📄 DocIntel Pro: Agentic Document Analysis Engine

DocIntel Pro is a high-throughput, Agentic Document Processing System designed to transform unstructured documents (PDFs, Images) into structured, actionable intelligence. It leverages state-of-the-art Visual Language Models (VLMs) like Qwen2-VL (via Fireworks AI) to perform unified layout analysis, OCR, and semantic understanding in a single pass.

"Visual First, Text Second." Unlike traditional OCR pipelines that lose context, DocIntel Pro understands the visual structure of your document—charts, tables, seals, and handwriting—preserving the semantic meaning of your data.

---

🏗️ System Architecture

The system follows a Cloud-Native Microservices Architecture, orchestrated by a central workflow engine.

High-Level Data Flow

flowchart TB
    User((User)) -->|Uploads Doc| UI["Frontend (Next.js)"]
    UI -->|POST /analyze| Orch{"Orchestrator Service"}
    
    subgraph Core_Services [Backend Services]
        direction TB
        Orch -->|1. Convert/Normalize| Preprocess[Preprocessing Service]
        Orch -->|2. Parallel Analysis| Visual[Visual Intelligence Service]
        
        subgraph Visual_Logic [Visual Logic]
            Visual -->|Extract| VLM["Fireworks AI (Qwen2-VL)"]
            VLM -->|JSON| Parser[Layout Parser]
        end
    end
    
    Preprocess -->|Clean Images| Orch
    Visual -->|Structured Blocks| Orch
    
    Orch -->|3. Aggregate & Order| Response[Final JSON Response]
    Response --> UI

Loading

Microservices Breakdown

Service	Port	Tech Stack	Responsibilities
Frontend	:3000	Next.js, React, Tailwind, Shadcn/UI	Modern, responsive UI for uploading documents and visualizing results with bounding boxes.
Orchestrator	:8000	FastAPI, Python, AsyncIO	Workflow engine. Handles file intake, routing to workers, error handling, and results aggregation.
Preprocessing	:8001	FastAPI, OpenCV, pdf2image	CPU-bound image operations: PDF-to-Image conversion, Denoising, Deskewing, Normalization.
Visual Intelligence	:8002	FastAPI, Fireworks AI SDK	GPU-accelerated inference. Interfaces with VLMs to detect layout, extract text, and recognize tables/figures in one step.

---

🚀 Key Features

• Unified Analysis: Uses Vision-Language Models to perform Layout Analysis and OCR simultaneously.
• Multi-Page Support: Handles multi-page PDFs with parallel processing for high performance.
• Visual Grounding: Returns precise bounding boxes for every detected element (Text, Tables, Figures).
• Resilient Architecture: Exponential backoff and retries for external API calls; isolated services preventing cascade failures.
• Modern UI: A beautiful, responsive interface with real-time feedback and visual overlays.

---

🛠️ Getting Started

Prerequisites

• Python 3.10+ (Recommended: Use Conda)
• Node.js 18+ & npm / bun
• PopplerUtils (Required for PDF processing)
  • Ubuntu: sudo apt-get install poppler-utils
  • MacOS: brew install poppler
• Fireworks AI API Key: Get one at fireworks.ai

1. Environment Setup

Clone the repository and create a virtual environment:

git clone https://github.com/yourusername/docintel-pro.git
cd docintel-pro

# Create Conda Environment
conda create -n doc_analysis_env python=3.10 -y
conda activate doc_analysis_env

# Install Backend Dependencies
pip install -r preprocessing_service/requirements.txt
pip install -r visual_service/requirements.txt
pip install -r orchestrator/requirements.txt (if present, or install commonly used libs)
# Note: Ensure you have standard libs: fastapi, uvicorn, python-multipart, httpx, opencv-python-headless, pdf2image, fireworks-ai

2. Configuration

Create a .env file in the root directory (or ensure services indicate where to look). For the Visual Service, export your API Key:

export FIREWORKS_API_KEY="your_api_key_here"

3. Run the Services

It is recommended to run each service in a separate terminal window or use a process manager (like Supervisord or Docker Compose - coming soon).

Terminal 1: Preprocessing Service

uvicorn preprocessing_service.main:app --port 8001 --reload

Terminal 2: Visual Intelligence Service

uvicorn visual_service.main:app --port 8002 --reload

Terminal 3: Orchestrator

uvicorn orchestrator.main:app --port 8000 --reload

Terminal 4: Frontend

cd frontend
npm install
npm run dev

The UI will be available at http://localhost:3000.

---

🔮 Roadmap

• Docker Compose: One-click startup for the entire stack.
• Database Integration: Persist analysis results (PostgreSQL/MongoDB).
• Streaming Responses: Real-time progress updates for long documents.
• Local Inference: Support for local VLMs (e.g., Llava, BakLLaVA) via Ollama.

---

🤝 Contributing

Contributions are welcome! Please fork the repository and submit a Pull Request.

📄 License

This project is licensed under the MIT License.