Major TOM
Major TOM is a standard and a growing family of open Earth observation datasets curated by Phi-lab at the European Space Agency, in collaboration with CloudFerro and asterisk labs. Its core image datasets are among the largest ML-ready collections of Copernicus Sentinel-2 and Sentinel-1 data ever published, covering more than half of the planet's surface and nearly all dry land. On top of these, a set of embedding expansions turns the raw imagery into dense vector representations: examples include Core-S2L2A-MMEarth (around 39.7 billion vectors) and Core-AlphaEarth-Embeddings (around 71.3 billion vectors), alongside smaller SSL4EO, DINOv2 and SigLIP embedding sets. Together the expansions run to well over 100 billion vectors, all tied to consistent geographic grid cells.
Everything is hosted on the Hugging Face Hub under the Major-TOM organisation, stored as Parquet (with GeoParquet metadata) so you can stream or bulk-download what you need rather than pulling entire multi-terabyte archives. Each patch carries its geospatial footprint, so embeddings can be joined back to locations, imagery and other modalities.
For RAG and AI work, Major TOM is best suited to geospatial and Earth observation retrieval: similarity search over places, building region-aware indexes, or supplying pre-computed features to downstream models for classification, change detection and land-cover tasks. It is a features-and-imagery resource, not a text corpus, so it suits multimodal or geospatial pipelines rather than conventional document RAG.
Watch-outs: the different expansions use different backbone models, so vectors from separate sets are not directly comparable, and you should match the embedding model to your task. Licensing is CC-BY-SA-4.0 on the datasets, which requires attribution and share-alike on derivatives; the underlying Copernicus Sentinel imagery also carries its own open access terms worth checking. If you need broader or non-satellite geospatial context, OpenStreetMap is a natural companion already in this directory.
Related sources
Cohere Wikipedia 22-12 Embeddings (per-language)
Wikipedia encoded with the Cohere multilingual-22-12 embedding model, with embeddings (numeric vectors that capture meaning) computed on each article's title plus text. Published per language, covering many including Arabic, German, English, Spanish, French, Hindi, Italian, Japanese, Korean, Simple English, and Chinese. You can stream it rather than download it in full, which matters given the size.
Cohere Wikipedia Multilingual Embeddings (2023-11)
The November 2023 Wikipedia dump across 300+ languages, split into passages and turned into embeddings (numeric vectors that capture meaning) with the Cohere Embed V3 multilingual model. Close to 250 million paragraph embeddings in total. Drop it into a vector database for semantic search over all of Wikipedia, or use it as a knowledge source for multilingual RAG, including cross-lingual search where a query in one language returns relevant results from another.
Google Satellite Embedding (AlphaEarth Foundations)
Global annual satellite embeddings produced by Google and Google DeepMind's AlphaEarth Foundations model. Every 10 metre pixel holds a 64 dimensional vector summarising a year of multi-sensor Earth observation, available yearly from 2017.
LAION-400M with CLIP Embeddings
400 million image-text pairs filtered from Common Crawl using CLIP, a model that scores how well an image matches a caption, and released alongside their pre-computed CLIP embeddings and kNN indices for fast similarity search. The ready-made embeddings make it unusually easy to work with. Note that it ships image URLs, not the images themselves, so some links have decayed.