Telco Network Troubleshooting Chatbot (LLM + RAG Prototype)

A simple, fully local chatbot that demonstrates the "Talk to Your Network" concept using a small LLM, Retrieval-Augmented Generation (RAG), and synthetic telecom logs.

You can ask natural-language questions about simulated network issues (e.g., "Why is latency high on Core-Router01?" or "Show me all ISIS neighbor drops") and get grounded answers with root-cause suggestions and recommended actions.

This is a prototype / proof-of-concept — it runs entirely on your laptop, uses fake data, and is not meant for production networks.

Features

• Conversational troubleshooting interface (Streamlit)
• RAG over synthetic syslog-style logs + a small telco knowledge base
• Local LLM inference with Ollama (no cloud, no API keys)
• Fast local embeddings (Hugging Face sentence-transformers)
• Easy data regeneration for testing different scenarios

Demo

Quick Start

1. Install Ollama

Download and install from https://ollama.com

Pull a small model (recommended for speed on laptops):

ollama pull llama3.2:3b

(Other small models like phi3:mini or gemma2:2b also work well)

2. Set Up the Python Environment

git clone https://github.com/afiren/telco-troubleshooting-chatbot.git
cd telco-troubleshooting-chatbot
pip install -r requirements.txt

3. Generate Synthetic Logs

python data_generator.py

This creates data/simulated_logs.csv with realistic telecom-style alerts, KPIs, and actions.

4. Run the Chatbot

streamlit run app.py

Open your browser at http://localhost:8501 and start asking questions.

Example Questions to Try

• "Show me all critical alerts"
• "What caused the ISIS neighbor drops?"
• "Any logs with high packet loss?"
• "Recommend actions for interface flaps"
• "Why is latency high?"

Project Structure

.
├── app.py                  # Main Streamlit app + RAG chain
├── data_generator.py       # Generates synthetic logs
├── data/
│   ├── simulated_logs.csv  # Generated logs
│   └── telco_manual.txt    # Domain knowledge for RAG
├── requirements.txt
└── README.md

Limitations

• Fully synthetic data — no real network integration
• Small local models can be less accurate than giant cloud LLMs
• No persistence beyond the current session
• Prototype only — not hardened for production use

Future Ideas (Contributions Welcome)

• Add agentic workflows (e.g., auto-diagnostics)
• Connect to network simulators (Mininet, GNS3)
• Support real (anonymized) log ingestion
• Fine-tune a telco-specific small model

License

MIT License — feel free to fork, modify, and use.

---

Built as a personal project to explore LLM applications in telecom network operations.