Adra-AI

Adra-AI is a multi-agent AI coding platform that generates new projects, edits existing repositories, and answers codebase questions through autonomous agent workflows. It leverages repository-aware RAG using Qdrant vector search, code-aware chunking, file hashing, and incremental indexing for context-driven code generation.

Problem Statement

Existing AI coding tools lack deep understanding of repository structure and context. They generate code without awareness of the codebase's architecture, leading to inconsistent and often incorrect suggestions.

Architecture Overview

The system uses a multi-agent architecture orchestrated by LangGraph, with specialized agents for code generation, editing, and Q&A. A RAG pipeline powered by Qdrant provides repository-aware context through code-aware chunking and incremental indexing. The FastAPI backend handles agent orchestration while the Next.js frontend provides real-time updates via WebSockets.

Technical Challenges Solved

• Designing code-aware chunking strategies that preserve semantic meaning across file boundaries
• Implementing incremental indexing with file hashing to avoid re-processing unchanged files
• Orchestrating multiple AI agents with LangGraph for complex multi-step code generation workflows
• Building real-time WebSocket communication for streaming agent responses to the frontend

Key Learnings

• RAG systems for code require fundamentally different chunking strategies than document RAG
• Incremental indexing dramatically reduces processing time for large repositories
• Multi-agent architectures benefit from clear separation of concerns and well-defined handoff protocols
• Real-time streaming of AI responses significantly improves user experience