Document reading turns a PDF or an image of a document, scanned ones included, into usable text: the text of each page in Markdown (headings, lists, tables), the extracted tables and, on request, specific fields (invoice number, date, amountā¦) following a schema provided by the feature that calls it. The features that use it live in the modules or projects that need them; this page describes the capability and its settings.
Turn it on
In AI configuration, Capabilities tab, Document reading section, choose a Provider and click Save. While it is set to None (AI disabled), the features that read documents are hidden or fall back to manual entry.
| Provider | Files read | Default model |
|---|---|---|
| OpenAI | PDF, PNG, JPEG, WEBP, GIF | gpt-4o |
| Anthropic | PDF, PNG, JPEG, WEBP, GIF | default chat model |
| Mistral (Mistral OCR) | PDF, PNG, JPEG | mistral-ocr-latest |
| Azure OpenAI | PNG, JPEG (no PDF) | none: deployment name required |
| AI provided by the publisher | PDF and images | chosen by the publisher |
The API key is the one from the provider's tab. A file of another type is refused before anything is sent ("This file type is not read by the configured AI provider.").
Size and content
- Maximum file size (MB): from 1 to 20 MB, 10 by default. A larger file is refused before anything is sent ("The file exceeds the maximum size accepted by the AI.").
- With Mistral OCR, extracted fields use the service's document annotation feature; without a field schema, OCR takes no additional instruction.
What is sent to the provider
The file itself, a reading instruction and, where relevant, the field schema and the context given by the feature. Nothing is kept on our side: the file is read in memory for the duration of the call, then forgotten; it is stored neither in the database nor in the logs. The usage log only keeps the feature, the provider, the model, the tokens and the number of pages read. Retention by the provider depends on its contract.
Consumption and caps
Each reading counts its pages and tokens in AI usage (By capability breakdown). The Monthly cap (pages read) of the Usage screen and a feature's own cap (Features tab) block once reached. With the publisher's AI, reading is deducted from the token credit.
See also
- "Configure AI": capabilities, providers, choice per feature.
- "Monitor AI usage": consumption and caps.