🧠 LangGraph Multi-Agent RAG System

🚀 Production-grade Retrieval-Augmented Generation (RAG) system powered by multi-agent orchestration, hybrid search, and real-time validation

---

🌟 Key Highlights

• 🤖 Multi-Agent Architecture using LangGraph for reasoning workflows
• 🔍 Hybrid Retrieval (Vector + BM25) for high recall
• 🎯 Cross-Encoder Reranking for precision improvement
• 🔁 Self-Healing Loop with validation + query refinement
• 🌐 Web Search Integration for real-time knowledge
• ⚡ Async FastAPI Backend with non-blocking execution
• 💻 Next.js Frontend with live reasoning trace

---

🧠 Problem Statement

Traditional RAG systems:

• Retrieve irrelevant chunks
• Hallucinate when context is weak
• Lack validation mechanisms

👉 This system solves:

• Retrieval accuracy
• Context validation
• Multi-step reasoning

---

🏗️ System Architecture

User Query
    ↓
Orchestrator Agent
    ↓
Retriever Agent (Hybrid Search: Vector + BM25)
    ↓
Reranker Agent (Cross-Encoder)
    ↓
Validator Agent (LLM-based evaluation)
    ├── Invalid → Query Refinement Loop
    └── Valid → Summarizer Agent
            ↓
        Final Answer

(Optional)
    ↓
Web Search Agent (DuckDuckGo)

---

🔍 RAG Pipeline

Retrieval Layer

• Vector DB: ChromaDB
• Embeddings: all-MiniLM-L6-v2
• Keyword Search: BM25
• Chunking: Recursive + hierarchical strategy

Ranking Layer

• Cross-Encoder: ms-marco-MiniLM-L-6-v2
• Improves semantic relevance significantly

Validation Layer

• LLM-based context validation
• Detects hallucination risk
• Triggers retry loop with refined query

---

🤖 Multi-Agent Workflow

Agent	Responsibility
Orchestrator	Query planning + routing
Retriever	Hybrid document retrieval
Reranker	Context scoring
Validator	Quality + hallucination check
Web Search	External knowledge retrieval
Summarizer	Final answer generation

---

⚙️ Tech Stack

Layer	Technology
Orchestration	LangGraph
Backend	FastAPI (Async)
Frontend	Next.js + Tailwind
Vector DB	ChromaDB
Embeddings	HuggingFace
Reranker	Cross-Encoder
Search	BM25 + DuckDuckGo

---

🚀 Backend Features

• POST /query → Execute RAG pipeline
• GET /status/{query_id} → Real-time progress
• GET /history/{user_id} → Query history

👉 Uses async background tasks for scalability

---

💻 Frontend Features

• Real-time Reasoning Trace
• Agent-level visibility
• Dynamic loading states
• Toggle for web search

---

📊 Performance & Optimizations

• Hybrid search improves recall
• Reranking improves precision
• Validation reduces hallucination
• Async execution avoids blocking

---

🧪 Example Query

Summarize all security audit findings related to OAuth token leakage and suggest mitigations

---

🔮 Future Improvements

• Add vector DB scaling (Pinecone / Weaviate)
• Implement semantic caching (Redis)
• Add evaluation metrics (RAGAS / TruLens)
• Streaming responses
• Multi-user session memory

---

📂 Project Structure

project/
├── backend/
│   ├── agents/
│   ├── scripts/
│   ├── api/
├── frontend/
│   ├── app/
│   ├── components/

---

👤 Author

Raj Kalash Tiwari GitHub: https://github.com/rjkalash

---

📈 Status

✅ Advanced RAG system with multi-agent reasoning ⚡ Designed for scalable AI applications

---

⭐ Star this repo if you found it useful!