Files & Documents

Attach images and documents to a request to analyze, compare or reference them.

Attach up to five supported images or documents to a request. Aivana can read them, answer questions about them, and compare several side by side. The API, the SDKs' attachment support and the Playground all follow the limits below.

• At most 5 files per request, in any mix, and 40 MiB in all.
• Word and CSV text together can't exceed 400,000 characters per request.
• A file that can't be read (an encrypted PDF or Word document, an old .doc file, a CSV that isn't text) is refused with an error that says why, and which file.

• A PDF is read as the original file, its pages as well as its text, so charts and scanned pages are read too, and answers cite pages, like (p. 4).
• A Word document is read as its text, with its headings, lists, tables and footnotes. Pictures inside it aren't read.
• A CSV file is read with totals, averages and ranges computed exactly over every row.
• Attach two or more documents to compare them: answers say which document each point comes from, by its title or its number in the order you attached them.

Send each file base64-encoded in the attachments field, with its mime_type. The SDKs take the same two fields (Node.js spells the second mimeType), so you don't build the request body yourself.

curl https://api.aivana.ai/v1/generate \
  -H "X-API-Key: $AIVANA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "prompt": "Summarize the key risks in this contract.",
    "attachments": [
      { "mime_type": "application/pdf", "data": "<base64 of the file>" }
    ]
  }'

Files are billed as input tokens, and are included in usage.input_tokens. Each image counts as its pixels divided by 750 after it is scaled to fit, at most 1,534 tokens. A PDF counts as 2,300 tokens a page. A Word or CSV file counts as one token for every 4 characters of its text. A request that doesn't finish isn't charged for its files.