Google Satellite Embedding (AlphaEarth Foundations)
The Satellite Embedding dataset packs a year of multi-source Earth observation into a single vector per pixel. Google DeepMind's AlphaEarth Foundations model assimilates optical and thermal imagery from Sentinel-2 and Landsat, cloud penetrating radar, elevation and 3D surface measurements, climate variables and other geospatial signals, then encodes them as a unit-length, 64 dimensional embedding at 10 metre resolution. Coverage is global across terrestrial surfaces and shallow waters, with one image per year from 2017 onward and, at time of writing, annual layers through 2024. Google describes it as one of the largest datasets of its kind, on the order of a trillion plus embedding footprints per year.
You access it mainly through Google Earth Engine, where it is an image collection with 64 bands (A00 to A63) that you can filter, sample and export. Google has also published the embeddings to a public Google Cloud Storage bucket and mirrored copies exist on the AWS Registry of Open Data and Source Cooperative, so you can pull GeoTIFF tiles without an Earth Engine account if you prefer a bulk workflow.
For RAG and AI work the value is that the heavy feature engineering is already done: the vectors are analysis ready inputs for classification, regression, change detection and similarity search, and they slot straight into a vector index for geospatial retrieval or nearest-neighbour lookups over places. It is best for land cover, agriculture, habitat and environmental monitoring rather than text retrieval. Watch-outs: this is imagery-derived numeric data, not documents, so pair it with textual or tabular sources for grounded answers; the embedding space is Google's own and not directly comparable across model versions; and the CC BY 4.0 licence requires you to credit Google and Google DeepMind. If you need raw pixels instead of embeddings, the underlying Sentinel-2 and Landsat collections (also in the Earth Engine catalogue) are the usual alternatives.
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.
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.
Major TOM
An open, globally dense collection of embeddings and ML-ready imagery derived from Copernicus Sentinel data, published by ESA Phi-lab. The embedding expansions run to well over 100 billion vectors covering most of the Earth.