PDF indexing fails on Windows: NamedTemporaryFile(delete=True) blocks PyMuPDFLoader from opening the temp file

[chirag127]

Problem

On Windows, PdfToEntries.extract_text cannot index any PDF because PyMuPDFLoader is passed the path of a still-open NamedTemporaryFile(delete=True), and Windows disallows a second open on that path.

Cite

src/khoj/processor/content/pdf/pdf_to_entries.py:100-105

with tempfile.NamedTemporaryFile(suffix=".pdf", delete=True) as tmpf:
    tmpf.write(pdf_file)
    tmpf.flush()
    loader = PyMuPDFLoader(tmpf.name)
    pdf_entries_per_file = loader.load()

Per Python docs: on Windows the name cannot be used to open the file a second time while the handle is still open.

Steps

1. Windows 11, pip install khoj (Python 3.10-3.12 per pyproject.toml).
2. Upload any PDF via /api/content?t=pdf.

Expected

PDF is indexed.

Actual

extract_text catches PermissionError: [WinError 32] The process cannot access the file because it is being used by another process and logs Unable to process file: <bytes>. This file will not be indexed. No entries created.

Environment

OS: Windows 11; Python 3.10-3.12; khoj 2.0.0-beta.28; pymupdf == 1.24.11, langchain-community == 0.3.31 (from pyproject.toml).

Suggested fix

Use delete=False + explicit os.unlink(tmpf.name) in a finally block, or (Python 3.12+) delete_on_close=False.

Thanks for maintaining khoj-ai/khoj!

[tianmind-studio]

I reproduced this on Windows 11 / Python 3.10: reopening the path from NamedTemporaryFile(delete=True) while its handle is live raises PermissionError. I'm working on the close-before-load + explicit-finally-cleanup fix, with a platform-independent regression test that asserts the temp handle is closed before the loader opens it and the path is removed afterward.