🧠 MEKA – Multi-Agent Expert Knowledge Assistant

🚀 Production-grade multi-agent AI system for incident analysis, root cause detection, and knowledge extraction across internal engineering data

---

🌟 Key Highlights

• 🤖 Multi-Agent Reasoning System using LangGraph
• 🔍 Hybrid Retrieval (ChromaDB + Web Search)
• 🎯 Cross-Encoder Reranking for high precision
• 🧠 Context-Aware Retrieval using ParentDocumentRetriever
• 🔁 Validation Layer to reduce hallucinations
• ⚡ Async FastAPI Backend with real-time status tracking

---

🧠 Problem Statement

Engineering teams struggle with:

• Scattered incident data (Jira, Slack, Docs)
• Slow root cause analysis
• Lack of contextual reasoning

👉 MEKA solves this by:

• Aggregating knowledge sources
• Running multi-agent reasoning workflows
• Producing structured, validated answers

---

🏗️ System Architecture

User Query
    ↓
Planner Agent (Query Decomposition)
    ↓
Retriever Agent
    ├── Vector DB (ChromaDB)
    └── Web Search (DuckDuckGo)
    ↓
Reranker Agent (Cross-Encoder)
    ↓
Summarizer Agent (LLM - Ollama)
    ↓
Validator Agent (Hallucination Check)
    ↓
Final Answer + Reasoning Trace

---

🤖 Multi-Agent Workflow

Agent	Responsibility
Planner	Query decomposition + routing
Retriever	Hybrid search (vector + web)
Reranker	Context scoring using cross-encoder
Summarizer	Generate final answer
Validator	Ensure factual alignment

---

🔍 Retrieval Pipeline

• Vector DB: ChromaDB (local persistence)
• Embeddings: BAAI/bge-small-en-v1.5
• Retriever: ParentDocumentRetriever (deep context)
• Web Search: DuckDuckGo integration

👉 Combines internal + external knowledge

---

🎯 Ranking & Validation

• Cross-Encoder (ms-marco-MiniLM-L-6-v2)

  • Improves top-k precision significantly

• Validation Layer

  • Detects hallucination risk
  • Ensures answer grounded in context

---

⚙️ Tech Stack

Layer	Technology
Orchestration	LangGraph
Backend	FastAPI (Async)
Frontend	React + Vite
Vector DB	ChromaDB
LLM	Ollama (Llama3 / Mistral)
Embeddings	BGE-small
Search	DuckDuckGo

---

🚀 Backend API

POST /api/query

Submit query asynchronously

GET /api/status/{query_id}

Track agent execution + reasoning trace

GET /api/history/{user_id}

Retrieve past queries

---

📊 Design Decisions

• ChromaDB

  • Lightweight, no external infra required

• ParentDocumentRetriever

  • Preserves long-context understanding

• Cross-Encoder

  • Higher precision vs bi-encoder

• Local LLM (Ollama)

  • Privacy-focused
  • No external API dependency

---

⚖️ Trade-offs

• Higher latency due to reranking
• Local LLM depends on hardware
• Sequential scoring impacts speed

👉 Optimized for accuracy over speed

---

🧪 Testing Strategy

• Generated synthetic:

  • Jira incidents
  • Slack threads
  • Confluence docs

• Example queries:

Summarize all Kubernetes outages caused by misconfiguration
Identify incidents related to OAuth token leakage

---

📂 Project Structure

backend/
frontend/
agents/
scripts/

---

👤 Author

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

---

📈 Status

✅ Advanced multi-agent RAG system ⚡ Designed for enterprise knowledge systems

---

⭐ Star this repo if you found it useful!