Skip to content

EVE: Earth Virtual Expert

EVE Avatar Earth Virtual Expert (EVE) An open science initiative funded and supported by the European Space Agency's Φ-lab, developed by Pi School in collaboration with Imperative Space and Mistral AI, aimed at advancing the use of Large Language Models (LLMs) for the Earth Observation (EO) and Earth Science (ES) communities.

Use EVE

EVE is open to everyone at eve-chat.chat. Sign up with the EVE identity provider and confirm your email address with the code it sends you. Sign-in, password reset, and optional two-factor authentication also happen at the identity provider.

Build with EVE

EVE exposes an OpenAI-compatible API at https://eve-chat.chat/api/v1. Authenticate with an API key that you create in the chat platform.

Components

This guide documents the components that power EVE. They are open source and designed to be reused in research on LLMs for Earth Observation.

  • Chat frontend: the web application of the EVE chat platform.
  • Chat backend: the API behind the chat platform.
  • Evalkit: an evaluation framework for LLMs on EO tasks and general domain benchmarks.
  • Data scraping: a framework to collect EO documents from trusted sources.
  • Data processing: a pipeline to clean, filter, chunk, and embed EO documents.

The project also publishes experiment repositories with hands-on notebooks. They are not production code and are not part of this guide:

  • EVE-RAG: build and benchmark Retrieval-Augmented Generation (RAG) systems.
  • EVE-Hallucination: experiments and analysis on the hallucination pipeline of EVE.

Project overview

EVE is a domain-specialized AI system that supports scientists, analysts, and decision-makers through natural-language interaction.

It assists with tasks such as:

  • EO and ES knowledge search and summarisation
  • Scientific Q&A and policy brief generation
  • Document understanding and retrieval
  • Technical and educational content creation
  • Search across private EO document collections through RAG

The project unites data curation, model development, evaluation, and deployment to deliver open, reusable tools for Europe's EO ecosystem.

Model

EVE-Instruct is trained on domain-curated corpora by fine-tuning Mistral Small 3.2 24B, through continued pretraining and instruction fine-tuning.

  • The model is optimized for factuality, reasoning, and integration with retrieval-augmented generation.
  • Training was carried out using EuroHPC MareNostrum, cloud environments, and Mistral infrastructure.
  • Evaluation targets both general domain and EO-specific benchmarks.

In the chat platform, EVE-Instruct is the default model. The Jülich Supercomputing Centre (JSC) serves it through its Blablador inference service, and backup providers answer when JSC is unavailable.

Licence

All EVE releases follow ESA and EU open-science principles, promoting transparency, reproducibility, and European AI sovereignty.
The project's outputs (models, datasets, pipelines, and documentation) are shared under permissive open-source licences (Apache 2.0 or CC-BY).

Contributors