Stories Worth Sharing:5-Day Gen AI Intensive Course with Google
For our Kaggle Capstone, we challenged ourselves to build something real: a chatbot that could answer FAQs for our own company using GenAI tools. The goal? Create a working prototype, learn as much as possible, and explore how powerful and accessible today’s AI really is.
Why This Project?
Every business deals with repetitive questions. A static FAQ page helps, but it’s not always user-friendly. We wanted something better — a conversational assistant that could understand and respond naturally.
The Tech Behind the Bot
Our project aimed to create exactly that. Here’s a breakdown of what the system does:
- Knowledge Base: A list of 40 company-specific Q&A pairs.
- Embeddings: We used Google’s
embedding-001model to convert these into vector representations (mathematical pattern lists) — the backbone for finding meaning in text. - User Input: A chat interface (built with
ipywidgets) lets users ask questions. - Semantic Search (Meaning-based search): User queries are converted into mathematical representations (embeddings) and compared to the FAQ list using cosine similarity (a method to measure semantic similarity) to find the closest match, even if different words are used.
- RAG (Retrieval-Augmented Generation): A technique where the AI (Gemini 1.5 Flash) answers based on factual search results from our knowledge base rather than making things up. If a match is strong enough, it generates a natural-sounding response.
- Fallback: No good match? The bot politely admits it doesn’t have the info and suggests speaking to a human.
All of this runs inside a Kaggle Notebook, using the google.generativeai library, with help from ChatGPT for brainstorming and code cleanup.
The Code (A Glimpse)
The entire project lives within a Kaggle Notebook.
The Python code uses the google.generativeai library to interact with the Gemini API (interface) for both embedding (vector) creation and answer generation. Key parts include:
- A cosine_similarity function (using numpy) to compare the mathematical similarity of embeddings.
- A core find_best_match function to perform the vector search.
- Separate LLM prompting functions: one for generating answers from context (RAG) and another for the "low confidence" fallback response.
- The main smart_faq_bot_gemini function orchestrating the process: embedding the query, finding the match, deciding whether to use RAG or the fallback based on the similarity score.
- A simple interactive UI using
ipywidgetsfor demonstration within the notebook. But that is rather for the Demo purpose and has improvement potential.
Honest Reflections: Successes and Shortcomings
Now, for some real talk:
Can the code be improved? Absolutely. It’s functional but could be more robust, modular, and optimized. Error handling is present but could be more sophisticated. Using a proper vector store would be better than a simple list for larger datasets.
Is the model usage economically wise? Probably not in its current state. Each query involves embedding calls and a generation call. Optimization (e.g., caching, potentially cheaper models, prompt tuning) would be crucial for real-world deployment.
Does it solve the real-world problem? It demonstrates a proof-of-concept and works for the defined 40 FAQs. However, for Peak Pioneers’ actual use, it would need significant refinement, more comprehensive data, better evaluation, and integration capabilities. It’s a starting point, not a finished product.
Does this project show GenAI empowers ordinary people?
YES. 100% YES.
This is the most crucial takeaway for us. Without AI assistance (especially Gemini for coding and ChatGPT for ideas), we, as a team that isn't composed exclusively of deep AI/ML developers, could not easily have built something like this. It democratizes development in a way we haven't experienced before.
Open Questions?
Let us know if you desire any additional info about the project or the concept. We are happy to share more details.
Contact us