This is a submission for the Google x Kaggle 5-Day Generative AI Event.
Our project demonstrates how to build an interactive airline assistant using Generative AI techniques including:
- RAG (Retrieval-Augmented Generation) for FAQ responses
- Embeddings-based similarity using
text-embedding-004
- Gemini 2.0 Flash for intent classification and smart field extraction
- Multi-turn interaction logic with running state
- A clean Gradio chat interface
The assistant interacts with users just like a real airline support agent:
- 🧠 Understands your input using AI-based classification
- 📘 Answers travel questions (e.g., "Can I bring a cat on the plane?")
- 🗃️ Differentiates between FAQ and personal information
- 📝 Collects key fields: name, flight number, issue, and email
- 🧠 Stores information across turns using a persistent user state
To run this notebook, you must add your Google Generative AI API key:
- Go to "Add-ons" > "Secrets" (in Kaggle)
- Add a new secret named:
GOOGLE_API_KEY
- Paste your Gemini API key as the value (from https://makersuite.google.com/app/apikey)
Component | Description |
---|---|
text-embedding-004 |
Generates vector embeddings for FAQ retrieval |
cosine similarity |
Finds the best-matching FAQ question |
Gemini Flash |
Classifies whether the user input is a question or a personal detail |
Field extraction |
AI-based slot-filling logic (e.g., extracting names from messy input) |
State tracking |
Stores each user field (name, flight, issue, email) during chat |
Gradio |
Provides an interactive chatbot interface for users |
This project simulates how a real-world AI-powered airline assistant might behave — blending FAQ search with intelligent form-filling, all within a lightweight notebook.
It showcases how we can combine LLMs, embeddings, and interface tools to build useful, responsive GenAI apps.
notebook.ipynb
– full interactive assistantassets
– demo pictures
Huge thanks to Kaggle, Google, and the GenAI team for organizing the event!