📄 AgenticDoc — Multi-Agent Document Understanding System

An end-to-end agent-based pipeline for extracting, structuring, and querying information from academic PDFs using layout analysis, semantic understanding, knowledge graphs, and hybrid retrieval (FAISS + KG).

🚀 Overview

AgenticDoc is designed to transform raw PDFs into structured, queryable knowledge. It combines multiple specialized agents to process documents step by step:

• Extract content from PDFs
• Detect layout (titles, paragraphs, figures, etc.)
• Understand semantics
• Build a knowledge graph
• Enable retrieval using FAISS + Hybrid search
• Generate answers using LLMs

🧠 Architecture

PDF → Extraction → Layout Detection → Semantic Understanding → Knowledge Graph → Retrieval (FAISS + Hybrid) → Answer Generation

Each stage is handled by a dedicated agent for modularity and extensibility.

⚙️ Features

• 📄 PDF text extraction (with OCR fallback)
• 🧩 Layout detection (titles, paragraphs, figures, tables)
• 🧠 Semantic classification (Method, Result, Definition, etc.)
• 🕸️ Knowledge Graph construction (GraphML output)
• 🔍 FAISS-based semantic search
• 🔗 Hybrid retrieval (FAISS + Graph reasoning)
• 🤖 LLM-based answer generation (Ollama / Grok support)
• 📊 Evaluation metrics for retrieval performance
• 🎨 Visualization tools (layout, alignment, graph)

📂 Project Structure

• agents/ # Core pipeline agents
• extractors/ # PDF & layout extraction logic
• output/ # Generated outputs (JSON, graph, * FAISS index)
• tests/ # Unit tests
• plots/ # Visualization outputs
• pdf/ # Input documents
• main.py # Entry point
• pipeline.py # Pipeline orchestration
• ui.py # Streamlit UI
• retrieval_evaluation.py # Evaluation script
• run_*.py # Individual stage runners

🛠️ Installation

1. Clone the repository

git clone https://github.com/saikiranvankudothu/agenticdoc2.git 
cd agenticdoc2

Create virtual environment with uv

uv init 
uv sync

Run full pipeline

python pipeline.py --pdf pdf/paper.pdf

output metrics