infly-ai/inf-mllm/infinity-parser

Infinity Parser Skill

Submit one PDF or supported image per request to get Markdown and layout blocks. Download URLs or capture webpage screenshots before sending the resulting file.

Copy .env.example to .env, then set INF_API_URL to the API base URL and INF_API_KEY to your key. The Python command reads .env automatically without printing the key. Environment variables with the same names override the file and can also be used without one.

Choose a tier

Input and options

• Send multipart/form-data with required file and optional tier (nano, flash, or pro; default flash). The same fields apply to /v1/parse and /v1/parse/stream.
• PDFs and single-frame PNG, JPEG, WEBP, BMP, TIFF, and GIF are supported. The Gateway identifies content from bytes, so a wrong extension or MIME type does not prevent parsing.
• Multi-frame images, HTML, URL strings, Office files, SVG, and archives are unsupported. Download a web URL or capture a webpage screenshot in the platform, then send the resulting PDF or image.

python3 skills/infinity-parser/scripts/parse.py /path/to/report.pdf \
  --tier pro \
  --pages 1-3,5 \
  --parse-chart true \
  --keep-header-footer false \
  -o /path/to/output

The command saves report.json and report.md in the -o directory. Omit -o to save them next to the input. Omit --pages to parse all pages. --tier defaults to Flash. Use --keep-header-footer true to include header/footer text in Markdown or --parse-chart false to skip extra chart extraction. Partial page failures still save both files and return exit code 2.

For a direct API request, send tier:

set -a
source skills/infinity-parser/.env
set +a
curl "${INF_API_URL%/}/v1/parse" \
  -H "Authorization: Bearer ${INF_API_KEY}" \
  -F 'file=@/path/to/report.pdf' \
  -F 'tier=pro' \
  -F 'pages=1-3,5' \
  -F 'parse_chart=true' \
  -F 'keep_header_footer=false'

Read references/api-reference.md for response and streaming event details.

Use the multipart endpoint in preference to the legacy /v1/chat/completions JSON endpoint. The OpenAI-compatible JSON endpoint still uses model; it accepts PDF Base64 only and requires data:application/pdf;base64,... plus a .pdf filename.