💻 Debug Assistant AI

An end-to-end AI system that analyzes system errors and logs, retrieves similar past incidents using vector search, and generates root cause analysis with suggested fixes using large language models. This project combines Retrieval-Augmented Generation (RAG), FastAPI, PostgreSQL with pgvector, a Streamlit frontend, and full Dockerized deployment.

---

📌 Overview

Users can:

• Input system errors, logs, or failure messages
• Retrieve similar past incidents using semantic search
• Get AI-generated root cause analysis and solutions
• Understand how to fix and prevent issues
• Interact with the system via a simple UI

---

🧠 How It Works

1. Input — User enters an error or issue via Streamlit UI
2. Send — Frontend sends query to backend via POST /analyze
3. Embedding — Query is converted into a vector using a sentence transformer model
4. Retrieval — PostgreSQL (pgvector) finds similar past incidents
5. Augmentation — Retrieved context is combined with the user query
6. Generation — LLaMA 3 (via Ollama) generates root cause and solution
7. Return — Response is sent back and displayed in the UI

---

🏗️ Tech Stack

Layer	Technology
Backend	FastAPI, LangChain, psycopg2
AI Models	sentence-transformers (embeddings), LLaMA 3 (Ollama)
Frontend	Streamlit
Database	PostgreSQL + pgvector
DevOps	Docker, Docker Compose

---

📂 Project Structure

incident_ai/
│
├── backend/
│   ├── main.py
│   ├── rag.py
│   ├── db.py
│   ├── seed.py
│   └── requirements.txt
│
├── frontend/
│   ├── app.py
│   └── requirements.txt
│
├── docker-compose.yml
└── README.md

---

🚀 Getting Started

git clone https://github.com/your-username/incident_ai.git
cd incident_ai
docker-compose up -d --build

This starts:

• Frontend → http://localhost:8501
• Backend → http://localhost:8000
• PostgreSQL (pgvector) → running in container
• Ollama (LLaMA 3) → running in container

---

🔗 API Reference

POST /analyze

Request:

{
  "query": "TimeoutError while calling Ollama model"
}

Response:

{
  "response": "Root cause: The model is not fully loaded or request timed out...\nFix: Increase timeout or preload model...\nPrevention: Ensure model is ready before requests."
}

---

🤖 Model Details

Component	Value
Embedding Model	sentence-transformers/all-MiniLM-L6-v2
LLM	LLaMA 3 (via Ollama)
Input	Error logs / system messages
Output	Root cause + fix + prevention
Framework	LangChain
Strategy	RAG (Retrieval-Augmented Generation)

---

📊 RAG Pipeline

User Query
→ Embedding (vector)
→ pgvector similarity search
→ Retrieve top-k incidents
→ Inject into prompt
→ LLM generates solution

---

🎯 Features

• ✅ Semantic error search using embeddings
• ✅ pgvector-powered similarity retrieval
• ✅ LangChain-based LLM orchestration
• ✅ Root cause + fix + prevention generation
• ✅ Streamlit UI for interaction
• ✅ Dockerized microservices architecture
• ✅ Real-world debugging use case

---

💡 Example Use Case

Input:

Request timeout while calling LLaMA 3 via FastAPI

Output:

Root cause: Model not ready or timeout threshold too low
Fix: Increase timeout or preload model using ollama pull
Prevention: Implement health checks before inference calls

---

🚀 Future Improvements

• Add conversation memory (multi-turn debugging)
• Integrate LangGraph for multi-step reasoning
• Expand dataset with real-world incidents
• Add confidence scoring
• Deploy on AWS EC2 with CI/CD

---

🏁 Summary

This project demonstrates how to build a production-style AI system that combines:

• Retrieval (pgvector)
• Reasoning (LLM)
• Orchestration (LangChain)

to solve real-world engineering problems like debugging and incident analysis.