Document Q&A SaaS Backend

A Node.js/TypeScript backend for a Document Q&A SaaS application that allows users to upload documents (PDF, DOCX, TXT) and ask questions about their content using AI.

Features

• Document Processing: Support for PDF, DOCX, and TXT files
• Text Chunking: Intelligent text splitting with overlap for better context
• Vector Embeddings: Using transformers.js for local embedding generation
• Q&A System: Local Llama model integration for intelligent question answering
• RAG Pipeline: Retrieval-Augmented Generation for accurate answers
• TypeScript: Full type safety with strict configuration
• SQLite Database: Prisma ORM with SQLite for development

Tech Stack

• Runtime: Node.js 18+ with TypeScript 5+
• Framework: Express.js with TypeScript
• Database: SQLite with Prisma ORM
• Document Processing: pdf-parse, mammoth
• Embeddings: @xenova/transformers (all-MiniLM-L6-v2)
• Q&A: Local Llama model
• File Upload: Multer with validation

Quick Start

Prerequisites

• Node.js 18+
• npm or yarn
• Local Llama model file

Installation

1. Clone and install dependencies:

   npm install

2. Set up environment variables:

   cp .env.example .env

   Edit .env and configure your local Llama model:

   LLAMA_MODEL_PATH="./models/llama-2-7b-chat.gguf"
   LLAMA_MODEL_TYPE="llama-2-7b-chat"

3. Set up the database:

   npm run db:generate
   npm run db:migrate

4. Start the development server:

   npm run dev

The server will start on http://localhost:3001

API Endpoints

Health Check

• GET /health - Server health status

Documents

• GET /api/documents - List all documents
• GET /api/documents/:id - Get specific document
• POST /api/documents - Upload new document
• DELETE /api/documents/:id - Delete document
• GET /api/documents/:id/chunks - Get document chunks

Q&A

• POST /api/qa/ask - Ask question about document
• GET /api/qa/models - Get AI model information
• POST /api/qa/batch - Ask multiple questions

Usage Examples

Upload a Document

curl -X POST http://localhost:3001/api/documents \
  -F "file=@document.pdf" \
  -F "userId=user123"

Ask a Question

curl -X POST http://localhost:3001/api/qa/ask \
  -H "Content-Type: application/json" \
  -d '{
    "documentId": "doc_id_here",
    "question": "What is the main topic of this document?",
    "userId": "user123"
  }'

Get Document List

curl http://localhost:3001/api/documents

Project Structure

src/
├── index.ts              # Main server entry point
├── types/                # TypeScript type definitions
│   └── index.ts
├── middleware/           # Express middleware
│   └── errorHandler.ts
├── services/             # Business logic services
│   ├── documentProcessor.ts
│   ├── embeddingService.ts
│   └── qaService.ts
├── routes/               # API route handlers
│   ├── documents.ts
│   └── qa.ts
└── prisma/               # Database schema
    └── schema.prisma

Development

Available Scripts

• npm run dev - Start development server with hot reload
• npm run build - Build for production
• npm run start - Start production server
• npm run type-check - TypeScript type checking
• npm run db:migrate - Run database migrations
• npm run db:studio - Open Prisma Studio
• npm test - Run tests

Environment Variables

Variable	Description	Default
PORT	Server port	3001
NODE_ENV	Environment	development
DATABASE_URL	Database connection	file:./dev.db
LLAMA_MODEL_PATH	Path to Llama model file	./models/llama-2-7b-chat.gguf
LLAMA_MODEL_TYPE	Llama model type	llama-2-7b-chat
MAX_FILE_SIZE	Max file size (bytes)	10485760 (10MB)
EMBEDDING_MODEL	Embedding model	Xenova/all-MiniLM-L6-v2

Database Schema

The application uses Prisma with the following models:

• User: User accounts (basic implementation)
• Document: Uploaded documents with metadata
• Chunk: Text chunks with embeddings for vector search
• Usage: Usage tracking for analytics

File Processing Pipeline

1. Upload: File validation and storage
2. Extraction: Text extraction based on file type
3. Chunking: Intelligent text splitting with overlap
4. Embedding: Vector embedding generation
5. Storage: Database storage with relationships

Q&A Pipeline

1. Question: User submits question
2. Embedding: Generate embedding for question
3. Search: Vector similarity search for relevant chunks
4. Context: Prepare context from relevant chunks
5. Answer: Generate answer using local Llama model
6. Response: Return structured response with sources

Error Handling

The application includes comprehensive error handling:

• Input validation with detailed error messages
• File processing error handling
• API error responses with appropriate HTTP status codes
• Database error handling
• Graceful server shutdown

Security

• Helmet.js for security headers
• CORS configuration
• File type validation
• File size limits
• Input sanitization

Performance

• Efficient text chunking with overlap
• Local embedding generation (no external API calls)
• Database indexing for fast queries
• Memory-efficient file processing

Next Steps

This is the MVP backend. Future enhancements include:

• User authentication and authorization
• Rate limiting
• Caching layer
• Advanced vector database (Pinecone/Chroma)
• Multi-document conversations
• Usage analytics and billing
• Admin dashboard
• API documentation with Swagger

Contributing

1. Fork the repository
2. Create a feature branch
3. Make your changes
4. Add tests
5. Submit a pull request

License

MIT License - see LICENSE file for details