File Type Support#

MinerU accepts a fixed set of file types defined as whitelisted suffix lists in mineru/cli/common.py:

These three lists are combined into the constant SUPPORTED_UPLOAD_SUFFIXES used by the FastAPI layer for upload validation.

File Type Detection#

Detection uses a three-stage pipeline implemented in mineru/utils/guess_suffix_or_lang.py:

1. OOXML structure inspection — For Office documents, MinerU parses the ZIP package structure (reading _rels/.rels and [Content_Types].xml) to identify docx/pptx/xlsx by their main content-type relationships. This runs before Magika to avoid false positives from embedded objects 1.

2. Magika detection — If OOXML inspection returns nothing, Magika is invoked. A global Magika() instance is initialized at module load time 2. Two entry points exist:

   • guess_suffix_by_bytes(file_bytes, file_path=None) — used when raw bytes are available (e.g. in-memory processing)
   • guess_suffix_by_path(file_path) — used for on-disk files (e.g. during upload validation)

3. PDF signature fallback — When Magika returns "ai" or "html" for a .pdf-extension file, the code reads the first 4 bytes and corrects the label to "pdf" if the %PDF magic bytes are present 3 4. This handles edge cases where Magika misclassifies valid PDFs.

Filtering Behavior by Context#

CLI / collect_input_documents: When processing a directory, the CLI silently skips any file whose detected suffix is not in pdf_suffixes + image_suffixes + office_suffixes 5. No error or warning is emitted for skipped files; they are simply excluded from the collected document list.

FastAPI upload (save_upload_files): Unsupported file types are explicitly rejected. After writing the upload to disk, guess_suffix_by_path() is called; if the result is not in SUPPORTED_UPLOAD_SUFFIXES, the file is deleted and an HTTP 400 is raised with "Unsupported file type: {suffix}" 6.

read_fn (used in direct API calls): Raises a plain Exception("Unknown file suffix: {suffix}") for any suffix not in the whitelist 7.

Office vs. PDF/Image Pipeline Split#

Office files (docx, pptx, xlsx) are routed to a separate _process_office_doc path inside do_parse(), which dispatches to format-specific analyzers (office_docx_analyze, office_pptx_analyze, office_xlsx_analyze). After office files are processed, they are removed from the remaining list; only then does the PDF/image pipeline run 8. Image files are converted to PDF bytes before PDF processing begins via images_bytes_to_pdf_bytes 9.