Radio-RAG

Retrieval-Augmented Generation (RAG) for radio regulations (e.g., ITU rules and spectrum management).
Index regulation PDFs with FAISS, retrieve the most relevant passages, and generate grounded answers with an LLM.

---

Table of Contents

• Overview
• Radio-RAG Custom GPT
• Features
• Quick Start
• Project Structure
• Usage
  • Examples
  • Common Arguments
• Experiments
• Hugging Face (ZeroGPU)
• Figures
• Troubleshooting
• Citation
• License

---

Overview

Radio-RAG implements a practical RAG pipeline tailored for telecom/spectrum regulations:

1. Ingest regulation PDFs and split them into chunks.
2. Embed those chunks and build a FAISS index.
3. Retrieve the most relevant passages for a user question.
4. Generate grounded answers with an LLM using the retrieved context.

---

Radio-RAG Custom GPT

Prefer a chat interface instead of running the code locally?

Use the Radio-RAG Custom GPT, a specialized assistant built on top of this RAG pipeline:

• ✅ Backed by the same PDF → chunks → FAISS retrieval used in this repo
• 📚 Tailored for ITU Radio Regulations & spectrum management use-cases
• 📎 Answers are grounded in the indexed documents, with references to the relevant provisions
• 🧪 Great for quick checks, exploration, and validating how RAG behaves before full deployment

👉 Open the Radio-RAG Custom GPT

---

Features

• 🔎 PDF → chunks → FAISS: simple, configurable ingestion pipeline
• 🧠 Model-agnostic: choose your embedding and LLM backends
• ⚙️ Tunable retrieval: chunk size, overlap, index type, top-K
• 🧪 Experiment ready: compare vanilla LLM vs. RAG-augmented runs

---

Quick Start

1) Install

git clone https://github.com/Zakaria010/Radio-RAG.git
cd Radio-RAG

# (optional) create a virtual environment
python -m venv .venv
source .venv/bin/activate    # Windows: .venv\Scripts\activate

pip install -r requirements.txt

2) Add your PDFs

Create the data/ folder (if it doesn’t exist) and put your regulation PDFs inside:

data/
├─ 2400594-RR-Vol 1-E-A5.pdf
├─ 2400594-RR-Vol 2-E-A5.pdf
├─ 2400594-RR-Vol 3-E-A5.pdf
├─ 2400594-RR-Vol 4-E-A5.pdf
└─ your_other_regulation_book.pdf

---

Project Structure

Radio-RAG/
├─ data/                 # Put regulation PDFs here
├─ tests/                # Evaluation / experiment scripts
├─ utils/                # Helpers (parsing, chunking, indexing, retrieval)
├─ local_rag.py          # CLI entry-point
├─ requirements.txt
├─ LICENSE
└─ README.md

---

Usage

Run the built-in help to see the exact flags supported by your current version.

python local_rag.py --help

Example

**A) Ask a question **

python local_rag.py \
  --pdf_folder ./data \

Then the app will run and you can ask your question

Common Arguments

• --pdf_folder (str, default: ./data) — directory of PDFs
• --chunk_size (int) — chunk length used for text splitting
• --overlap (int) — overlap between adjacent chunks
• --index_type (str) — FAISS index (flatl2, hnsw, ivfflat, ivfpq, …)
• --model_name (str) — LLM ID/name
• --top_k (int) — number of retrieved chunks

---

Experiments

Evaluation utilities live in tests/. A typical pattern:

python -m tests.evaluation  --chunk_size 400 --overlap 50 --index_type innerproduct --top_k 5 --model_name "deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B" --test "splits/eval.json"

This evaluation.py script include a switch like --norag to compare vanilla LLM vs RAG.

---

Hugging Face (ZeroGPU)

Prefer a hosted demo? Try the app on Hugging Face Spaces (ZeroGPU spins up on demand):

---

Figures

Figure 5 — Vanilla vs. RAG (Qualitative)

---

Troubleshooting

• No/irrelevant answers → Confirm PDFs parse correctly; try larger --top_k; adjust --chunk_size / --overlap
• Index performance → Start with flatl2 (baseline) or hnsw (fast). IVF variants can help at larger scale
• Changed models/params → Rebuild the index to avoid stale vectors

---

Citation

If you use this repository, please cite the paper:

• Paper: arXiv

@misc{kassimi2025retrieval,
      title={Retrieval-Augmented Generation for Reliable Interpretation of Radio Regulations}, 
      author={Zakaria El Kassimi and Fares Fourati and Mohamed-Slim Alouini},
      year={2025},
      eprint={2509.09651},
      archivePrefix={arXiv},
      primaryClass={cs.IR},
      url={https://arxiv.org/abs/2509.09651}, 
}