smartpdfplumber 0.1.5

A wrapper around LangChain’s PDFPlumber integration with added support for image-aware data extraction.

Smart PDF Plumber

A lightweight wrapper over LangChain’s PDFPlumber integration that extends PDF parsing with image understanding—extracting not just text, but also contextual insights from embedded images.

It is designed for two common cases:

• Extract plain text from PDFs page by page.
• FEATURE: describe embedded images and include those descriptions in the page text.

Features

• Page-level PDF parsing with pdfplumber.
• Optional character deduplication for PDFs with repeated text.
• Optional image description support using either Google Gemini or Hugging Face vision-language models.
• LangChain-friendly output: a list of Document objects with metadata such as source path, page number, and total pages.

Installation

Install from PyPI:

pip install smartpdfplumber

The project currently targets Python 3.13 or newer.

Quick Start

from smartpdfplumber.loader import SmartPDFLoader

loader = SmartPDFLoader("path/to/file.pdf", describe_image=True, inference="groq_ai")
documents = loader.load()

for document in documents:
	print(document.metadata)
	print(document.page_content)

text_kwargs

Extra keyword arguments passed to pdfplumber.Page.extract_text().

dedupe

Set dedupe=True to call page.dedupe_chars() before extracting text. This can help with PDFs that repeat characters in the output.

describe_image

Set describe_image=True to include image descriptions inline in the page text.

When this is enabled, you must also provide inference.

inference

Supported values:

• gemini
• hf_transformers
• groq_ai

Image Descriptions

from smartpdfplumber.loader import SmartPDFLoader

loader = SmartPDFLoader(
	"path/to/file.pdf",
	dedupe=True,
	describe_image=True,
	inference="huggingface",
	model="Qwen/Qwen3.5-0.8B" # Optional: Default("Qwen/Qwen3.5-0.8B")
)
documents = loader.load()

Notes

• If you enable image descriptions without passing model, the parser raises a ValueError.
• If you use model="hf_transformers", the model is loaded lazily and cached for reuse.

License

MIT