Researching a Humanitarian Disaster Situation Report Chatbot — Using GPT-4-Turbo and full-context prompting | by Matthew Harris | Nov, 2023

In this article, we explore OpenAI’s new GPT-4-Turbo model, which has an increased context window of 128k tokens. This allows us to pass in a full document corpus for information retrieval. Unlike other methods like Retrieval Augmented Generation (RAG), which require document embeddings and semantic search, this brute force approach is simple to implement and only possible with larger context windows. We applied this method to humanitarian disaster situation reports on ReliefWeb, compressed using Sparse Priming Representations (SPR), and found that GPT-4-Turbo is able to answer basic questions about recent disasters. However, this approach is expensive and the model’s response time is slow. It’s important to implement a validation framework to control for hallucination and information omission. Despite these challenges, GPT-4-Turbo offers great capabilities and is a step forward in the rapidly expanding Large Language Models (LLMs) toolkit.



Source link