langchain-opendataloader-pdf

LangChain document loader for OpenDataLoader PDF — parse PDFs into structured Document objects for RAG pipelines.

For the full feature set of the core engine (hybrid AI mode, OCR, formula extraction, benchmarks, accessibility), see the OpenDataLoader PDF documentation.

Features

• Accurate reading order — XY-Cut++ algorithm handles multi-column layouts correctly
• Table extraction — Preserves table structure in output
• Multiple formats — Text, Markdown, JSON (with bounding boxes), HTML
• Per-page splitting — Each page becomes a separate Document with page number metadata
• AI safety — Built-in prompt injection filtering (hidden text, off-page content, invisible layers)
• 100% local — No cloud APIs, your documents never leave your machine
• Fast — Rule-based extraction, no GPU required

Requirements

• Python >= 3.10
• Java 11+ available on system PATH

Installation

pip install -U langchain-opendataloader-pdf

Quick Start

from langchain_opendataloader_pdf import OpenDataLoaderPDFLoader

loader = OpenDataLoaderPDFLoader(
    file_path="document.pdf",
    format="text"
)
documents = loader.load()

print(documents[0].page_content)
print(documents[0].metadata)
# {'source': 'document.pdf', 'format': 'text', 'page': 1}

Usage Examples

Batch Processing

from langchain_opendataloader_pdf import OpenDataLoaderPDFLoader

# Single file, multiple files, or directories — all in one call
loader = OpenDataLoaderPDFLoader(
    file_path=["report1.pdf", "report2.pdf", "documents/"]
)
docs = loader.load()

Output Formats

# Plain text (default) — best for simple RAG
loader = OpenDataLoaderPDFLoader(file_path="doc.pdf", format="text")

# Markdown — preserves headings, lists, tables
loader = OpenDataLoaderPDFLoader(file_path="doc.pdf", format="markdown")

# JSON — structured data with bounding boxes for source citations
loader = OpenDataLoaderPDFLoader(file_path="doc.pdf", format="json")

# HTML — styled output
loader = OpenDataLoaderPDFLoader(file_path="doc.pdf", format="html")

Tagged PDF Support

For accessible PDFs with structure tags (common in government/legal documents):

loader = OpenDataLoaderPDFLoader(
    file_path="accessible_document.pdf",
    use_struct_tree=True  # Use native PDF structure
)

Table Detection

loader = OpenDataLoaderPDFLoader(
    file_path="financial_report.pdf",
    format="markdown",
    table_method="cluster"  # Better for borderless tables
)

Sensitive Data Sanitization

# Replace emails, phone numbers, IPs, credit cards, URLs with placeholders
loader = OpenDataLoaderPDFLoader(
    file_path="document.pdf",
    sanitize=True
)

Extract Specific Pages

loader = OpenDataLoaderPDFLoader(
    file_path="document.pdf",
    pages="1,3,5-10"
)

Include Headers and Footers

# By default, headers and footers are excluded for cleaner RAG output
loader = OpenDataLoaderPDFLoader(
    file_path="document.pdf",
    include_header_footer=True
)

Password-Protected PDFs

loader = OpenDataLoaderPDFLoader(
    file_path="encrypted.pdf",
    password="secret123"
)

Image Handling

# Images are excluded by default (image_output="off")
# This is optimal for text-based RAG pipelines

# Embed images as Base64 (for multimodal RAG)
loader = OpenDataLoaderPDFLoader(
    file_path="doc.pdf",
    format="markdown",
    image_output="embedded",
    image_format="jpeg"  # or "png"
)

# Save images as files to a local directory
loader = OpenDataLoaderPDFLoader(
    file_path="doc.pdf",
    format="markdown",
    image_output="external",
    image_dir="./images",   # images saved here; defaults to temp dir if not set
    image_format="png"
)

Hybrid AI Mode

For complex documents (tables, charts, scanned content), hybrid mode routes pages to an AI backend for better accuracy while keeping simple pages on the fast local engine:

# Requires a running docling-fast server (default: localhost:5002)
loader = OpenDataLoaderPDFLoader(
    file_path="complex_report.pdf",
    format="markdown",
    hybrid="docling-fast",          # Enable hybrid extraction
    hybrid_mode="auto",             # Auto-triage: only complex pages go to backend
    hybrid_url="http://localhost:5002",
)
documents = loader.load()

# Document metadata shows which backend was used
print(documents[0].metadata)
# {'source': 'complex_report.pdf', 'format': 'markdown', 'page': 1, 'hybrid': 'docling-fast'}

Suppress Logging

loader = OpenDataLoaderPDFLoader(
    file_path="doc.pdf",
    quiet=True
)

RAG Pipeline Example

from langchain_opendataloader_pdf import OpenDataLoaderPDFLoader
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain_openai import OpenAIEmbeddings
from langchain_community.vectorstores import FAISS

# Load PDF
loader = OpenDataLoaderPDFLoader(
    file_path="knowledge_base.pdf",
    format="markdown",
    quiet=True
)
documents = loader.load()

# Split into chunks
splitter = RecursiveCharacterTextSplitter(
    chunk_size=1000,
    chunk_overlap=200
)
chunks = splitter.split_documents(documents)

# Create vector store
embeddings = OpenAIEmbeddings()
vectorstore = FAISS.from_documents(chunks, embeddings)

# Query
results = vectorstore.similarity_search("What is the main topic?")

Parameters Reference

Document Metadata

Each returned Document includes metadata:

doc.metadata
# {'source': 'document.pdf', 'format': 'text', 'page': 1}

# When hybrid mode is active:
# {'source': 'document.pdf', 'format': 'text', 'page': 1, 'hybrid': 'docling-fast'}

When split_pages=False, the page key is omitted.

License

Apache License 2.0. See LICENSE for details.

Links

• Documentation — Full documentation (hybrid mode, benchmarks, accessibility)
• GitHub — Core engine source code
• LangChain Docs — LangChain integration reference
• PyPI Package