Spaces:
Sleeping
Sleeping
title: Airline FAQ RAG Project | |
emoji: π | |
colorFrom: indigo | |
colorTo: pink | |
sdk: gradio | |
sdk_version: 5.29.1 | |
app_file: app.py | |
pinned: false | |
# Airline FAQ RAG Project | |
Welcome to the **Airline FAQ RAG Project**! This repository houses an innovative pet project exploring the creation of airline-related FAQ data using an AI agent and building a Retrieval-Augmented Generation (RAG) application to provide accurate and context-aware responses. Whether you're an AI enthusiast, a developer, or someone curious about natural language processing (NLP) and information retrieval, this project offers insights into data generation, embedding strategies, and RAG system optimization. | |
## π― Project Overview | |
This project has two primary components: | |
1. **FAQ Data Generation Agent**: An AI-driven agent that generates high-quality, airline-related FAQ data based on predefined topics, exploring the impact of prompt engineering on output quality. | |
2. **RAG Application**: A Retrieval-Augmented Generation system that leverages the generated FAQ data to answer queries, with experiments on different chunking strategies to optimize retrieval performance. | |
The goal is to build a foundation for an end-to-end conversational AI agent for airline customer support, starting with a robust RAG system. | |
## π Features | |
- **Dynamic FAQ Generation**: Automatically creates comprehensive airline FAQs covering topics like booking, baggage, cancellations, and in-flight services. | |
- **RAG Implementation**: Combines retrieval and generation to provide accurate, context-aware answers to user queries. | |
- **Chunking Experiments**: Evaluates multiple chunking strategies (e.g., full FAQ file vs. question-answer pair chunking) to optimize embedding and retrieval performance. | |
- **Prompt Engineering Insights**: Explores how different prompts affect the quality and relevance of generated FAQ data. | |
- **Scalable Design**: Lays the groundwork for extending the system into a fully autonomous airline support agent. | |
## π Repository Structure | |
- `data_generation/`: Scripts for generating airline FAQ data using an AI agent. | |
- `rag_application/`: Implementation of the RAG system, including embedding creation and retrieval logic. | |
- `data/`: Sample FAQ datasets and embeddings. | |
- `experiments/`: Notebooks and scripts comparing chunking strategies and their performance. | |
- `docs/`: Additional documentation and analysis of findings. | |
## π οΈ Getting Started | |
### Prerequisites | |
- Python 3.8+ | |
- Libraries: `transformers`, `faiss`, `numpy`, `pandas`, `langchain` (or your preferred RAG framework) | |
- Optional: GPU for faster embedding generation | |
### Installation | |
1. Clone the repository: | |
```bash | |
git clone https://github.com/your-username/airline-faq-rag.git | |
cd airline-faq-rag | |
``` | |
2. Install dependencies: | |
```bash | |
pip install -r requirements.txt | |
``` | |
3. Generate FAQ data: | |
```bash | |
python data_generation/generate_faq.py --topics "booking,baggage,cancellations" | |
``` | |
4. Run the RAG application: | |
```bash | |
python rag_application/run_rag.py | |
``` | |
## π‘ Key Components | |
### 1. FAQ Data Generation | |
The FAQ generation agent creates airline-related question-answer pairs based on user-specified topics (e.g., booking, baggage, cancellations). Key features: | |
- **Prompt Engineering**: Experimented with various prompts to control tone, detail, and accuracy. For example: | |
- Prompt 1: "Generate 10 FAQs about airline baggage policies in a formal tone." | |
- Prompt 2: "Create concise FAQs for airline cancellations with customer-friendly language." | |
- **Learnings**: | |
- Specific prompts with clear instructions (e.g., "include examples") yield more relevant and detailed FAQs. | |
- Iterative prompt refinement improves output consistency and reduces hallucination. | |
- Adding context (e.g., airline-specific policies) enhances realism but requires careful prompt design to avoid bias. | |
### 2. RAG Application | |
The RAG system retrieves relevant FAQ answers for user queries using embeddings and generates responses. Key experiments: | |
- **Embedding Strategies**: | |
- **Full FAQ File Embedding**: Treated the entire FAQ dataset as a single document, creating one embedding per file. | |
- **Question-Answer Pair Chunking**: Split FAQs into individual question-answer pairs, creating embeddings for each pair. | |
- **Custom Chunking**: Grouped related FAQs (e.g., all baggage-related questions) into chunks to balance context and granularity. | |
- **Performance Evaluation**: | |
- **Full FAQ Embedding**: Fast but less precise, as it struggles with fine-grained retrieval for specific questions. | |
- **Question-Answer Pair Chunking**: Best performance for precise queries (e.g., "What is the baggage allowance?"), with higher relevance scores in retrieval. | |
- **Custom Chunking**: Improved context for complex queries but increased retrieval latency. | |
- **Findings**: Question-answer pair chunking outperformed other methods in precision and recall, making it the preferred approach for this use case. | |
## π Key Learnings | |
- **Prompt Sensitivity**: Small changes in prompt wording significantly affect FAQ quality. For example, specifying "customer-friendly" vs. "formal" tones altered the output's readability and tone. | |
- **Chunking Matters**: Fine-grained chunking (question-answer pairs) improves retrieval accuracy but requires careful indexing to manage scale. | |
- **Embedding Trade-offs**: Dense embeddings (e.g., using BERT-based models) offer better semantic understanding but are computationally expensive compared to sparse methods. | |
- **Scalability Challenges**: Large FAQ datasets require efficient indexing (e.g., FAISS) to maintain low-latency retrieval. | |
## π Next Steps | |
To evolve this project into an end-to-end airline support agent: | |
1. **Context-Aware Generation**: Integrate user context (e.g., booking details) into the RAG pipeline for personalized responses. | |
2. **Multi-Turn Conversations**: Enhance the agent to handle follow-up questions and maintain conversation history. | |
3. **Real-Time Data Integration**: Incorporate live airline data (e.g., flight status APIs) to provide dynamic answers. | |
4. **Model Fine-Tuning**: Fine-tune the language model on airline-specific data to improve response accuracy and domain knowledge. | |
5. **Evaluation Metrics**: Implement automated evaluation (e.g., BLEU, ROUGE, or human-in-the-loop feedback) to quantify response quality. | |
6. **Deployment**: Package the RAG system as a web or mobile app for real-world testing. | |
## π License | |
This project is licensed under the MIT License. See `LICENSE` for details. | |
## π¬ Contact | |
For questions or feedback, reach out via GitHub Issues or email at [[email protected]]. | |
--- | |
βοΈ *Let's build the future of airline customer support together!* | |