Supported formats
Entries
Auto detection
Withinput_format=auto (default) the order is:
1
Extension
The extension of
filename is looked at.2
Content-Type
If the extension is not enough, the header
content-type is used.3
Error 415
If nothing matches, it responds
415 asking you to set input_format.ZIP
When you send a ZIP, the first supported document it finds is extracted.- Hidden folders and files (those starting with
.) are ignored. - If ZIP has more entries than allowed,
413responds. - If no member is a supported format, reply
415. - If the internal document exceeds the size limit, respond
413.
The ZIP is for packaging one document, not batch processing. For multiple documents, make a request for each one.
CSV to Markdown
The CSV is converted to a Markdown table. Cells are escaped (| and line breaks) and trimmed if they are very long, to keep the table readable.
What is preserved and what is lost due to formatting
Converting to Markdown always means losing some visual fidelity — Markdown It has no concept of a page, exact source or absolute position. This table summarizes what survives reasonably well and what is flattened or discarded:Real use cases by format
- PDF +
fast: bulk ingestion for a RAG search index where Speed matters more than preserving every nuance of layout. - DOCX for legal contracts:
clean=true(default) is especially useful here because contracts usually have repeated headers/footers in each page that dirty the Markdown if they are not normalized. - XLSX of financial reports: each sheet becomes a Markdown table independent — useful for a LLM to reason about a sheet of “Q3 Revenue” without having to parse a binary file.
- PPTX from sales decks: extracting only the text from the svalides is enough for an agent to summarize the content of a presentation without need to process images.
