RAG Telecom Chatbot

Telecom_chat_bot streamlit-app-Telecom-RAG-System-Qwen3-27B on Groq.webm

A Retrieval-Augmented Generation (RAG) customer care chatbot for telecom support. It answers questions about mobile connectivity, billing, SIM issues, and roaming by retrieving relevant context from three knowledge sources and generating responses with Qwen3-27B via Groq.

Architecture

User question
     │
     ▼
Merged Retriever (top-k from each store)
  ├── ChromaDB · faq        (FAQ entries from CSV)
  ├── ChromaDB · tickets    (resolved support tickets from SQLite)
  └── ChromaDB · guides     (PDF guide chunks)
     │
     ▼
ChatPromptTemplate → Qwen3-27B (Groq) → Answer

Embedding model: sentence-transformers/all-MiniLM-L6-v2 (runs locally via HuggingFace)
LLM: qwen/qwen3-27b served by Groq

Project Structure

rag-telecom-chatbot/
├── app.py              # Streamlit web UI
├── main.py             # CLI entry point
├── rag_chain.py        # Builds the LangChain RAG chain
├── retriever.py        # Merges the three Chroma retrievers
├── ingest_faq.py       # Loads data/faq.csv → Chroma 'faq' collection
├── ingest_tickets.py   # Loads data/tickets.db → Chroma 'tickets' collection
├── ingest_pdf.py       # Loads data/telecom_guide.pdf → Chroma 'guides' collection
├── data/
│   ├── faq.csv             # FAQ question/answer pairs
│   ├── tickets.db          # SQLite database of resolved support tickets
│   ├── telecom_guide.pdf   # Telecom user guide (chunked at ingest)
├── chroma_store/       # Persisted Chroma vector database (created at ingest)
├── evaluate_retrieval.py   # Evaluate retrieval quality    
├── pyproject.toml
├── uv.lock
└── .env.example

Prerequisites

• Python 3.11+
• uv (recommended) or pip
• A Groq API key
• A HuggingFace token (for downloading the embedding model)

Setup

1. Clone and install dependencies

git clone <repo-url>
cd rag-telecom-chatbot
uv sync          # or: pip install -e .

2. Configure environment variables

cp .env.example .env

Edit .env and fill in your keys:

GROQ_API_KEY=your_groq_api_key_here
HF_TOKEN=your_huggingface_token_here

3. Ingest data into Chroma

Run the three ingestion scripts once to build the vector store:

python ingest_faq.py
python ingest_tickets.py
python ingest_pdf.py

Each script embeds the source data and persists it to chroma_store/. Re-run a script only when its source data changes.

Running the App

Streamlit web UI

streamlit run app.py

Opens at http://localhost:8501. The sidebar has one-click sample questions and a button to clear the conversation history.

CLI

python main.py

Interactive prompt — type a question and press Enter. Type quit to exit.

Data Sources

Collection	Source file	Granularity
faq	data/faq.csv	1 document per FAQ row
tickets	data/tickets.db	1 document per resolved ticket
guides	data/telecom_guide.pdf	Chunks of 600 chars with 100-char overlap

The retriever fetches the top 3 results from each collection (9 context documents total) for every query.