Project Story: Training Wheels

What Inspired Us

Our team was inspired to create Training Wheels by the challenges many small businesses and creators face in implementing custom chatbots. We saw an opportunity to simplify chatbot creation by allowing users to train bots on their own content, all without needing to code. Our vision was to make embedding a chatbot as easy as embedding a YouTube video, empowering users of any background to add a smart, content-specific chatbot to their site.

What We Learned

Throughout the project, we deepened our understanding of AI and NLP systems, especially in the context of building accessible and user-friendly platforms. We explored how to use Retrieval-Augmented Generation (RAG) to enhance chatbot responses, employing LangChain and vector search for efficient information retrieval. Additionally, working as a team taught us the importance of balancing complexity with ease of use, ensuring that both our backend processes and user interface were streamlined for a seamless user experience.

How We Built the Project

We built Training Wheels with the following technology stack:

Frontend: React for the user interface, designed to be intuitive and accessible, guiding users through the process of uploading their resources and embedding the chatbot.
Backend: Python and Flask to handle file processing, training, and chatbot interactions.
Data Storage: PostgreSQL to store and organize user data and resources.
NLP and AI: Leveraging LangChain, Topic Modeling, and large language models (LLMs) for chatbot training, supported by RAG (Retrieval-Augmented-Generation) to enable the chatbot to answer questions based on specific user content.
Vector Search: Implemented for efficient content retrieval within the chatbot, allowing it to provide relevant and contextually accurate responses based on user queries.

In this setup, users upload documents such as PDFs or web pages, which are processed and indexed through vector search to build a content-rich knowledge base. LangChain and NLP models handle the chatbot’s training, and RAG enables the bot to pull answers directly from the uploaded content.

Challenges We Faced

Building Training Wheels brought several technical and design challenges.

File Parsing and Consistency: Supporting various document types (PDFs, Word documents, etc.) was complex, as each format required different parsing techniques to ensure the chatbot accurately interpreted content.
Balancing Performance and Ease of Use: We needed to ensure our backend was scalable to handle multiple requests while also keeping the front-end interface simple for users who may not have a technical background.
Embedding Flexibility: Designing an embedding system that worked smoothly with diverse websites and platforms required careful coding and testing to make sure the chatbot could be added to sites without issues.

Overcoming these challenges as a team taught us valuable lessons in collaboration, problem-solving, and delivering on a vision. We’re excited to see how Training Wheels can empower more people to create and share chatbots based on their own unique content!