As with nearly all aspects of the work we did at FloodLAMP during the COVID pandemic, the recent advances in AI invite a complete reimagining of this type of hands-on technical training program. Even incorporating *just* a custom chatbot, implemented with a current GPT-4 level model, is game changing. This gives the trainee real-time, interactive support that is immensely helpful in mastering the material that is both procedurally intensive and involves intimidating scientific terminology. Extending from the chatbot, we envision a comprehensive multimodal system that intelligently delivers tailored content and guidance throughout the training process, and then continuing to support throughout all operations. For the lab setting, a hands-free system with voice-activated input and audio-visual output is ideal, and achievable with current technology.
We've implemented proof-of-concept custom chatbots on several platforms. This initial effort utilizes a subset of the testing and protocol documents we have available. Notably, with these basic, unoptimized RAG ("retrieval augmented generation") tools, there is degradation in answer accuracy when adding the course video transcripts to the baseline test training protocol document. This may be due to dilution of the most relevant chunks in retrieval. For more information, refer to
this link for details on this initial manual performance evaluation.
The POC custom chatbots developed and hosted using Streamlit which is an open-source Python library for data science and machine learning web apps. We invite you to test them.
FloodLAMP Test Training Custom Chatbot - LangChain using GPT-4 (v1.1)FloodLAMP Test Training Custom Chatbot - OpenAI Assistants using GPT-4 (v1.2)