langchain-docuweave

LangChain integration for DocuWeave — a PDF loader that reconstructs the document's heading hierarchy before chunking, so each chunk knows which section it came from.

Installation

pip install -U langchain-docuweave

Usage

from langchain_docuweave import DocuWeaveLoader

loader = DocuWeaveLoader("paper.pdf", max_tokens=512)
docs = loader.load()

print(docs[0].metadata["section_path"])
# "3 Methods > 3.2 Experimental Setup"

# or stream one chunk at a time
for doc in loader.lazy_load():
    print(doc.page_content[:80])
    print(doc.metadata["section_path"])

Metadata

Each returned Document includes:

Why not PyMuPDFLoader?

PyMuPDFLoader extracts text per page. DocuWeaveLoader builds a section tree from heading signals across the full document and cuts chunks at section boundaries. In a benchmark of 390 PDFs / 3,927 QA pairs, this improved Recall@1 by +23.4% over recursive character splitting (p<0.01, bge-base-en-v1.5).