Vector embeddings have quickly become a key building block in GeoAI, powering new forms of search, analysis, and data integration. But the landscape is young, fast-moving, and badly conflated: "geo embeddings" gets used for at least three fundamentally different things, and "open" gets used for at least five.

This talk is a practical field guide. It starts with a simple framework that separates the pieces that usually get lumped together: foundation models that produce embeddings (like Clay), pre-computed embedding representations you can use directly (like TESSERA's global temporal grid), and the systems used to store and query billions of vectors at scale (vector databases and indexes). Keeping these straight is the difference between evaluating a model, adopting a dataset, and choosing infrastructure.

From there we compare how the approaches actually differ in their inputs (single image, time series, multimodal), their training strategies, and their intended uses (similarity search, classification, change detection), and we look at how general-purpose vision models are being adapted for geospatial work, where that helps and where geo-specific pretraining still wins.

Then the FOSS4G question: what is genuinely open here? "Open" can mean open weights, open training data, an open license, open inference, or merely a public demo, and these rarely all come together. We'll give an honest read on what's truly open-source, what's only partially accessible, and where the real gaps remain, so you can tell marketing from reality.

You'll leave with a mental model for the geo-embedding ecosystem, concrete criteria for evaluating any model, representation, or system you come across, and a clearer sense of what you can actually build with what's freely available today. No prior embedding experience required, just curiosity about where GeoAI is heading.