OCR Translator

Agentic pipeline that extracts text from images and PDFs, translates it to English, and saves polished output as .docx and .txt.

How It Works

1. Load — PDFs are split into pages; native text is extracted directly, scanned pages are rendered as images
2. OCR — Pages are processed in parallel (5 workers) using a vision LLM as the OCR engine
3. Translate & Refine — Full text is translated to English and cleaned of OCR artifacts in a single LLM call
4. Format & Save — Output is structured as Markdown, then saved as both .docx and .txt

Setup

pip install google-genai openai ollama python-dotenv Pillow PyMuPDF python-docx opencv-python numpy

For the AutoGen framework:

pip install autogen-agentchat autogen-ext autogen-core

Create a .env file with credentials for your chosen engine:

# Google Gemini
GEMINI_API_KEY=...
GEMINI_MODEL=gemini-1.5-pro        # optional

# Azure OpenAI
AZURE_OPENAI_API_KEY=...
AZURE_OPENAI_ENDPOINT=...
AZURE_OPENAI_DEPLOYMENT=gpt-4o     # optional

# Ollama (cloud)
OLLAMA_API_KEY=...                  # omit for local Ollama

# Ollama / DeepSeek (local)
OLLAMA_HOST=http://localhost:11434  # optional
OLLAMA_MODEL=deepseek-r1:latest    # optional

Usage

# Basic — OCR + translate an image or PDF
python main.py document.pdf

# Specify source language hint and engine
python main.py scan.jpg --lang Sanskrit --engine azure

# Use the AutoGen multi-agent framework
python main.py document.pdf --engine google --framework autogen

Engines

Engine	Provider	Requires
google	Google Gemini	GEMINI_API_KEY
azure	Azure OpenAI	AZURE_OPENAI_API_KEY, AZURE_OPENAI_ENDPOINT
ollama	Ollama (cloud/local)	OLLAMA_API_KEY or local Ollama running
deepseek	DeepSeek via Ollama	Local Ollama with models pulled

Frameworks

• simple (default) — Linear pipeline orchestrated directly in main.py
• autogen — Microsoft AutoGen agent that autonomously decides whether to use native text extraction or OCR based on file analysis

Project Structure

main.py                        # Entry point and simple framework pipeline
src/
  engines/
    google/  ocr.py, agents.py # Google Gemini backend
    azure/   ocr.py, agents.py # Azure OpenAI backend
    ollama/  ocr.py, agents.py # Ollama SDK backend
    deepseek/ocr.py, agents.py # DeepSeek (local Ollama) backend
  autogen_flow.py              # AutoGen multi-agent orchestration
  input_handler.py             # PDF/image loading (PyMuPDF)
  image_preprocessing.py       # OpenCV preprocessing for OCR
  docx_saver.py                # Markdown-to-DOCX converter

Supported Formats

Input: PDF, JPG, JPEG, PNG, BMP, TIFF, WebP

Output: .docx and .txt (named with engine suffix, e.g. document_gcp_refined.docx)