Welcome to our blog where we delve into the exciting journey we embarked on with MEKSI to revolutionize their educational chatbot for medical students. Let’s explore how we navigated through various technologies to achieve a more robust, engaging, and insightful educational tool.
MEKSI initially had a chatbot that was powered by GPT-2 and IBM Watson. These technologies provided a solid foundation but lacked in delivering a more nuanced and conversational experience that the client desired.
To elevate the chatbot’s capabilities, we explored fine-tuned versions of OpenAI’s GPT-3.5 and GPT-4, along with BERT and Llama. GPT-3.5 and GPT-4 were particularly impressive with their advanced language understanding and generation capabilities, making them ideal candidates for a conversational chatbot. BERT’s strength in understanding context and Llama’s versatility in handling diverse queries were also notable.
To enhance the chatbot’s ability to provide accurate and relevant information, we employed the Retrieval-Augmented Generation (RAG) technique. RAG combines the generative power of LLMs with external knowledge sources, significantly improving the chatbot’s response quality, especially in a complex field like medicine.
A significant challenge arose when the OpenAI API experienced downtime, leading to the chatbot becoming non-functional. This incident underscored the critical need for a less dependent and more resilient solution.
In response, we developed a multi-model approach. This involved setting up a system where the chatbot could seamlessly switch to an alternative model if the primary API became unavailable. This approach ensured continuous service and reliability.
Through careful selection and optimization of our AI models and cloud services, we achieved notable cost savings for MEKSI. This allowed for more budget to be allocated to other areas of development and improvement.
The revamped chatbot offered a significantly improved conversational experience. It also provided enhanced learning analytics, offering deeper insights into how students interacted with the content and helping MEKSI to tailor the educational experience more effectively.
We compared offerings from Amazon Comprehend and Google Vertex AI to determine the best fit for our needs. This analysis was crucial in ensuring we chose a platform that offered the right blend of features, scalability, and cost-effectiveness.
After thorough evaluation, the final solution was deployed on Google Cloud. This platform provided the scalability, reliability, and advanced AI capabilities we needed.
We decided to use a fine-tuned GPT-3.5 (or an equivalent model) as one of the core inference engines. This choice was driven by its advanced conversational abilities, making it ideal for a chatbot aimed at medical education.
In summary, our journey with MEKSI in developing an advanced LLM-based educational chatbot was filled with challenges, innovations, and significant achievements. The final product not only enhanced the chat experience but also provided valuable insights through learning analytics, all while achieving cost savings and ensuring reliability.
This blog post shares our journey and experience with MEKSI and the various technologies we explored. The outcomes of similar projects may vary based on specific requirements and implementations.