SurfSense/surfsense_backend/README.md

118 lines
5.4 KiB
Markdown

# Surf Backend
## Technology Stack Overview
This application is a modern AI-powered search and knowledge management platform built with the following technology stack:
### Core Framework and Environment
- **Python 3.12+**: The application requires Python 3.12 or newer
- **FastAPI**: Modern, fast web framework for building APIs with Python
- **Uvicorn**: ASGI server implementation, running the FastAPI application
- **PostgreSQL with pgvector**: Database with vector search capabilities for similarity searches
- **SQLAlchemy**: SQL toolkit and ORM (Object-Relational Mapping) for database interactions
- **FastAPI Users**: Authentication and user management with JWT and OAuth support
### Key Features and Components
#### Authentication and User Management
- JWT-based authentication
- OAuth integration (Google)
- User registration, login, and password reset flows
#### Search and Retrieval System
- **Hybrid Search**: Combines vector similarity and full-text search for optimal results using Reciprocal Rank Fusion (RRF)
- **Vector Embeddings**: Document and text embeddings for semantic search
- **pgvector**: PostgreSQL extension for efficient vector similarity operations
- **Chonkie**: Advanced document chunking and embedding library
- Uses `AutoEmbeddings` for flexible embedding model selection
- `LateChunker` for optimized document chunking based on embedding model's max sequence length
#### AI and NLP Capabilities
- **LangChain**: Framework for developing AI-powered applications
- Used for document processing, research, and response generation
- Integration with various LLM models through LiteLLM
- Document conversion utilities for standardized processing
- **GPT Integration**: Integration with LLM models through LiteLLM
- Multiple LLM configurations for different use cases:
- Fast LLM: Quick responses (default: gpt-4o-mini)
- Smart LLM: More comprehensive analysis (default: gpt-4o-mini)
- Strategic LLM: Complex reasoning (default: gpt-4o-mini)
- Long Context LLM: For processing large documents (default: gemini-2.0-flash-thinking)
- **Rerankers with FlashRank**: Advanced result ranking for improved search relevance
- Configurable reranking models (default: ms-marco-MiniLM-L-12-v2)
- Supports multiple reranking backends (FlashRank, Cohere, etc.)
- Improves search result quality by reordering based on semantic relevance
- **GPT-Researcher**: Advanced research capabilities
- Multiple research modes (GENERAL, DEEP, DEEPER)
- Customizable report formats with proper citations
- Streaming research results for real-time updates
#### External Integrations
- **Slack Connector**: Integration with Slack for data retrieval and notifications
- **Notion Connector**: Integration with Notion for document retrieval
- **Search APIs**: Integration with Tavily and Serper API for web search
- **Firecrawl**: Web crawling and data extraction capabilities
#### Data Processing
- **Unstructured**: Tools for processing unstructured data
- **Markdownify**: Converting HTML to Markdown
- **Playwright**: Web automation and scraping capabilities
#### Main Modules
- **Search Spaces**: Isolated search environments for different contexts or projects
- **Documents**: Storage and retrieval of various document types
- **Chunks**: Document fragments for more precise retrieval
- **Chats**: Conversation management with different depth levels (GENERAL, DEEP)
- **Podcasts**: Audio content management with generation capabilities
- **Search Source Connectors**: Integration with various data sources
### Development Tools
- **Poetry**: Python dependency management (indicated by pyproject.toml)
- **CORS support**: Cross-Origin Resource Sharing enabled for API access
- **Environment Variables**: Configuration through .env files
## Database Schema
The application uses a relational database with the following main entities:
- Users: Authentication and user management
- SearchSpaces: Isolated search environments owned by users
- Documents: Various document types with content and embeddings
- Chunks: Smaller pieces of documents for granular retrieval
- Chats: Conversation tracking with different depth levels
- Podcasts: Audio content with generation capabilities
- SearchSourceConnectors: External data source integrations
## API Endpoints
The API is structured with the following main route groups:
- `/auth/*`: Authentication endpoints (JWT, OAuth)
- `/users/*`: User management
- `/api/v1/search-spaces/*`: Search space management
- `/api/v1/documents/*`: Document management
- `/api/v1/podcasts/*`: Podcast functionality
- `/api/v1/chats/*`: Chat and conversation endpoints
- `/api/v1/search-source-connectors/*`: External data source management
## Deployment
The application is configured to run with Uvicorn and can be deployed with:
```
python main.py
```
This will start the server on all interfaces (0.0.0.0) with info-level logging.
## Requirements
See pyproject.toml for detailed dependency information. Key dependencies include:
- asyncpg: Asynchronous PostgreSQL client
- chonkie: Document chunking and embedding library
- fastapi and related packages
- fastapi-users: Authentication and user management
- firecrawl-py: Web crawling capabilities
- langchain components for AI workflows
- litellm: LLM model integration
- pgvector: Vector similarity search in PostgreSQL
- rerankers with FlashRank: Advanced result ranking
- Various AI and NLP libraries
- Integration clients for Slack, Notion, etc.