v0.2.1

Fixed

• docai-ingest names each Markdown file <pdf name>-<hash>.md, where the hash comes from the full PDF path. Two PDFs with the same file name in different folders were written to the same <pdf name>.md, so the second one was skipped and never indexed. Markdown written by earlier versions is not reused, so the first run after upgrading OCRs every PDF again.
• docai-classify keeps .txt and .md files when neither python-magic nor the file command is available. The built-in signature sniff has no signature for plain text, so these files were quarantined as content type undetermined. A .txt or .md file with no NUL bytes in its first 8 KB is now treated as text/plain.

Upgrade notes

• The first run of docai-ingest after upgrading OCRs every PDF again, because Markdown files written by 0.2.0 under the old <pdf name>.md name are not reused.

Documentation

• The README and package metadata match the code. The project URL points to github.com/novatechflow/docai, the description names OpenAI-compatible and custom OCR endpoints, and the classifiers list Python 3.14, which CI tests and the Docker image uses.
• The README shows how to keep only the PDFs from the docai-classify --doc-list before passing it to docai-ingest.
• The README Docker example runs docai_toolkit.classify instead of pytest, which the image does not include. The test section lists the commands CI runs.
• The case study section that described DOCX loading and text cleaning was removed, because the code does neither.