Docvoice

Transform Documentation into Intelligent Voice Conversation

Comment

💡 Inspiration

Our inspiration came from the growing need for intelligent, accessible knowledge management in today's digital world. We observed that:

  • Customer Support Teams spend hours searching through documentation to answer questions
  • Developers struggle to find relevant information in vast codebases and documentation
  • End Users prefer natural voice interactions over complex search interfaces
  • Businesses need scalable solutions that can handle growing content without performance degradation

We were inspired by the potential of TiDB Cloud's Vector Search to revolutionise how people access and interact with information, combined with the power of OpenAI's language models and real-time voice communication.

🎯 What it does

docvoice - Turn docs into voice transforms static websites into intelligent, voice-enabled knowledge systems:

Core Functionality

  • 🕷️ Intelligent Web Crawling: Automatically crawls websites and extracts meaningful content
  • 🔍 Semantic Vector Search: Uses TiDB Cloud's vector search for context-aware information retrieval
  • 🤖 AI-Powered Q&A: Generates intelligent answers using OpenAI's GPT-4 model
  • 🎤 Voice Interface: Natural voice conversations with AI agents
  • 📱 Widget Integration: Easy-to-embed voice agents on any website
  • 📊 Real-time Indexing: Live content processing and vector database updates

Use Cases

  • Customer Support: Instant voice-based support using company documentation
  • Developer Documentation: Voice search through technical documentation and APIs
  • Educational Content: Interactive learning through voice conversations
  • Knowledge Management: Intelligent search through internal company knowledge bases
  • E-commerce Support: Voice assistance for product information and FAQs

🏗️ How we built it

Architecture Overview

We built a dual-architecture system with frontend and backend components working together:

┌─────────────────┐    ┌─────────────────┐    ┌─────────────────┐
│   Frontend      │    │    Backend      │    │   TiDB Cloud    │
│   (Next.js)     │◄──►│   (Python)      │◄──►│   Vector DB     │
│                 │    │                 │    │                 │
│ • Web Interface │    │ • Voice Agent   │    │ • Vector Search │
│ • Widget System │    │ • Speech-to-Text│    │ • Embeddings    │
│ • Content Mgmt  │    │ • AI Processing │    │ • Real-time     │
└─────────────────┘    └─────────────────┘    └─────────────────┘

Frontend Development

  • Next.js 14: Modern React framework with App Router for optimal performance
  • TypeScript: Type-safe development ensuring code reliability
  • Tailwind CSS: Responsive design system for beautiful UI/UX
  • LiveKit Integration: Real-time voice communication components
  • Widget System: Iframe-based integration for easy website embedding

Backend Development

  • Python Backend: High-performance voice processing server
  • Deepgram Integration: Advanced speech-to-text capabilities
  • OpenAI API: GPT-4 integration for intelligent responses
  • LiveKit Server: Real-time communication infrastructure
  • TiDB Integration: Vector search and database operations

Vector Search Implementation

  • Content Chunking: Intelligent splitting of content into semantic chunks
  • OpenAI Embeddings: text-embedding-3-large for vector generation
  • TiDB Cloud: Enterprise-grade vector database with real-time indexing
  • Semantic Search: Context-aware information retrieval

Voice Processing Pipeline

  1. Speech Input → Deepgram STT → Text
  2. Text Query → OpenAI Embedding → Vector
  3. Vector Search → TiDB Cloud → Relevant Content
  4. Content Processing → OpenAI GPT-4 → Intelligent Response
  5. Response → Text-to-Speech → Voice Output

📊 How docvoice Uses TiDB Cloud

docvoice leverages TiDB Cloud's vector search as its core AI knowledge system:

Database Schema & Storage:

  • enhanced_chunks: Stores website content chunks with 1536-dimensional vector embeddings

    • content: Raw text chunks from web pages
    • embedding VECTOR(1536): OpenAI text-embedding-3-large vectors for semantic search
    • metadata JSON: Page titles, URLs, and indexing information
    • url & page_title: Source attribution for answers

  • url_sources: Manages websites to be indexed

    • indexing_mode: Simple vs enhanced crawling options
    • max_pages & max_depth: Crawling limits and depth control
    • chunks_count & pages_count: Indexing statistics

  • agents: Stores AI voice agent configurations

    • personality, conversation_style, system_prompt
    • llm_model, stt_model, tts_voice_id settings
    • search_limit, context_window for response generation

  • agent_url_assignments: Links agents to their knowledge bases

    • assignment_type: Primary vs secondary URL sources
    • search_priority: Order of importance for content retrieval

  • indexing_jobs: Tracks website crawling progress and status

Vector Search Implementation:

  • Real-time Embedding: Content is chunked and embedded using OpenAI's latest model
  • Semantic Retrieval: Vector similarity search finds most relevant content chunks
  • Hybrid Search: Combines vector search with traditional text search for comprehensive results
  • Performance: Sub-second queries with TiDB Cloud's distributed architecture

TiDB Cloud Account: satyasandeep786@gmail.com

Built With

Share this project:

Updates