One block per sheet, rows as comma-separated values. |
| Image | .png, .jpg, .gif, .webp | Triagic can't read pictures yet, so the description is required and is all the agent sees. |

Limits: 10 MB per file, 50 documents per organization, and 200,000 characters of
text per document (the rest is dropped). Each row on the page shows how much text
was kept, so an empty count on a PDF is the sign of a scan.

## Generate a doc from a prompt [#generate-a-doc-from-a-prompt]

Admins can also have Triagic write a doc. On a cloud-managed install, open the
desktop app's **Knowledge** page, press **Generate from a prompt**, and describe what
the doc should cover, for example *"Late delivery complaints"*. Under **Make**, pick
**Knowledge doc**, **Playbook**, or **Both, linked** (see
[Managing playbooks](/docs/admin/playbooks#generate-a-playbook-from-a-prompt)).

Under **It may read**, choose what the AI may look at:

- **Data sources** you pick: for a database, its schema and a small sample of rows.
- **Triagic tickets**: titles, resolutions and agent notes.
- **Existing knowledge docs**, so it adds to what is already written instead of
  repeating it.

The result is a draft. A draft lives on your desktop, is private to you, and never
syncs until you publish it. Each section names what it was written from, such as
*41 tickets* or *Postgres schema*, and the side panel lists everything the run read
and what it cost.

**Revising.** Type into **Ask for changes** (for example *"Shorter"* or *"Add an
example reply"*) and press **Revise**. The changes come back as a redline: each
changed, added or removed section is shown on its own, and you press **Keep change** or
**Discard** on each one. Every change needs a decision before you can publish. **Edit
by hand** saves your own text as a new version at no cost.

**Publishing** is free. **Publish to org** writes the doc to the knowledge base (or
replaces the text of the doc it came from), records it in [Audit](/docs/admin/audit),
and reaches the organization's desktop apps on their next config sync, within 15
minutes. The published doc stays as it is while you keep revising the draft, until
you publish again.

To revise a doc that already exists, an admin clicks it in the desktop **Knowledge**
list. That opens a draft holding its current text, ready for **Ask for changes**.

Generating and revising spend AI through the provider your desktop uses; the buttons show
the price first. See [Spending and budget](/docs/admin/spending#per-run-cap-for-generation).
Installs that aren't centrally managed have no knowledge base, so there the generator
makes playbooks only.

## Enabling, editing, deleting [#enabling-editing-deleting]

Every document has an **Enabled** switch. A disabled document stays listed but is
not sent to the desktop apps and is not searched. Delete removes it for good;
there is no undo.

Every upload, edit and delete is recorded in [Audit](/docs/admin/audit) and
pushed to the organization's desktop apps on their next config sync, within
15 minutes.

## What the agent does with it [#what-the-agent-does-with-it]

The desktop app keeps a copy of the enabled documents and exposes two tools to the
agent: one that lists them and one that searches them by keyword. The
investigator's instructions tell it to search the knowledge base before answering
a product or process question. Document text is treated as untrusted content, the
same as ticket text: instructions written inside a document can't change how the
agent behaves.

Members can see the list and the descriptions. Only admins can upload, edit,
delete, generate or revise.
