Telecom RAG: Future-Ready Telecom RAN Assistant

A Retrieval-Augmented Generation (RAG) application built for telecom Radio Access Network (RAN) tasks — 3GPP spec Q&A, root cause analysis, and anomaly detection — combining domain-specific knowledge retrieval with LLM-based reasoning.

Problem Statement

Telecom RANs are growing in complexity, making root cause analysis, anomaly detection, and optimization increasingly hard to manage manually. This project automates and accelerates those decisions using a RAG pipeline grounded in real telecom data — 3GPP standards (via TeleQnA), O-RAN network telemetry, and 5G core network fault data — while staying explainable, efficient, and faithful to its sources.

Use Cases

• 3GPP Spec Q&A — ask a telecom standards question, get an answer grounded in TeleQnA's question/explanation pairs, with sources cited.
• Root Cause Analysis — given an alarm or fault description, retrieve relevant context and explain likely causes.
• Anomaly Detection — flag abnormal network behavior using O-RAN performance telemetry.

Architecture

TeleQnA.txt ──▶ ingest.py ──▶ chunker.py ──▶ embeddings.py ──▶ FAISS index
                                                                     │
                                                                     ▼
                          user question ──▶ retriever.py ──▶ generator.py ──▶ answer + sources
                                                                     ▲
                                                              llm_setup.py

O-RAN telemetry (slice_mixed, slice_traffic) and the Zenodo fault/network-element data are loaded via ingest.py for direct pandas-based analysis (the anomaly-detection / RCA modules) rather than embedded into the vector store — that data is structured telemetry, not retrievable text.

Tech Stack

Component	Choice
Language	Python
Embedding model	BAAI/bge-large-en (sentence-transformers)
Vector store	FAISS
LLM	Mistral-7B-Instruct
Backend (planned)	FastAPI
Demo UI (planned)	Streamlit
Evaluation (planned)	RAGAS

Datasets

Dataset	Purpose	Source
TeleQnA	3GPP Q&A knowledge base	github.com/netop-team/TeleQnA (zip password: teleqnadataset)
O-RAN COMMAG	Network performance telemetry for anomaly detection	github.com/wineslab/colosseum-oran-commag-dataset
Zenodo fault data	Network element + performance metrics for RCA	zenodo.org/records/7003755

Note: TeleQnA's full dataset has 10,000 questions across 5 categories. A 1,827-question subset (Standards overview + Standards specifications only) matches a narrower "3GPP standards" scope if you want to filter to just that.

Project Structure

Telecom RAG/
├── data/
│   ├── raw/            # downloaded datasets -- NOT tracked in git, see setup below
│   └── processed/      # FAISS index + metadata -- tracked, built by embeddings.py
├── RAG-pipeline/
│   ├── ingest.py        # loads TeleQnA, O-RAN CSVs, Zenodo Excel files
│   ├── chunker.py       # splits oversized text (no-op for TeleQnA's short entries)
│   ├── embeddings.py    # builds the FAISS index, checkpointed/resumable
│   ├── retriever.py     # searches the FAISS index for a given query
│   ├── llm_setup.py     # loads the LLM (in progress)
│   └── generator.py     # combines retrieval + LLM into a full answer (in progress)
├── venv/                # NOT tracked in git
├── requirements.txt
├── .gitignore
└── README.md

Setup

git clone <your-repo-url>
cd "Telecom RAG"
python3 -m venv venv
source venv/bin/activate      # Windows: venv\Scripts\activate
pip install -r requirements.txt

Then download the datasets above into data/raw/:

• data/raw/TeleQnA.txt
• data/raw/slice_mixed/, data/raw/slice_traffic/
• data/raw/7003755/ (the two Zenodo Excel files)

Running the Pipeline

python RAG-pipeline/ingest.py        # confirms all three data sources load correctly
python RAG-pipeline/chunker.py       # verifies chunk sizes (mostly a no-op here)
python RAG-pipeline/embeddings.py    # builds the FAISS index (resumable if interrupted)
python RAG-pipeline/retriever.py     # test retrieval against the index

Target KPIs

Metric	Target
Mean Reciprocal Rank (MRR)	Above 75%
Top-k Accuracy	Above 85%
Accuracy	Above 80%
Recall	Above 85%
Faithfulness	Above 90%

Status

• Data ingestion (TeleQnA, O-RAN, Zenodo)
• Chunking
• Embeddings + FAISS index
• Retriever
• LLM setup
• Full generation pipeline (retrieval + LLM)
• FastAPI backend
• Streamlit demo UI
• RAGAS evaluation against KPI targets

Contributors

• Nancy
• Aryan Bansal