DeepSeek Visor Agent

Production-ready wrapper for DeepSeek-OCR - Convert documents to structured data in 3 lines of code

Keywords: DeepSeek OCR, DeepSeek-OCR wrapper, document OCR, AI agent vision tool, LangChain OCR, LlamaIndex OCR

---

⚠️ GPU Requirements (CRITICAL)

NVIDIA GPU with Turing+ architecture required

✅ Supported	❌ Not Supported
RTX 20/30/40 series (Turing/Ampere/Ada)	GTX 10 series (Pascal - no FlashAttention)
Tesla T4, A10, A100	GTX 1080 Ti, GTX 1660
Minimum: RTX 2060 (6GB VRAM)	CPU-only mode
Recommended: RTX 3090 (24GB VRAM)	AMD GPUs (ROCm)

Why? DeepSeek-OCR requires FlashAttention 2.x, which only supports compute capability 7.5+ (Turing and newer).

No GPU? Join our hosted API waitlist (planned for future release).

📖 Detailed compatibility guide: GPU_COMPATIBILITY.md

---

🎯 What is This?

DeepSeek Visor Agent is a production-ready Python wrapper for DeepSeek-OCR, the state-of-the-art open-source OCR model by DeepSeek AI.

Built on DeepSeek-OCR, this wrapper makes document understanding effortless for AI agents by handling all the complexity:

• ✅ Auto device detection (CUDA with Turing+ GPUs)
• ✅ Automatic fallback (Gundam mode → Base mode → Tiny mode when OOM)
• ✅ Structured output (Markdown + extracted fields)
• ✅ Agent-ready (LangChain, LlamaIndex, Dify compatible)

⚡ Quick Start

Prerequisites

Before installation, ensure you have:

1. NVIDIA GPU with Turing+ architecture (RTX 20/30/40 series, Tesla T4/A100)
2. CUDA 11.8+ installed and configured
3. Python 3.9+

Installation

Step 1: Install the package

pip install deepseek-visor-agent

Step 2: (First-time only) Model download

The first time you run the tool, it will automatically download the DeepSeek-OCR model (~6.2 GB) from HuggingFace:

from deepseek_visor_agent import VisionDocumentTool

# This will trigger model download on first run
tool = VisionDocumentTool()

The model will be cached in ~/.cache/huggingface/ and reused for subsequent runs.

Step 3: (Optional) Install FlashAttention for better performance

# For RTX GPUs with compute capability 7.5+
pip install flash-attn --no-build-isolation

Basic Usage

Process Images:

from deepseek_visor_agent import VisionDocumentTool

# Initialize the tool (auto-detects best device and model)
tool = VisionDocumentTool()

# Process a document image
result = tool.run("invoice.jpg")

print(result["fields"]["total"])  # "$199.00"
print(result["fields"]["date"])   # "2024-01-15"
print(result["document_type"])    # "invoice"

Process PDFs:

# PDF files work the same way - automatically converts pages to images
result = tool.run("contract.pdf")

print(f"Processed {result['pages']} pages")
print(result["markdown"])  # Multi-page PDFs have <--- Page Split ---> separators

# Process specific pages only
result = tool.run("long_document.pdf", pdf_start_page=0, pdf_end_page=2)

That's it! No configuration needed.

📖 Complete User Journey

Scenario 1: Standalone Python Script

Use Case: Extract invoice data for accounting automation

from deepseek_visor_agent import VisionDocumentTool
import json

# Initialize once
tool = VisionDocumentTool()

# Process multiple invoices
invoices = ["invoice1.jpg", "invoice2.pdf", "invoice3.png"]
results = []

for invoice_path in invoices:
    result = tool.run(invoice_path, document_type="invoice")
    results.append({
        "file": invoice_path,
        "total": result["fields"]["total"],
        "vendor": result["fields"]["vendor"],
        "date": result["fields"]["date"]
    })

# Export to JSON
with open("extracted_invoices.json", "w") as f:
    json.dump(results, f, indent=2)

print(f"Processed {len(results)} invoices")

