Replacing a langchain-community loader

LangChain sunset langchain-community on 22 May 2026 and archived its repository on 19 June (the announcement). Its loaders now come as standalone packages, each kept by the people closest to the library it wraps. They are still LangChain loaders: they build on langchain-core, return LangChain Document objects and fit the same splitters and vector stores. What changes is the import line, and sometimes what comes out.

PyPDFLoader and PDFPlumberLoader still install from the frozen 0.4.2 release, with a deprecation warning and no further updates. Before switching, compare the old loader and its replacement on your own documents.

Where each loader went

Library	Standalone package	Loader
PyMuPDF	langchain-pymupdf4llm	PyMuPDF4LLMLoader
Docling	langchain-docling[local]	DoclingLoader
Unstructured	langchain-unstructured	UnstructuredLoader
OpenDataLoader PDF	langchain-opendataloader-pdf	OpenDataLoaderPDFLoader
pypdf, pdfplumber	none yet	PyPDFLoader, PDFPlumberLoader in langchain-community 0.4.2
Azure Document Intelligence	none; langchain-azure-ai has no loader for it	complydoc’s preset calls Azure’s SDK

PyMuPDF, and so langchain-pymupdf4llm, is licensed AGPL. complydoc does not depend on it; check the licence before shipping it in your own product.

Compare the old loader and the new one

compare_langchain_loaders.py

"""Compare a retired langchain-community loader with its standalone replacement."""

# requires: langchain_community, langchain_pymupdf4llm

from langchain_community.document_loaders import PyPDFLoader  # archived June 2026
from langchain_pymupdf4llm import PyMuPDF4LLMLoader

import complydoc as cd

path = "src/complydoc/sample/vendor-assessment.pdf"
report = cd.compare_loaders(
    {"pypdf": PyPDFLoader(path), "pymupdf4llm": PyMuPDF4LLMLoader(path)},
    facts=["Two administrator accounts have no multi-factor authentication"],
)

comparison = report.loader_comparison
for row in comparison.loaders:
    print(
        f"{row.name:<12} {row.characters:>6,} characters  "
        f"{row.identifiers_in_text:>2} identifiers in text  "
        f"{len(row.metadata_keys):>2} metadata keys  "
        f"network attempts: {len(row.network_attempts)}"
    )

for document in report.documents:
    for reading in document.extractions[1:]:
        print(
            f"\n{document.relative_path}: {reading.extractor} matches the baseline "
            f"{reading.similarity:.0%}"
        )

for difference in comparison.identifier_differences:
    where = "text" if difference.location == "text" else f"metadata ({', '.join(difference.keys)})"
    print(
        f"{difference.label} in {where}: found by {', '.join(difference.found_by)}, "
        f"missed by {', '.join(difference.missed_by)}"
    )

for check in comparison.facts:
    print(check.fact, check.found)

# `complydoc ui` opens what is in .complydoc, where the command line writes too.
cd.write_json(report, ".complydoc/loaders.json")

The first loader is the baseline, so put the one you have today first. Then look at what your pipeline depends on:

• Text: how much of each page the two readings share, with the words that differ marked in the report.
• Facts: pass the sentences your retrieval has to find. Each is reported as kept whole, split or missing, per loader.
• Identifiers: any found in one loader’s output and not the other’s.
• Metadata keys: a filter or citation that reads a key the new loader does not return will stop matching.
• Time and network: load time per loader, and any connection one attempted.

From the command line, with the same comparison in a file:

loaders.yaml

loaders:
  pypdf: langchain_community.document_loaders:PyPDFLoader
  pymupdf4llm: langchain_pymupdf4llm:PyMuPDF4LLMLoader
paths: ./contracts
facts:
  - Any change to the scope must be agreed in writing

complydoc compare-loaders loaders.yaml
complydoc ui

What changes, on a two-column contract

Measured on master-services-agreement.pdf in the repository’s sample documents, eight pages set in two columns:

Loader	Words shared with pypdf	Section 7 on page 4
PyPDFLoader	baseline	Read column by column, heading on its own line
PDFPlumberLoader	30%	Read across both columns: the heading lands mid-sentence in section 6
PyMuPDF4LLMLoader	46%	Read column by column, as Markdown: # **7. Confidentiality**

PyMuPDF4LLM shares under half its words with pypdf because it rejoins each paragraph and writes Markdown, not because it lost text: its character count is within 2% of pypdf’s, and a chunker that splits on headings now sees them. Its metadata drops page_label and adds file_path, format, creationDate and modDate. Both number pages from 0 in page.

The presets

cd.parsers.docling() goes through langchain-docling, and needs its local extra to convert on your machine. cd.parsers.azure_document_intelligence() used langchain-community’s loader and now calls Azure’s SDK, azure-ai-documentintelligence, with the same request. In mode="page" it reports Azure’s page numbers, which start at 1, as page_number.