Google’s official Embeddings documentation explains that the Gemini API can convert text into embeddings: numerical vector representations of content. These vectors are used to represent semantic information in text and can support use cases such as finding related content, retrieval-oriented workflows, classification, and clustering.
In retrieval, documents and user queries can both be embedded, but they play different roles. A document representation describes the content that may be retrieved, such as a passage, record, or knowledge snippet. A query representation describes the user’s question, search phrase, or information need. The documentation distinguishes these roles with task types such as RETRIEVAL_DOCUMENT for document-side embeddings and RETRIEVAL_QUERY for query-side embeddings.
This distinction should not be read as a promise of a fixed retrieval score or guaranteed result. Its purpose is to align the embedding generation step with the role the text plays in the application: document embeddings represent searchable content, while query embeddings represent the retrieval request. An application can then compare query vectors with document vectors according to its own retrieval design to find semantically related content.
Implementation details should follow the official documentation for supported models, parameters, and task types. Embeddings are vector representations that can support search, matching, and text organization, but the final retrieval behavior also depends on data preparation, indexing, similarity comparison, and application logic.
Break the goal into input, action, and output before adding review checkpoints.
Keep examples, failure conditions, and review criteria so the tutorial stays reusable.