docin 0.1.2

A Python package for performing OCR and document indexing on legacy documents using the Mistral Ocr API.

README for docin OCR Tool

Overview

docin is a lightweight document processing toolkit that combines OCR (Optical Character Recognition) and intelligent document analysis. It features two main components:

1. OCR Engine: Powered by the Mistral API, it extracts text and images from PDF and image files, converting them into clean, structured Markdown format.

2. LangQuery: An intelligent document analysis tool that uses LangChain and spaCy to extract structured information through natural language queries and visualize named entities.

---

Features

OCR Capabilities

• Automatically detects PDF or image input
• Performs OCR using the Mistral API
• Exports results as Markdown (.md)
• Optionally includes extracted images
• Displays real-time progress for multi-page documents
• Prevents accidental overwriting of output files

Document Analysis

• Extract structured information using natural language queries
• Visualize named entities with interactive highlighting
• Customize prompts and examples for specific use cases
• Return results in JSON format for easy processing

---

Requirements

• Python 3.8+
• A valid Mistral API key
• A LangChain-compatible language model (for document analysis)
• spaCy with 'en_core_web_sm' model (auto-installed)

---

Installation

pip install docin

The spaCy model 'en_core_web_sm' will be automatically downloaded during installation.

---

Usage

OCR Processing

from ocr import MistOcr

# Initialize with your Mistral API key
ocr = MistOcr(api_key='your_mistral_api_key')

# Run OCR on a PDF or image file
ocr.doc_to_md(
    filename='path/to/document.pdf',
    output_filename='output/result.md',
    include_image=False,        # Include embedded or saved images (optional)
    return_response=False      # Return OCR response (optional)
)

Document Analysis with LangQuery

from ocr.query import LangQuery

# Initialize with your LangChain model
query = LangQuery(llm=your_llm)

# Load and analyze a document
query.load_document("Your document text here")

# Extract information using natural language
response = query.query_document("Find all company names and locations")

# Visualize results (in Jupyter/IPython)
query.render(response)

You can customize the analysis by setting different examples:

# Set custom examples for specific entity types
query.set_examples("""
{
    'companies': ['Example Corp', 'Tech Inc'],
    'dates': ['2023-01-01'],
    'locations': ['New York']
}
""")

---

Output

OCR Output

• Saves extracted text in a Markdown (.md) file
• Creates an images/ folder in the same directory for any extracted images
• Displays progress during export
• Returns an OCR response object when return_response=True

LangQuery Output

• Returns structured JSON with extracted entities
• Provides interactive entity highlighting in Jupyter/IPython
• Supports customizable response formats through examples

---

Supported File Types

• PDF (.pdf)
• Image formats: .jpg, .jpeg, .png, .bmp, .tiff
• LangQuery works with any text content

---

Error Handling

• Raises ValueError for unsupported file types
• Prompts before overwriting existing files
• Logs warnings for missing or invalid image data
• Validates document loading for LangQuery
• Ensures proper model initialization

---

Notes

• For OCR: Use high-resolution images (≥300 DPI) for best accuracy
• Supports multi-page PDFs and large documents
• Extracted images are saved with unique IDs in the images/ directory
• LangQuery works best with well-structured text content
• Entity visualization requires Jupyter/IPython environment

---

Example Output

OCR Output

Markdown file:

# Page 1
Extracted text...

# Page 2
More extracted text...

Images folder:

images/
 ├── image_1.png
 ├── image_2.jpg

LangQuery Output

JSON Response:

{
    "companies": ["Acme Corp", "TechStart Inc"],
    "locations": ["New York", "San Francisco"],
    "dates": ["2023-01-15"]
}

Visual Output:

• Interactive highlighting of found entities in the document text
• Color-coded entity types for easy identification

---

Author

Ime Inyang

License

MIT

Version

0.1.2