# Documents and extraction profiles

Document workflows can read selected files, extract text and fields, render page images, and produce records for review or downstream processing. Start by binding the correct document source, then define what should be extracted.

## Pick documents

1. Select the document-reading or extraction job.
2. Choose its source connection.
3. Open **Documents** and browse that source.
4. Select the files the job should process.
5. Confirm the selection and validate the workflow.

The picker is scoped to the step's own source and, for an app, its permitted document area. It is not a global picker that silently moves files between stores. If the picker asks for a source first, choose the connection before browsing.

A stored file key is an exact reference, not a filename search. Moving or replacing a file can change whether the reference remains valid. Check file access and source selection when a run reports a missing document.

## Configure extraction

A document profile describes the structure or extraction behavior expected for a document family. Review its purpose and fields rather than choosing a profile only because its name sounds similar. Use the profile editor to select or refine the profile supported by the job.

The profile editor offers three modes:

- **Automatic detection** chooses a suitable document profile and uses open discovery when none matches. Browsing mappings in this mode does not force that profile.
- **Selected document type** applies the profile you choose. Search by name, family, or purpose, then review its column mapping.
- **Free discovery** starts without a predefined document type and is useful for unfamiliar layouts.

Review the mapping's column keys, types, inclusion rules, and meanings. Core columns remain in the proposal even when their values are missing; recommended and optional columns need useful source content. Missing core values remain empty rather than preventing saving.

Choose **Customize this profile in this job** for a local override. Use **Advanced JSON and column definitions**, edit the override, then **Apply profile configuration**. Job overrides take precedence over the app profile library and bundled catalog without replacing shared defaults for other apps. **Table purpose** explains the intended result and can guide the proposal.

Document understanding can suggest a profile, but suggestions require review. Field names, types, repeated groups, and source evidence should match the intended downstream table or app. Some profile-building workflows require manual execution because they support an interactive decision.

## Verify evidence and output

Test with representative documents, including alternate layouts and missing fields. Inspect extracted values together with their source references or anchors when available. A valid structured response can still be semantically incorrect.

For review workflows, preserve the link between the source revision and the proposed records. Replacing a document should lead to a new analysis rather than silently treating old evidence as current.

Model and extraction settings can come from app parameters. When the job declares that inheritance, the app's selection governs the run. See [documents](/docs/documents/overview), [outputs](/docs/workflows/outputs), and [data sources](/docs/workflows/data-sources).