Ollama-OCR: Extract Text from Images and PDFs with Vision Models

Overview

Ollama-OCR is a Python OCR tool that sends images and PDFs to vision-language models available through Ollama. It turns visual documents into text or structured output, with options for formats such as Markdown, JSON, tables, and key-value pairs.

It is intended for developers and users who want to run document extraction through an Ollama model rather than use a conventional OCR service. Choose it when you need a Python workflow or a local web interface, and can provide and run a compatible vision model.

Key Features

• Processes images and PDFs using vision models supported through Ollama.
• Provides a Python API for processing one file or batches of files.
• Offers output formats including Markdown, plain text, JSON, structured data, key-value pairs, and tables.
• Supports custom extraction prompts and a requested text language.
• Includes batch options for parallel processing, progress tracking, recursive file discovery, and image preprocessing.
• Provides a Streamlit web application with uploads, previews, batch processing, and downloadable results.

Use Cases

• Developers can add document text extraction to Python applications that already use Ollama.
• Teams can extract dates, names, or other specified fields from scans and images using custom prompts.
• Users processing folders of images can run batch extraction and review success and failure statistics.
• Analysts can turn tables, charts, and other visual document content into structured outputs for further use.
• Non-developers can try document extraction through the Streamlit interface instead of integrating the package directly.

Getting Started

Install the package:

pip install ollama-ocr

Install Ollama and pull a compatible vision model before processing files. See the README for model setup, usage examples, batch processing, and Streamlit instructions.

Alternatives

• Marker: Marker converts PDFs and images into structured formats with its own document-processing pipeline, rather than relying on Ollama-served vision-language models.
• PaddleOCR: PaddleOCR provides a production-oriented OCR and document AI engine with broad language support, rather than a workflow centered on Ollama VLMs.

Considerations

• OCR depends on an Ollama installation and a compatible vision-language model, which must be available separately.
• Output quality can vary by model. The README specifically cautions that LLaVA can sometimes produce incorrect output.
• Processing speed and resource needs depend on the chosen model and runtime; the repository does not provide benchmark figures.
• The repository reports Jupyter Notebook as its language, even though the documented package and web app are used from Python.