ESA’s Earth Virtual Expert puts scientific retrieval inside the AI workflow

ESA introduced its Earth Virtual Expert, EVE, on 8 October. Developed with Pi School, Mistral and Imperative Space, the system is intended to help users question Earth-observation and Earth-science literature.

Its current scope is text. ESA describes scientific sources, including material from space agencies and peer-reviewed literature, combined with retrieval to bring additional documents into an answer. Image search and presentation are planned for 2027; users should not assume that a text answer is already an analysis of a satellite image.

For European environmental and resilience teams, the useful development is a route through a large scientific knowledge base. Our analysis focuses on the publicly documented retrieval method, rather than treating fluent answers as verified scientific conclusions.

What the public pipeline makes visible

The eve-rag repository describes a sequence: split documents into retrievable passages, embed and store them, retrieve relevant material and pass it to the language model. Its examples include hierarchical chunking to preserve document structure, a query rewriter for follow-up questions and a reranker to reorder retrieved passages.

That separation matters. A system can miss the right paper before generation begins, or retrieve the right paper but the wrong passage. Improving the generator alone cannot establish that the evidence presented with an answer supports its claim. The repository is a public implementation reference, not an independent inspection of the live service’s exact configuration.