> For the complete documentation index, see [llms.txt](https://docs.artific.nl/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.artific.nl/en/for-administrators/assistants/knowledge-and-sources.md).

# Giving an assistant knowledge

How an assistant gets access to your organisation's documents.

An assistant with no knowledge can only rely on the general ability of its model and whatever you wrote in its instructions. Attaching your own documents is what turns it from a generic chatbot into something that knows your organisation.

## How it works

Documents live in the **Knowledge base**, organised into collections. An assistant does not read a collection directly. It uses a **search tool** built over one or more collections. That indirection is what lets you decide precisely what a given assistant can and cannot find.

```
Documents  →  Collection  →  Search tool  →  Assistant
```

## Setting it up

1. **Put the documents somewhere.** Create a collection and upload the files. See [Collections](/en/for-administrators/knowledge-base/collections.md).
2. **Make them searchable.** Create a search tool over that collection. See [Search tools](/en/for-administrators/knowledge-base/search-tools.md).
3. **Attach it to the assistant.** Either from the last step of the search tool wizard, or later from the assistant's **Tools** tab, where the search tool appears as a card you switch on.
4. **Tell the assistant to use it.** In **Instructions for AI**, be explicit: *"Answer only from the documents available to you. If the answer is not there, say so."*
5. **Turn on Show references** so users can see which document an answer came from.

## Getting good answers

The quality of an assistant's answers depends more on the documents than on any setting.

* **Curate, do not dump.** A focused collection of current documents beats a large collection containing three versions of the same policy. The assistant cannot tell which is the current one.
* **Remove superseded documents.** They are the most common cause of confidently wrong answers.
* **Scans are read for you, but text is still better.** Scanned PDFs are processed so their text can be searched. Recognition is not perfect, so where you have a choice, add the original text version rather than a scan of it.
* **Give documents meaningful names.** File names are shown to users in the references, so `Leave policy 2026.pdf` is far more useful than `scan_0043.pdf`.
* **Split very large documents** where it is natural to do so. A well-structured document with clear headings retrieves better than one long unbroken wall of text.

## Uploads inside a conversation

Separately from the knowledge base, users can attach a file to a single message when **Allow file upload in chat messages** is on. That file is used for that conversation only. It is not added to your knowledge base and other users never see it.

Use this for one-off work ("summarise this report"). Use the knowledge base for anything the assistant should know permanently.

## If your organisation uses Sources

Some organisations are on an earlier version of the knowledge feature, which appears in the menu as **Sources** rather than **Knowledge base**. It does the same job with a different structure: you link assistants directly on the source itself, and there are no collections or search tools.

See [Sources](/en/for-administrators/knowledge-base/sources.md). Your account manager can tell you which version your organisation uses and what moving to the newer one involves.