Gemini Embedding

Gemini Embedding is Google’s text-embedding offering, served through the Gemini API. Like OpenAI’s endpoint, it turns text into vectors with one call, but it has a distinctive feature set: a task_type parameter that optimizes the embedding for its intended use (RETRIEVAL_DOCUMENT, RETRIEVAL_QUERY, SEMANTIC_SIMILARITY, CLASSIFICATION, and more), strong multilingual coverage, and Matryoshka representation, which lets you truncate vectors to a smaller output_dimensionality without retraining or re-ranking tricks.

For you, the interesting part is the task-aware design. Most embedding APIs use one representation for everything; Gemini lets you tell the model that a document is being stored versus a query being asked, which measurably improves retrieval because questions and answers are phrased differently. If you’re already on Google’s stack — Gemini for generation, Vertex AI for deployment — using the matching embedding model keeps billing, auth, and SDKs in one place, and the model has consistently ranked near the top of public retrieval benchmarks.

You’ll call it via the google-genai SDK with client.models.embed_content(...), passing your model name, text, and task_type. The workflow mirrors every other provider: batch-embed your corpus with the document task type, embed queries with the query task type, and compare with cosine similarity. Free-tier access through Google AI Studio makes it an easy model to experiment with before committing.