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

# Primary Sources

> DABAR reasons only over sources you define — documents, PDFs, audio, video, databases, APIs. No open-web crawling. No AI citing AI. Every claim traced to source and page.

DABAR is built on a simple principle: **an answer is only as trustworthy as its sources**. Every output is traced back to a source you, the organization, have explicitly approved.

## What counts as a primary source

DABAR accepts structured and unstructured inputs across a wide range of formats:

<CardGroup cols={2}>
  <Card title="Documents" icon="file">
    PDFs, Word documents, plain text, spreadsheets.
  </Card>

  <Card title="Media" icon="photo-film">
    Audio recordings, video, images, maps.
  </Card>

  <Card title="Live data" icon="database">
    SQL databases, internal APIs, approved external APIs.
  </Card>

  <Card title="Web" icon="globe">
    Specific URLs you whitelist — never unverified open-web content.
  </Card>
</CardGroup>

## The "AI citing AI" problem

Most AI research tools pull from the open internet. That means:

* They may cite pages that were themselves written by AI.
* They cannot distinguish authoritative from unauthoritative sources.
* Their outputs cannot be audited — the ground truth changes daily.

DABAR eliminates this by design. **Sources are an explicit allowlist.** If your organization hasn't approved it, DABAR won't use it.

## Where sources live

Sources in DABAR exist at two levels:

### Policy-level sources

Defined on a **Politic** via `knowledgeFiles`. These are the authoritative references the policy reasons **against** — regulations, playbooks, internal guidelines.

```json theme={null}
{
  "knowledgeFiles": [
    { "name": "risk-guidelines.pdf", "type": "pdf", "url": "https://..." }
  ]
}
```

### Project-level sources

Defined on a **Project** via `files`. These are the specific inputs DABAR is being asked to analyze — a loan application, a contract, a due diligence packet.

```json theme={null}
{
  "files": [
    { "name": "application-001.pdf", "type": "document", "url": "https://..." },
    { "name": "analyst-notes",       "type": "text",     "content": "..." }
  ]
}
```

Supported project file types: `document`, `url`, `photo`, `video`, `audio`, `text`.

## How DABAR uses them

When a project runs, DABAR:

1. **Ingests** every source, preserving structure (page numbers, timestamps, cell references).
2. **Grounds** its reasoning against the policy's `knowledgeFiles`.
3. **Cites** the exact source and location for every claim it makes.
4. **Flags** anything that cannot be grounded — never fabricates a citation.

## Source control = audit trail

Because every output is traced to a source you approved, DABAR produces a **complete audit trail** by default. That makes outputs:

* **Defensible** in legal and compliance contexts.
* **Reproducible** — re-run the same project and the citations still point where they pointed.
* **Portable** — the evidence travels with the answer, not hidden in a model's weights.

## Next

<CardGroup cols={2}>
  <Card title="Policy Engine" icon="scale-balanced" href="/concepts/policy-engine">
    How DABAR decides what a source means.
  </Card>

  <Card title="Autonomous Agents" icon="bolt" href="/concepts/agents">
    Turn a sourced knowledge base into an agent that takes action.
  </Card>
</CardGroup>