Retrieval-Augmented Conversational Agents for Literary Discussion

A sophisticated literary discussion platform that combines the power of Large Language Models (LLMs) with retrieval-augmented generation to provide intelligent, context-aware discussions about books.

Backend Architecture

Technology Stack

• FastAPI: Modern, fast web framework for building APIs
• ChromaDB: Vector database for efficient similarity search and storage
• Sentence Transformers: For generating high-quality text embeddings
• Google Gemini API: For generating contextual responses
• Wikipedia API: For fetching book-related information
• Google Books API: For book metadata and search

Core Components

1. Book Data Preparation Pipeline

The system follows a three-step process to prepare books for discussion:

1. Book Search & Selection

   • Uses Google Books API to search and retrieve book metadata
   • Endpoint: /search-books
   • Returns book details including title, authors, description, and thumbnail

2. Wikipedia Data Fetching

   • Retrieves comprehensive book information from Wikipedia
   • Endpoint: /books/fetch-wiki
   • Stores book summaries and content in JSON format
   • Handles disambiguation and not-found cases gracefully

3. Content Embedding

   • Processes book content into vector embeddings
   • Uses SentenceTransformer ("all-MiniLM-L6-v2") for embedding generation
   • Chunks content into 500-character segments for optimal processing
   • Stores embeddings in ChromaDB for efficient retrieval
   • Endpoint: /books/embed

2. Vector Storage & Retrieval

• ChromaDB Integration

  • Persistent vector storage in vectorstore directory
  • Each book gets its own collection
  • Efficient similarity search for context retrieval
  • Automatic collection management and persistence

• Embedding Process

  • Content chunking for manageable segments
  • High-quality embeddings using SentenceTransformer
  • Unique IDs for each chunk
  • Automatic collection creation and management

3. Query Processing System

The system implements a sophisticated query processing pipeline:

1. Context Retrieval

   • Converts user questions into embeddings
   • Retrieves most relevant book chunks using similarity search
   • Combines retrieved context with conversation history

2. Response Generation

   • Uses Google Gemini API for response generation
   • Incorporates:
     • Retrieved book context
     • Conversation history
     • Book metadata
     • Wikipedia information
   • Generates concise, contextually relevant responses

3. Response Management

   • Maintains conversation history
   • Compresses responses for efficient storage
   • Handles error cases gracefully

API Endpoints

Book Management

• GET /search-books: Search for books using Google Books API
• POST /books/fetch-wiki: Fetch Wikipedia data for a book
• POST /books/embed: Generate and store embeddings for a book
• GET /books/check: Check if a book is prepared for discussion
• POST /books/prepare: Prepare a book for discussion (combines wiki fetch and embedding)

Discussion

• POST /chat/query: Process user questions and generate responses
  • Input: book_id, question, conversation history
  • Output: AI response with updated history

Data Flow

1. Book Preparation

   User Request → Book Search → Wikipedia Fetch → Content Embedding → Vector Storage
   

2. Query Processing

   User Question → Embedding → Context Retrieval → Response Generation → User Response
   

Security & Environment

• Environment variables for API keys
• CORS middleware for secure cross-origin requests
• Error handling and validation at each step
• Persistent storage with proper directory management

Frontend Overview

The frontend is a modern web application that provides:

• Book search and selection interface
• Real-time chat interface for book discussions
• Progress tracking for book preparation
• Responsive design for various devices

Getting Started

Prerequisites

• Python 3.10+
• Required API keys:
  • Google Books API
  • Google Gemini API
  • (Optional) Wikipedia API

Environment Setup

1. Create a .env file with required API keys:

   GOOGLE_API_KEY=your_google_books_api_key
   GEMINI_API_KEY=your_gemini_api_key
   

2. Install dependencies:

   pip install -r requirements.txt

3. Start the backend server:

   uvicorn backend.main:app --reload

Technical Approach

Retrieval-Augmented Generation (RAG)

The system implements RAG to enhance LLM responses with relevant book context:

1. Retrieval: Uses vector similarity search to find relevant book passages
2. Augmentation: Combines retrieved context with user query
3. Generation: Uses Gemini to generate contextually relevant responses

Vector Search Optimization

• Chunk size optimization (500 characters)
• Efficient embedding storage and retrieval
• Persistent vector database for quick access
• Automatic collection management

Error Handling & Resilience

• Graceful handling of API failures
• Proper error messages for users
• Automatic retry mechanisms
• Data validation at each step

Future Improvements

1. Enhanced Context Retrieval

   • Implement more sophisticated chunking strategies
   • Add metadata-aware retrieval
   • Implement hybrid search (keyword + semantic)

2. Performance Optimization

   • Implement caching for frequent queries
   • Optimize vector search parameters
   • Add batch processing for embeddings

3. Additional Features

   • Support for multiple languages
   • Enhanced book metadata integration
   • Advanced conversation management
   • User feedback integration