Climate and earth observation data covers the physical planet: satellite imagery, weather records, ocean and atmosphere measurements, and long environmental time series. It is large, well-structured, and mostly public domain, because much of it comes from government agencies whose remit is to publish. For a RAG system, the setup that answers questions by retrieving relevant material and then generating a reply from it, the value is usually in the descriptive and tabular layers rather than the raw imagery: dataset documentation, station records, and the summary indicators built on top of the measurements.
Choosing within this category means being honest about what you are building. If you want a system that answers questions in words, sources like the World Resources Institute, which publish analysis and indicators, retrieve far better than a satellite archive full of pixels. If you genuinely need the underlying observations, providers like Copernicus, NASA Earthdata, and NOAA give you enormous, authoritative archives, but you will do real engineering to turn them into passages worth retrieving. Weather providers such as the Met Office sit in between, with both bulk data and readable forecasts and reports.
Two things to watch. First, formats: a lot of this data arrives in scientific containers like NetCDF or GeoTIFF that are not text at all, so plan a conversion step before you can embed it, the process of turning text into searchable numerical vectors. Second, access: several archives are free but need registration or an API key, and some throttle heavy downloads. The sources below run from ready-to-read environmental analysis through to the raw observation archives underneath it.