Skip to content

replit.md

Latest commit

 

History

History
80 lines (62 loc) · 4.93 KB

replit.md

File metadata and controls

80 lines (62 loc) · 4.93 KB

Overview

LangGraph-Dev-Navigator is a framework designed to enhance AI-assisted development of LangGraph applications by grounding AI agents with accurate, version-controlled documentation and code knowledge. The system combines Retrieval-Augmented Generation (RAG) with Knowledge Graph validation to reduce hallucinations and provide reliable, runnable code generation. The project treats AI development assistance as a scientific endeavor with quantifiable metrics for improvement.

User Preferences

Preferred communication style: Simple, everyday language.

System Architecture

Frontend Architecture

The system uses a multi-channel approach for AI assistance integration:

  • Multiple AI Assistant Support: Configured for Claude, Cursor, Windsurf, GitHub Copilot, and other popular AI coding assistants
  • Instruction Distribution: Automated setup scripts distribute LangGraph-specific rules and instructions to different AI assistants via their respective configuration files
  • Web Interface: Express.js-based API server with static file serving for frontend components

Backend Architecture

The core architecture follows a microservices pattern with specialized components:

Main Application Layer

  • Node.js/Express API Server: Handles HTTP requests, CORS, security middleware, and serves static content
  • Admin Authentication System: Password-protected access to admin dashboard and all /api/admin/* endpoints via Replit secrets
  • Python Tools Layer: Command-line utilities for web scraping, search, screenshot capture, and LLM API interactions
  • Configuration Management: Environment-based configuration with .env file support and VS Code debug configurations

Knowledge Processing Layer

The system implements a dual-pipeline approach for different types of knowledge:

RAG Pipeline

  • Web Crawling: Playwright-based scraping with HTML5 parsing for content extraction
  • Content Chunking: Smart markdown chunking with configurable chunk sizes (default 5000 characters)
  • Vector Embeddings: OpenAI API integration for generating semantic embeddings
  • Semantic Search: Vector similarity search with optional hybrid search combining keyword and vector approaches

Knowledge Graph Pipeline

  • Code Analysis: Python AST parsing to extract classes, methods, functions, and imports
  • Graph Database Storage: Neo4j for storing code structure and relationships
  • Hallucination Detection: Validation of AI-generated code against actual library structure

Data Storage Strategy

The architecture uses a polyglot persistence approach:

  • PostgreSQL Database: Persistent storage for waitlist signups, survey responses, and user data with automatic reconnection handling
  • Vector Database: Supabase (PostgreSQL with pgvector) for RAG content and embeddings
  • Graph Database: Neo4j for code structure and validation
  • In-Memory Storage: JavaScript Map objects for development/testing scenarios

Security Architecture

  • Admin Authentication: Password-based protection for admin dashboard and API endpoints
    • Password stored securely in Replit environment variables (ADMIN_PASSWORD)
    • Authentication required for all /api/admin/* routes
    • Frontend login overlay prevents unauthorized access to admin dashboard
    • CSV export functionality protected with authentication
  • API Protection: All admin endpoints protected by middleware authentication
  • Database Security: Robust error handling and automatic reconnection for database operations

External Dependencies

Cloud Services

  • Supabase: PostgreSQL-based vector database with pgvector extension for semantic search
  • Neo4j: Graph database for code structure analysis and hallucination detection
  • OpenAI API: Embedding generation and LLM interactions
  • Google Generative AI: Alternative LLM provider support
  • Anthropic: Claude API integration

Development Tools

  • Playwright: Headless browser automation for web scraping and screenshot capture
  • DuckDuckGo Search: Privacy-focused search engine integration
  • UV Package Manager: Python dependency management and virtual environment handling

Framework Dependencies

  • LangGraph Submodule: Git submodule pointing to official LangGraph repository for ground truth documentation
  • MCP Server Integration: Model-Context Protocol server (mcp-crawl4ai-rag) as a separate submodule providing RAG and graph capabilities
  • Express.js: Web server framework with security middleware (Helmet, CORS)

Testing and Quality Assurance

  • PyTest: Python testing framework with async support
  • VS Code Integration: Debug configurations and task automation
  • JSON Schema Validation: Request validation using jsonschema library

The system follows a "grounded AI" approach where all code generation is validated against actual, version-controlled source code rather than potentially outdated training data, ensuring higher reliability and accuracy in AI-generated solutions.