Introducing ESA’s Earth Virtual Expert

EVE is a platform that provides information in a question-and-answer format, making it accessible for all levels of expertise. Users can ask questions such as: How has vegetation in southern Europe changed over the past decade? or how do satellites help monitor greenhouse gas emissions?

Earth observation data comes in many forms, including scientific papers, reports, technical documentation, models and repositories. Finding, comparing and understanding this information can be difficult, often requiring specialist expertise. This limits its accessibility to policymakers and the wider public.

To help bridge this gap, ESA Φ-lab partnered up with Pi School – an AI research and innovation company – as well as with two private companies: French-based AI company Mistral and Imperative Space, a multi-disciplinary company that works in innovation, communications and education for the space sector. Together they developed EVE, a large language model (LLM), which summarises, translates and presents huge amounts of data in accessible ways. Following on from this collaboration with ESA Φ-lab, Mistral has now signed a letter of intent with ESA.

EVE is adapted for Earth observation and Earth sciences using text-only sources including ESA, NASA and Copernicus websites, high-quality manuscripts and peer-reviewed research through a collaboration with Wiley, ensuring the information it provides is trustworthy.

Rather than relying only on the information it was trained on, EVE uses a Retrieval-Augmented Generation system that can search additional sources, including the latest articles published by Wiley, and incorporate relevant information into its responses. This allows EVE to answer questions using the most recent scientific evidence available.

One of the challenges of large language models is that they can sometimes generate information that is not supported by scientific evidence. To reduce this risk, EVE includes a safeguard to improve the reliability of its answers: its ‘LLM-as-a-judge’ feature reviews generated responses and helps identify content that should be revised or filtered before being presented to users.

The system is now live for all interested users. The model, code, curated training datasets and benchmarks are publicly available, with new releases and updates published on EVE’s Hugging Face and GitHub repositories. Looking ahead, EVE will continue to evolve through regular updates, with new releases planned in the coming months.

EVE has now entered its second development phase (as of September 2026): future updates will allow EVE to retrieve and analyse Earth observation data and carry out tasks with multiple steps.

The model will be able to automatically access and combine different types of information and will add the capability of searching and presenting satellite imagery in 2027. It will also include environmental data, weather reports and numerical models, providing users with a more complete view of the topics they explore.

As EVE grows and gains new capabilities, it makes Earth observation a more transparent and accessible field, helping turn scientific expertise into information that anyone can access and use.

EVE has been developed under the Foresight element of ESA's FutureEO programme.