> ## Documentation Index
> Fetch the complete documentation index at: https://www.activepieces.com/docs/llms.txt
> Use this file to discover all available pages before exploring further.

# Agent Knowledge

> Give an agent something to look things up in

Attach a document or a table and the agent looks the answer up instead of guessing at it.

## Two sources

<CardGroup cols={2}>
  <Card title="Files" icon="file" color="#8143E3">
    Upload a document. PDF, DOCX, TXT, or CSV.
  </Card>

  <Card title="Tables" icon="table" color="#0EA5E9" href="/docs/tables/with-agents">
    Point at a table already in your project.
  </Card>
</CardGroup>

Both live in the **Knowledge Base** section of the agent editor, below Agent Tools. Add one with **Add File Source** or **Add Table Source**.

<Note>
  Two things have to be in place or the section does not appear at all. Your Postgres needs the `pgvector` extension, which Activepieces installs on startup when the database offers it — Cloud and the standard self-hosted images are covered, but a managed Postgres without `pgvector` hides knowledge entirely. And the agent's model provider has to support embeddings: OpenAI, Google, Azure, and OpenRouter do.
</Note>

## Which one

|                | File                        | Table            |
| -------------- | --------------------------- | ---------------- |
| **Holds**      | Documents, policies, guides | Rows of data     |
| **Changes**    | Rarely                      | Constantly       |
| **Updated by** | Re-uploading                | Your flows, live |

<Tip>
  If a flow already writes the data, use a table. The agent then reads what's true right now, with nothing to re-upload.
</Tip>

## Why not just paste it in

Two reasons, and the first is size.

Everything in an agent's instructions has to fit the model's context window, and all of it is read on **every single run**. That caps you at a few pages, and makes every run slower and more expensive whether the content was relevant or not.

Knowledge works the other way round. A file is split into chunks and indexed, and the agent searches it only when it needs something, pulling back the passages that match. A three-hundred-page policy manual costs nothing until a question actually touches it.

The second reason is freshness. A product list pasted into instructions works until the list changes, and then you have two versions of the truth and one of them is wrong.

Instructions are for **how to behave**. Knowledge is for **what's true**.

## Say when to use it

Knowledge tells the agent what it can look up, not when to bother. Put that in the [instructions](/docs/agents/create):

> Before answering pricing questions, check the pricing table. Never quote a figure that isn't in it.