A comprehensive Retrieval-Augmented Generation system built with LangChain and LangGraph frameworks, supporting multi-format document parsing, advanced chunking strategies, hybrid retrieval mechanisms, and structured answer generation with comprehensive logging and performance monitoring.

This RAG system provides a complete document processing and question-answering pipeline that can handle PDF, PPTX, and Excel files. The system utilizes ChromaDB for vector storage, OpenAI GPT-4.1-mini for language generation, text-embedding-3-small for embeddings, and implements sophisticated text chunking with cross-page awareness. The workflow is orchestrated using LangGraph state machines for reliable and scalable processing.

graph TD
    A[Document Input] --> B[Document Parsing]
    B --> C[Text Chunking]
    C --> D[Vector Embedding]
    D --> E[Vector Database Storage]
    
    F[User Query] --> G[Vector Retrieval]
    G --> H[Parent Page Aggregation]
    H --> I[LLM Reranking]
    I --> J[Answer Generation]
    J --> K[Structured Response]
    K --> L[Logging & Monitoring]
    
    E --> G
    
    subgraph "Document Processing Pipeline"
        B1[PDF Parser] --> B
        B2[PPTX Parser] --> B
        B3[Excel Parser] --> B
    end
    
    subgraph "Retrieval Pipeline"
        G1[Vector Search] --> G
        H1[Cross-page Aggregation] --> H
    end
    
    subgraph "Generation Pipeline"
        I1[Context Assembly] --> I
        I2[Prompt Engineering] --> I
        I3[Response Validation] --> I
    end

1. Document Ingestion: Multi-format document parsing with layout detection and content extraction
2. Text Chunking: Advanced chunking with cross-page awareness and parent-child relationships
3. Vector Embedding: Batch processing with token limit management using text-embedding-3-small
4. Storage: Persistent vector database with metadata preservation using ChromaDB
5. Retrieval: Vector-based semantic search with similarity scoring
6. Parent Aggregation: Cross-page chunk aggregation to parent page level
7. Reranking: LLM-based relevance scoring for optimal context selection
8. Generation: Structured answer generation with confidence scoring and source attribution

# Create conda environment
conda create -n rag python=3.10
conda activate rag

# Install dependencies
pip install -r requirements.txt

Create a .env file in the project root:

OPENAI_API_KEY=your_openai_api_key_here
ENABLE_TELEMETRY=false  # Optional: disable telemetry for privacy

Key libraries include:

• langchain and langchain-openai for LLM integration
• langgraph for workflow orchestration
• chromadb for vector database
• openai GPT-4.1-mini for language generation
• text-embedding-3-small for document embeddings
• pdfplumber and pypdf for PDF processing
• python-pptx for PowerPoint processing
• pandas for Excel processing
• tiktoken for token counting

# Single question processing
python main.py "What are the key financial metrics mentioned in the documents?"

# Start interactive session
python main.py

# Follow prompts to add documents and ask questions
# Type 'quit' to exit

The system supports multiple document formats:

• PDF: Advanced layout detection with multi-column support
• PPTX: Complete content extraction including tables, charts, and images
• Excel: Multi-sheet processing with data preservation

• Environment variable handling
• Model configuration and constants
• System-wide settings management

• PDFParser: Advanced PDF text extraction with layout analysis
• PPTXParser: Comprehensive PowerPoint content extraction
• ExcelParser: Multi-sheet Excel data processing
• UnifiedDocumentParser: Format detection and routing

• CrossPageTextSplitter: Context-aware chunking across page boundaries
• ParentPageAggregator: Hierarchical chunk organization
• Token-aware splitting with configurable overlap

• VectorStoreManager: Persistent storage with metadata recovery
• Batch processing for large document sets
• Automatic retry logic for API rate limits

• VectorRetriever: Semantic similarity search using embeddings
• ParentPageAggregator: Cross-page chunk aggregation to parent pages
• LLMReranker: GPT-4.1-mini based relevance scoring and reranking
• HybridRetriever: Complete retrieval pipeline orchestration
• Configurable retrieval parameters and scoring weights

• AnswerGenerator: Structured response generation using GPT-4.1-mini
• Confidence scoring and uncertainty handling
• Source attribution and reasoning chains

• LangGraph state machine implementation
• Separate pipelines for document processing and querying
• Error handling and state management

• Multi-column PDF layout detection
• Table and chart extraction from presentations
• Cross-page text chunking with context preservation
• Metadata-rich document representation

• Vector-based semantic search with similarity scoring
• Parent page aggregation for cross-page chunk handling
• LLM-powered relevance reranking using GPT-4.1-mini
• Configurable retrieval parameters and scoring weights
• Source document tracking with page-level attribution

• Confidence-scored responses using GPT-4.1-mini
• Step-by-step reasoning chains
• Source attribution with page references
• Uncertainty acknowledgment

• Comprehensive query logging
• Processing time tracking
• Token usage monitoring
• Error rate analysis

• Batch processing for large document sets
• Persistent vector database with incremental updates
• Automatic retry logic for API failures
• Memory-efficient chunking strategies

The system follows a modular architecture with clear separation of concerns:

• Data Layer: Document parsing and storage management
• Processing Layer: Text chunking and vector embedding
• Retrieval Layer: Vector search, parent aggregation, and LLM reranking
• Generation Layer: LLM-based answer synthesis using GPT-4.1-mini
• Orchestration Layer: Workflow management and error handling

Each module is designed for independent testing and maintenance, with well-defined interfaces and comprehensive error handling.

• Batch processing prevents OpenAI API token limit violations
• Persistent vector storage eliminates reprocessing overhead
• Vector-based retrieval provides fast semantic search
• Parent page aggregation reduces redundant content
• Token counting prevents context window overflow
• Incremental document addition for large knowledge bases

The system provides comprehensive logging including:

• Query processing times and token usage
• Retrieval effectiveness metrics
• Generation quality indicators
• Error tracking and debugging information
• Performance analytics for optimization

The modular design supports easy extension:

• Additional document format parsers
• Custom chunking strategies
• Alternative embedding models
• Enhanced retrieval algorithms
• Specialized generation pipelines