Listen: Analysis of Gemini Embed Task-Based Dimensionality Deltas

When you generate vector embeddings for your text using Gemini, you have several optimization modes to choose from. Whether you are working on classification, clustering, semantic similarity, or document retrieval, each mode shapes your embeddings to fit the task at hand.

While each option gives you a slightly different result, some are more distinct than others. For instance, embeddings designed for semantic similarity are the most unique of the bunch. On the other hand, embeddings for tasks like search queries, document retrieval, and fact verification tend to look a lot more alike.

If you analyze the underlying data across all these different tasks, you can see how they compare. A close look reveals that the embeddings remain remarkably consistent. There is only a slight shift in values from one task to another, showing up as faint but perceptible lanes across the different dimensions.