PDF and document input

The gateway extracts the text request-locally before the model sees it and keeps neither the file nor the text. Accepted: PDF, DOCX, PPTX, XLSX, XLS, plain text. Images go as an image part.

The `file` Part

{
  "role": "user",
  "content": [
    { "type": "text", "text": "Summarize this document." },
    {
      "type": "file",
      "file": {
        "file_url": "https://example.org/report.pdf"
      }
    }
  ]
}

Field	Type	Notes
`type`	string	`"file"` or `"input_file"`.
`file.file_url`	string	A URL to fetch the document from. With `input_file`, the fields sit at the top level of the part.
`file.file_data`	string	Inline document bytes as a data URL, as an alternative to `file_url`.
`file.filename`	string	Used with `file_data` to hint the format.

file_id is rejected; there is no Files API. Scanned or image-only PDFs have no text to extract; send page images to a vision-capable model. Extraction runs in a bounded process, so very large documents can fail rather than return shortened text; split them first. The Responses facade accepts input_file parts; the Messages facade accepts document blocks. Audio and video parts are rejected.

Extract text without inference

Upload a multipart file to POST /api/v1/attachments/parse. The API returns extracted text, a model-independent token estimate, and the document type. No model ID is required; document parsing is billed separately.

Basics

Features

Server-Side Tools

Multimodal

Best Practices

SDKs

API Reference

PDF and document input

The `file` Part

Extract text without inference

​The file Part

​Related

​Extract text without inference

The `file` Part

Related

Extract text without inference