A modular multi-agent AI system that performs deep scientific research using a supervisor-worker architecture. It combines foundational and specialized language models to reason, plan, and execute tasks for document and chart analysis in scientific domains.

Agentic Document Intelligence with GPT-4 & SmolDocling

This project is an agentic AI pipeline that uses GPT-4 as the primary agent and integrates SmolDocling as a tool to deeply analyze uploaded documents. It allows users to upload various document formats and extract structured content automatically, with an intelligent evaluation and feedback loop.

🚀 Features

• Upload documents in multiple formats (PDF, Word, etc.)
• Automatic conversion to PDF if needed
• Dual extraction using GPT-4 and SmolDocling
• Evaluation of extracted content using BLEU, overlap, and Jaccard similarity
• Iterative feedback to SmolDocling to improve accuracy
• Final structured output in Word and PDF formats
• User prompt execution on final extracted document
• Streamlit UI for ease of use
• Dockerized for simple deployment

🧠 Pipeline Overview

1. Upload Document: User uploads a file and provides a prompt.

2. Preprocessing: The document is converted to PDF (if not already).

3. GPT-4 Extraction: Extracts text, tables, images, and structural elements.

4. SmolDocling Extraction: Sends static prompt to SmolDocling backend for extraction.

5. Evaluation: Compares GPT-4 vs SmolDocling outputs using:

   • Textual overlap ratio
   • BLEU score
   • Jaccard similarity

6. Consistency Check:

   • ✅ If consistent: build final doc and apply user prompt.
   • ❌ If inconsistent: identify differences and retry SmolDocling with feedback.

7. Final Output: Assemble and export the cleaned, structured document.

🗂️ Folder Structure

.
├── app.py                  # Streamlit UI
├── Dockerfile              # Docker configuration
├── requirements.txt
│
├── graph/                  # LangGraph logic
│   ├── graph_builder.py
│   └── nodes/              # Nodes in the pipeline
│       ├── user_input.py
│       ├── preprocess_doc.py
│       ├── gpt_extract.py
│       ├── smoldocling_call.py
│       ├── evaluate.py
│       ├── retry_node.py
│       ├── fina