langchain-pdfmuse

LangChain document loader for pdfmuse — a deterministic PDF/DOCX parser for RAG. Same file in → same Documents out, with exact coordinates, tables and section structure.

pip install langchain-pdfmuse

Usage

from langchain_pdfmuse import PdfmuseLoader

# one Document per page (default)
docs = PdfmuseLoader("report.pdf").load()

# RAG-optimized: one Document per block, with section-aware metadata
elements = PdfmuseLoader("report.pdf", mode="elements").load()
for e in elements:
    print(e.metadata["category"], e.metadata["heading_path"], e.metadata["bbox"])

Modes

Metadata (elements mode)

Each Document.metadata carries:

• source, source_kind (Pdf / Docx)
• page — 0-based page index
• category — Title · NarrativeText · Table
• heading_path — the section breadcrumb, e.g. ["Experience", "Alibaba"]
• bbox — {x0, y0, x1, y1} on the page (top-left origin, points)

heading_path + bbox let your retriever return "which section, where on the page" — not just a blob of text.

Why deterministic matters for RAG

Non-deterministic extractors make your chunks (and therefore embeddings and eval results) drift between runs. pdfmuse has no probabilistic models in its core path, so your index is reproducible. OCR / layout inference are opt-in backends, kept out of the deterministic core — scanned pages surface a NeedsOcr warning instead of guessing.

MIT OR Apache-2.0 · part of the pdfmuse project.