OCR

POST /v1/ocr extracts text, tables, and layout from a PDF, JPEG, or PNG in a single synchronous request. The response is the selected provider's native OCR output, returned unchanged. Use GET /v1/models to find the OCR models available to your project.

Send a document by URL

curl -X POST "https://api.llmapi.ai/v1/ocr" \
  -H "Authorization: Bearer $LLM_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "mistral-ocr",
    "document": {
      "type": "document_url",
      "document_url": "https://example.com/invoice.pdf"
    }
  }'

document takes one of three forms:

`type`	Field	Use for
`document_url`	`document_url`	A PDF, as an `http(s)` URL or a `data:application/pdf;base64,...` URL.
`image_url`	`image_url`	A JPEG or PNG, as an `http(s)` URL or a `data:image/...;base64,...` URL.
`file`	`file_id`	A file uploaded earlier with `POST /v1/files` and `purpose=ocr`.

Upload the file directly

Send multipart/form-data with a file part and a model field. Options can be sent as individual form fields or as one JSON object in a request field.

curl -X POST "https://api.llmapi.ai/v1/ocr" \
  -H "Authorization: Bearer $LLM_API_KEY" \
  -F "model=mistral-ocr" \
  -F "file=@invoice.pdf" \
  -F "pages=0,1,2"

Reuse an uploaded file

Upload once with purpose=ocr, then reference the returned id from any number of OCR requests in the same project:

curl -X POST "https://api.llmapi.ai/v1/files" \
  -H "Authorization: Bearer $LLM_API_KEY" \
  -F "purpose=ocr" \
  -F "file=@invoice.pdf"

curl -X POST "https://api.llmapi.ai/v1/ocr" \
  -H "Authorization: Bearer $LLM_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "mistral-ocr", "document": {"type": "file", "file_id": "file-abc123"}}'

Request fields

Field	Description
`model`	Required. OCR model ID from the catalog.
`document`	Required for JSON requests. The document source, as described above.
`pages`	Zero-based page indexes to process, for example `[0, 1, 2]`. Omit to process the whole document.
`include_image_base64`	Return extracted images (or cropped regions) as base64.
`image_limit`, `image_min_size`	Cap the number of extracted images and skip images below a minimum size.
`table_format`	`markdown` or `html`.
`extract_header`, `extract_footer`	Include page headers and footers in the output.
`include_blocks`	Return layout blocks alongside the text.
`confidence_scores`, `confidence_scores_granularity`	Return confidence scores, grouped by `word`, `block`, or `page`.
`document_annotation_format`, `bbox_annotation_format`	JSON schemas for structured extraction over the whole document or each bounding box.
`document_annotation_prompt`	Instructions for document annotation.
`need_layout_visualization`	Return layout visualization images, where supported.

Options are model-specific. Mistral OCR models accept the full set above. GLM-OCR models support pages (which must be a contiguous range), include_image_base64, and need_layout_visualization, and reject most other options with a 400 that names the parameter.

Response

The body is the provider's native OCR response. Mistral OCR models return per-page markdown:

{
  "model": "mistral-ocr-latest",
  "pages": [
    {
      "index": 0,
      "markdown": "# Invoice 1042\n\n| Item | Qty | Price |\n| --- | --- | --- |\n| ...",
      "images": [],
      "dimensions": { "dpi": 200, "height": 2200, "width": 1700 }
    }
  ],
  "usage_info": { "pages_processed": 1, "doc_size_bytes": 48213 }
}

GLM-OCR models return md_results, layout_details, and token usage instead. Check the model you requested and parse the matching shape.

Limits and billing

Supported formats are PDF, JPEG, and PNG; the type is detected from the file content.
Uploaded files: PDF up to 50 MB, images up to 10 MB. The maximum page count per document depends on the model's provider (100 pages for some); use pages to process large documents in parts.
Billing depends on the model's pricing mode:
- Page-priced models charge per processed page. Requests with document_annotation_format or bbox_annotation_format are charged at the annotated-page price.
- Token-priced models (such as GLM-OCR) charge for the prompt and completion tokens the provider reports, at the model's input and output token rates. The final charge does not depend on the page count.
In both modes, credits are reserved up front from the page price and the number of pages requested (a default ceiling when pages is omitted), then settled on actual usage. For token-priced models the reservation is only an estimate, so don't use it to predict the final charge.
Provider fallback and batch OCR are not supported; each request runs synchronously on the selected model.

Errors

Status	Meaning
`400`	Invalid request, unsupported format, or an option the model doesn't support.
`402`	Insufficient credits.
`404`	The referenced `file_id` doesn't exist in this project.
`413`	File exceeds the size limit.
`502`	The OCR provider failed or is unavailable.