Overview

This is a fully operational Intelligent Agentic RAG (Retrieval-Augmented Generation) System built with FastAPI and LangGraph. The system provides advanced document classification, ontological mapping, and comparative analysis capabilities for PDF documents. It uses AI agents to perform multi-step reasoning for intelligent document retrieval, analysis, and cross-document comparison.

The system integrates multiple AI services and frameworks to create a comprehensive document intelligence platform that can understand, classify, and analyze documents across HealthCare Benefits domain while maintaining semantic relationships through ontological structures.

System Architecture

Frontend Architecture

• Static Web Interface: Bootstrap-based responsive UI with JavaScript for dynamic interactions
• Real-time Updates: Event-driven interface with modals for loading states and error handling
• Document Management: Interactive document list with selection capabilities for comparison
• Progressive Enhancement: Fallback support for various browser capabilities

Backend Architecture

• FastAPI Framework: Asynchronous web framework with automatic API documentation
• LangGraph Agents: Multi-step reasoning agents using state graphs for complex workflows
• Service-Oriented Design: Modular services for PDF parsing, document classification, chunking, and vector operations
• Dependency Injection: FastAPI dependency system for service management and lifecycle

AI Agent System

• RAG Agent: Multi-step retrieval and reasoning using LangGraph state machines
• Comparative Agent: Cross-document analysis with structured comparison workflows
• Agentic Workflows: State-based processing with validation and iteration capabilities

Document Processing Pipeline

• PDF Parser: Advanced extraction using pdfplumber with table recognition
• Semantic Chunking: Context-aware document segmentation with overlap management
• Document Classification: AI-powered domain and type classification using Gemini
• Ontological Mapping: Knowledge graph integration for semantic relationships

Vector Storage and Retrieval

• ChromaDB: Persistent vector database for document embeddings
• Hierarchical Indexing: Multi-level indexing by domain, type, and ontological concepts
• Semantic Search: Embedding-based similarity search with metadata filtering

Knowledge Management

• Ontology Manager: Combines RDFLib-based programmatic ontology creation with manual refinement using Protégé, a widely adopted tool for editing and visualizing OWL and RDF schemas.
• Domain-Specific Structures: Maintains distinct ontologies tailored to healthcare, legal, and financial domains, enabling modular reasoning and compliance-aware modeling.
• Concept Mapping: Automates the linking of document content to ontological concepts, with Protégé facilitating validation, class hierarchy adjustments, and semantic consistency checks.

External Dependencies

AI and ML Services

• Google Gemini API: Primary LLM for classification, analysis, and response generation
• Sentence Transformers: Local embedding generation using all-MiniLM-L6-v2 model
• spaCy: Natural language processing for entity recognition and text analysis

Vector Database

• ChromaDB: Persistent vector storage with collection management
• Local Storage: File-based persistence for vector embeddings and metadata

Document Processing

• pdfplumber: PDF parsing with table extraction capabilities
• pandas: Data manipulation for structured content processing

Knowledge Graphs

• RDFLib: RDF graph creation and SPARQL querying
• owlready2: OWL ontology manipulation and reasoning

Web Framework

• FastAPI: Asynchronous web framework with OpenAPI support
• Jinja2: Template rendering for web interface
• Uvicorn: ASGI server for production deployment

Frontend Libraries

• Bootstrap 5: UI component framework
• Font Awesome: Icon library for interface elements

Language Processing

• NLTK: Text processing utilities and stopword management
• scikit-learn: Machine learning utilities for similarity calculations

Agent Framework

• LangGraph: State graph framework for multi-step agent workflows
• LangChain: Tool integration and agent utilities

Environment Variables

Create a .env file in the main directory:

# Required
GEMINI_API_KEY=your_gemini_api_key_here

Development Environment using Python’s native virtual environment

1. Setup and activate Virtual Environment:

   # Create virtual environment
   python -m venv venv
   
   # Activate and you should see '(venv)'
   source venv/Scripts/activate

2. Install dependencies:

   # Navigate to Backend ## Deployment using local Virtual Machine (venv)folder
   ./setup.sh

3. Set environment variables:

   # The "GEMINI_API_KEY" is automatically loaded into the environment from the .env file via python-dotenv, so manual declaration is not required.

4. Run the server:

   # Navigate to the folder containing main.py
   uvicorn main:app --host 0.0.0.0 --port 5000 --reload

5. Close out the server and Virtual Environment:

   # Ctrl + C to terminate API, and run the following in base to terminate virtual envorment
   deactivate
   
   # Remove 'venv' folder under project
   rm -rf venv/

Development Environment using Docker

Docker setup instructions are available in the 'Docker' folder's README.md

1. Start development container:

   cd docker
   docker-compose -f docker-compose.dev.yml up --build

2. Access the application:

   • API: http://localhost:5000
   • Documentation: http://localhost:5000/docs