PDF-to-LLM pipeline that actually reads the room

What it does

PyMuPDF4LLM is a convenience layer over the venerable PyMuPDF C engine. Feed it a PDF (or Office docs via the paid Pro add-on) and it emits Markdown, JSON, or plain text with reading-order reconstruction, table detection, and heading hierarchy intact. The target audience is obvious: RAG pipelines, vector stores, and anything else that needs LLM-ready text without calling a cloud API.

The interesting bit

The hybrid OCR strategy is the standout. Instead of blanket-OCRing every page or ignoring scanned regions entirely, it inspects each page first—checking for illegible characters, vector graphics masquerading as text, image-covered areas, and existing OCR layers—then applies OCR only where needed. The README claims this cuts OCR time by roughly 50% versus full-document approaches, and avoids degrading clean digital text with recognition errors. It auto-selects between Tesseract and rapidocr_onnxruntime at runtime, or accepts a custom OCR function.

Key highlights

• Three output formats from one import: to_markdown(), to_json(), to_text()
• Layout-aware extraction: multi-column reconstruction, table-to-Markdown, header/footer stripping, font-size-based heading detection
• Page chunking mode returns per-page dicts with metadata, bounding boxes, and TOC items—ready for vector stores
• Drop-in integrations for LlamaIndex (LlamaMarkdownReader) and LangChain (PyMuPDFLoader, MarkdownTextSplitter)
• Runs fully offline; no GPU, no tokens, no cloud bill

Caveats

• The “10–250× cheaper than vision-based LLM extraction” claim is in the README but unsourced—treat as marketing unless you verify against your own pipeline costs
• Office document support (Word, Excel, PowerPoint, HWP) requires the separate PyMuPDF Pro package, which is not open source
• Legacy layout mode exists (use_layout(False)) with different header detection behavior; the README doesn’t clarify when you’d still need it

Verdict

Worth a look if you’re building document ingestion for RAG and want to stay local. Skip it if you need deep PDF editing or already have a mature extraction stack you’re happy with.