Leveraging Large Language Models for Providing User Support and Access to Forecast Products

Prompt engineering is a new discipline that utilizes large language models (LLMs), like ChatGPT, to build applications and services. We explore two techniques:

1- Retrieval Augmented Generation (RAG) for providing help and documentation: For handling user queries, we have scraped our documentation and indexed it in a vector database. This context is used by the LLM to offer assistance and provide relevant information.

2 - Reasoning and Acting (ReAct) for access to forecast products: In this case, the LLM is tasked with selecting the most appropriate tool from a predefined set and determining the relevant parameters for the tool. One of these tools is an API that retrieves a point forecast at a given location for a specified date. By employing the LLM's capabilities, we can seamlessly transform a user's request like "Will it rain tomorrow in Reading?" into the corresponding API call.

Finally, we make use of ChatGPT's multi-lingual aspect to answer questions in the numerous languages spoken by ECMWF’s Member and Cooperating States.