Timeline:

• First run: ~30 seconds (model download + first inference)
• Subsequent runs: ~5-7 seconds per page

Scenario 2: LangChain AI Agent

Use Case: Build a chatbot that can answer questions about uploaded documents

from langchain.tools import tool
from langchain.agents import initialize_agent, AgentType
from langchain_openai import ChatOpenAI
from deepseek_visor_agent import VisionDocumentTool

# Initialize OCR tool
ocr_tool = VisionDocumentTool()

@tool
def analyze_document(image_path: str) -> dict:
    """Analyze any document image and extract structured data"""
    return ocr_tool.run(image_path, document_type="auto")

# Create agent
llm = ChatOpenAI(model="gpt-4", temperature=0)
agent = initialize_agent(
    [analyze_document],
    llm,
    agent=AgentType.ZERO_SHOT_REACT_DESCRIPTION,
    verbose=True
)

# User interaction
response = agent.run("What is the total amount in invoice.jpg?")
print(response)
# Output: "The total amount in the invoice is $199.00, dated 2024-01-15 from Acme Corp."

User Flow:

1. User uploads document image via chat interface
2. Agent calls analyze_document tool
3. DeepSeek-OCR extracts text + fields
4. LLM interprets results and responds naturally

Scenario 3: Batch PDF Processing

Use Case: Process hundreds of multi-page contracts

from deepseek_visor_agent import VisionDocumentTool
from pathlib import Path
import json

tool = VisionDocumentTool()

# Find all PDFs
contracts_dir = Path("./contracts/")
pdf_files = list(contracts_dir.glob("*.pdf"))

results = []
for pdf_path in pdf_files:
    print(f"Processing {pdf_path.name}...")

    result = tool.run(
        str(pdf_path),
        document_type="contract",
        pdf_start_page=0,  # Process first 3 pages only
        pdf_end_page=2
    )

    results.append({
        "filename": pdf_path.name,
        "parties": result["fields"]["parties"],
        "effective_date": result["fields"]["effective_date"],
        "pages_processed": result["pages"]
    })

# Save results
with open("contracts_summary.json", "w") as f:
    json.dump(results, f, indent=2)

print(f"Processed {len(results)} contracts")

Performance: ~6-7 seconds per page on Tesla T4

Scenario 4: REST API for No-Code Platforms (Dify/Flowise)

Use Case: Integrate with Dify for visual workflow builder

See Dify Integration Guide for complete setup.

High-level flow:

1. Deploy FastAPI wrapper (provided in examples)
2. Configure Dify HTTP node with OCR endpoint
3. Build visual workflow: Upload → OCR → Parse → Respond
4. No Python code needed for end users

🔗 Integrations

LangChain

from langchain.tools import tool
from deepseek_visor_agent import VisionDocumentTool

ocr_tool = VisionDocumentTool()

@tool
def extract_invoice_data(image_path: str) -> dict:
    """Extract structured data from invoice images"""
    return ocr_tool.run(image_path, document_type="invoice")

# Use in your agent
from langchain.agents import initialize_agent, AgentType
from langchain.llms import OpenAI

tools = [extract_invoice_data]
agent = initialize_agent(tools, OpenAI(temperature=0), agent=AgentType.ZERO_SHOT_REACT_DESCRIPTION)

response = agent.run("Extract the total from invoice.jpg")

LlamaIndex

from llama_index.tools import FunctionTool
from deepseek_visor_agent import VisionDocumentTool

tool = VisionDocumentTool()

def ocr_document(image_path: str) -> dict:
    """Process documents with OCR"""
    return tool.run(image_path)

llama_tool = FunctionTool.from_defaults(fn=ocr_document)

Dify / Flowise

See integration guide for REST API setup.

