✈️ Aviation Engine Maintenance RAG

A Retrieval-Augmented Generation (RAG) system for aircraft engine maintenance manuals.
Built on FAA Aviation Maintenance Technician Handbooks, it enables accurate, context-grounded question answering and evaluation.

---

---

🔄 RAG Pipeline Flow

flowchart LR
    A["FAA PDFs in data/pdfs/"] --> B["Chunking\nbuild_corpus.py"]
    B --> C["Embeddings + FAISS Index\nembed_index.py"]
    D["User Query (q)"] --> E["Retrieval of Relevant Chunks"]
    C --> E
    E --> F["Answer Generation\nrag_pipeline.py"]

Loading

🚀 Demo

You can try the pipeline in one click:

• Run on Google Colab (recommended)
• Or open the notebook locally: notebooks/demo_rag.ipynb

A lightweight Gradio/Streamlit app version is planned — see placeholder badge above if you want to host on Hugging Face Spaces later.

---

✨ Features

• Ingests FAA handbooks in PDF form and chunks text
• Embeds with sentence-transformers/all-MiniLM-L6-v2
• Vector retrieval with FAISS
• Context-grounded answering via LLMs (Gemma-2, LLaMA-3.1, Flan-T5)
• Evaluation with EM, ROUGE, and RAGAS metrics
• Ready-to-run Colab notebook and Windows batch scripts

⚠️ Scope: Powerplant Only This project currently includes only the FAA Aviation Maintenance Technician Handbook – Powerplant (FAA-H-8083-32B) as the knowledge base. Non-Powerplant handbooks (General, Airframe) are not included to keep repository size manageable. Queries outside the Powerplant domain may not be adequately answered or may return generic/uninformed responses. You can refresh or update Powerplant content using scripts/fetch_pdfs.py or setup_and_fetch.bat — but those commands only affect Powerplant materials.

---

💡 Example

Question: What does the accessory gearbox drive?
Retrieved Context: “The accessory gearbox provides drive for the starter, fuel pump, oil pump, hydraulic pump, and generators.”
Answer (RAG): starter, fuel pump, oil pump, hydraulic pump, generator

---

📊 Results

We evaluated the pipeline on a gold Q&A set (data/qna_eval.jsonl) using RAGAS, Exact Match (EM), and ROUGE metrics.

Metric	Score	Notes
Exact Match (EM)	1.00	All 8 predictions exactly matched references
ROUGE-L	1.00	Perfect overlap with reference answers
Faithfulness (RAGAS)	–	Not computed yet
Relevance (RAGAS)	–	Not computed yet

📌 Based on 8 Q&A pairs. RAGAS scores can be added once evaluated with src/eval_ragas.py.

---

📚 Source Material

This project uses the official FAA Aviation Maintenance Technician Handbooks – Aircraft:

• FAA Handbooks & Manuals – Aircraft

---

📖 Overview

• Ingest PDF manuals/logs → text chunks
• Embed chunks with sentence-transformers
• Build FAISS vector index
• Retrieve + answer via LLM
• Evaluate with RAGAS, EM, ROUGE

---

🛠️ Tech Stack

• Embeddings: sentence-transformers/all-MiniLM-L6-v2
• Index: FAISS
• LLMs: google/gemma-2-2b-it, meta-llama/Llama-3.1-8B-Instruct, google/flan-t5-base
• Evaluation: RAGAS, EM, ROUGE
• Utilities: PyPDF, Pandas, NumPy

---

📂 Structure

aviation-engine-maintenance-rag/
├── data/
│   ├── pdfs/                # FAA handbook PDFs
│   └── qna_eval.jsonl       # gold Q&A set
├── notebooks/
│   └── demo_rag.ipynb       # Jupyter demo
├── scripts/
│   └── fetch_pdfs.py        # downloads FAA PDFs
├── src/
│   ├── build_corpus.py      # PDF → text chunks
│   ├── embed_index.py       # embeddings + FAISS index
│   ├── rag_pipeline.py      # retrieval + generation
│   ├── eval_ragas.py        # evaluation metrics
│   └── utils.py             # helper functions
├── setup_and_fetch.bat      # one-click setup (Windows)
├── run_end_to_end.bat       # one-click demo run
├── requirements.txt
├── README.md
└── LICENSE

---

📂 Knowledge Base: FAA PDFs

The PDFs placed in data/pdfs/ are the heart of this project – they form the source knowledge base.
All retrieval and answering is grounded in these FAA Aviation Maintenance Technician Handbooks.

When you run the pipeline, it processes the PDFs in three steps:

1. Chunking – src/build_corpus.py splits the PDFs into manageable text blocks.
2. Embedding & Indexing – src/embed_index.py converts chunks into dense embeddings and builds a FAISS index.
3. Retrieval + Answer Generation – src/rag_pipeline.py retrieves the most relevant chunks for a query and generates an answer.

This ensures that every answer is derived from the FAA handbooks rather than model hallucination.
The FAA PDFs are included in this repo under data/pdfs/ so the demo works out of the box.
(You can also refresh them anytime using scripts/fetch_pdfs.py or setup_and_fetch.bat.)

---

⚡ Quickstart

🪟 Quickstart Windows

# Clone repo
git clone https://github.com/suniltyagi/aviation-engine-maintenance-rag.git
cd aviation-engine-maintenance-rag

# Create virtual environment
python -m venv .venv
.venv\Scripts\activate

# Install dependencies
pip install -r requirements.txt

# (Optional) Refresh FAA PDFs
setup_and_fetch.bat

# Run the full pipeline (index → RAG → evaluation)
run_end_to_end.bat

🐧 Linux / macOS

# Clone repo
git clone https://github.com/suniltyagi/aviation-engine-maintenance-rag.git
cd aviation-engine-maintenance-rag

# Create virtual environment
python3 -m venv .venv
source .venv/bin/activate

# Install dependencies
pip install -r requirements.txt

# (Optional) Refresh FAA PDFs
python scripts/fetch_pdfs.py

# Run pipeline manually (index → RAG → evaluation)
python src/build_corpus.py
python src/embed_index.py
python src/rag_pipeline.py
python src/eval_ragas.py