📘 Local LLM OCR using Ollama (Free & Offline Image-to-Text Extraction)

This project demonstrates OCR (Optical Character Recognition) using locally running LLMs via Ollama — completely free, private, and offline. No API keys, no external calls, no cloud costs.

It works with any vision-enabled Ollama model such as:

• qwen2.5vl:3b
• llava models
• moondream
• any future Ollama models supporting image input

---

🚀 Features

• 100% free, offline OCR

• Works with any vision-enabled LLM in Ollama

• Supports:

  • 🖼️ Local images
  • 🌐 Online image URLs (download → base64 → LLM)

• Preserves text order and completeness

• Easy to modify for structured JSON output

• Privacy-friendly: image never leaves your machine

---

🛠️ Requirements

• Python 3.8+
• Ollama installed locally 👉 https://ollama.com/download
• A vision-capable model pulled in Ollama:

ollama pull qwen2.5vl:3b
# or any other vision-enabled model

Install Python dependencies:

pip install requests

(Ollama's Python client comes built into the package when Ollama is installed.)

---

📦 Project Structure

✔️ image_to_base64(image_path)

Converts any local image to a Base64 string.

✔️ image_to_text_from_url(image_url)

Downloads the image → converts to Base64 → sends to local LLM.

✔️ image_to_text_from_base64(image_base64)

Sends Base64 image directly to the LLM for OCR.

---

🧠 Why Local LLM OCR?

Traditional OCR tools struggle with:

• Small text
• Handwritten notes
• Blurry/low-quality images
• Mixed text layouts

LLM-based OCR:

• Understands context
• Reconstructs partial text
• Keeps reading order
• Works even on messy images

And with Ollama, you get all that fully offline.

---

🧪 Usage Example

Extract text from a local image

local_image_path = "image.jpg"
image_base64 = image_to_base64(local_image_path)

text = image_to_text_from_base64(image_base64)
print(text)

Extract text from an online image

image_url = "https://example.com/sample.jpg"
text = image_to_text_from_url(image_url)
print(text)

Save output to a text file

with open("extracted_text.txt", "w", encoding="utf-8") as f:
    f.write(text)

---

🔁 Replace the Model Easily

Change this:

model="qwen2.5vl:3b"

To any Ollama vision model:

model="llava:13b"
model="moondream:latest"
model="bakllava"
model="llama3.2-vision"

No other changes needed.

---

🧩 Full Script Included

This repository contains:

• Local image → OCR
• URL image → OCR
• Base64 utilities
• Text saving to .txt

Everything ready to use out of the box.

---

🎯 Use Cases

• Extract text from scanned documents
• Read PDFs (after converting PDF → image)
• OCR for receipts & invoices
• Handwritten note transcription
• Desktop automation
• Data extraction & cleanup

---

🤝 Contributing

Pull requests are welcome! You can extend this to:

• OCR → JSON structuring
• Multi-image batch processing
• CLI tool
• GUI for drag-and-drop OCR

---

📄 License

MIT License — free for personal & commercial use.