📊 Features

Automatic Device Management

The tool automatically detects your hardware and selects the optimal configuration:

Hardware	Inference Mode	Memory Usage
RTX 4090 (24GB)	Gundam	~10GB
RTX 3090 (24GB)	Base	~6GB
RTX 2060 (6GB)	Tiny	~3GB
CPU only	Not Supported	-

Automatic Fallback

If inference fails (OOM, CUDA errors), automatically falls back to lower-resolution modes:

Gundam mode (OOM) → Large mode → Base mode → Small mode → Tiny mode (Success!)

Supported Document Types

• ✅ Invoices - Extracts total, date, vendor, line items
• ✅ Contracts - Extracts parties, effective date, terms
• ✅ PDF Documents - Multi-page PDFs with automatic page splitting
• 🚧 Resumes - Coming soon
• 🚧 Forms - Coming soon

PDF Support

Based on DeepSeek-OCR official implementation:

• ✅ Multi-page PDF processing
• ✅ Automatic page-to-image conversion (PyMuPDF)
• ✅ Configurable DPI (default: 144, same as official)
• ✅ Page range selection
• ✅ Same API as image processing

# Process entire PDF
result = tool.run("contract.pdf")

# Process specific pages (0-indexed)
result = tool.run("doc.pdf", pdf_start_page=0, pdf_end_page=2)

# Adjust quality
result = tool.run("scan.pdf", pdf_dpi=200)  # Higher DPI = better quality

Output Format

{
    "markdown": "# Invoice\n\nDate: 2024-01-15\n...",
    "fields": {
        "total": "$199.00",
        "date": "2024-01-15",
        "vendor": "Acme Corp"
    },
    "document_type": "invoice",
    "confidence": 0.95,
    "metadata": {
        "inference_mode": "tiny",
        "device": "cuda",
        "inference_time_ms": 1823
    },
    "pages": 1  # Number of pages processed (1 for images, N for PDFs)
}

⚡ Performance

✅ GPU-Tested on Tesla T4 (16GB VRAM) - 2025-10-21

Inference Mode	Inference Time	Test Environment	Notes
Tiny	5.35s/page	Tesla T4, Simple Doc	Fastest, 64 tokens
Small	6.53s/page	Tesla T4, Simple Doc	100 tokens
Base	6.77s/page	Tesla T4, Simple Doc	256 tokens, Most Common
Large	6.35s/page	Tesla T4, Simple Doc	400 tokens
Gundam	6.67s/page	Tesla T4, Simple Doc	Crop mode, 256+400 tokens

⚠️ Note: Performance tested on simple text documents. Real-world complex documents (tables, images, forms) may vary.

📚 Documentation

🚀 Getting Started

• 📚 Documentation Center - Complete documentation hub
• GPU Compatibility Guide
• Hardware Limitations

🏗️ For Developers

• Hardware Limitations
• Dify Integration
• LangChain Example
• LlamaIndex Example
• PDF Processing Example

🛣️ Roadmap

• Core OCR engine with auto-fallback
• Invoice parser
• Contract parser (basic)
• PDF support (via PyMuPDF, official DeepSeek-OCR method)
• Resume parser
• Multi-language support
• Hosted API (Cloud version)
• LlamaIndex native tool
• Dify plugin

🤝 Contributing

We welcome contributions! Areas where help is needed:

1. New parsers - Add support for new document types
2. Testing - More test cases and edge cases
3. Documentation - Improve guides and examples
4. Performance - Optimization suggestions

Please submit issues or pull requests on GitHub.

📖 Citation

Built on top of DeepSeek-OCR:

@misc{deepseek-ocr,
  author = {DeepSeek AI},
  title = {DeepSeek-OCR},
  year = {2025},
  url = {https://huggingface.co/deepseek-ai/DeepSeek-OCR}
}

📄 License

Apache License 2.0 - see LICENSE file for details.

🙏 Acknowledgments

• DeepSeek AI team for the amazing OCR model
• Hugging Face for model hosting
• LangChain and LlamaIndex communities for inspiration

📬 Contact

• GitHub Issues: Report bugs or request features
• Email: jack_ai@qq.com

---

Star ⭐ this repo if you find it useful!