Abstract:

This document serves as a technical guide to the text embedding models available through the Google Gemini API, with a focus on the newest model, gemini-embedding-001. It details how these models convert text into numerical vector representations to power advanced Natural Language Processing (NLP) applications such as semantic search, classification, clustering, and particularly Retrieval-Augmented Generation (RAG). The guide outlines key features for optimizing performance, including the use of task_type to tailor embeddings for specific functions and the ability to control embedding size via output_dimensionality. This dimensionality control, enabled by Matryoshka Representation Learning (MRL), allows for more efficient storage and computation. The document provides code examples for generating embeddings, explains the necessity of normalizing smaller-dimension vectors, lists common use cases, and recommends vector database solutions for storage in production environments.

A Guide to Gemini API Text Embeddings

• Core Function: The Gemini API provides text embedding models, with gemini-embedding-001 being the newest, to convert text (words, phrases, sentences, and code) into numerical vectors. These embeddings are foundational for advanced AI tasks.
• Primary Use Case (RAG): Embeddings are crucial for building Retrieval-Augmented Generation (RAG) systems, where they are used to retrieve relevant information from a knowledge base to provide factual, context-rich responses from language models.
• Generating Embeddings: Embeddings are generated using the embedContent method. It can process a single string of text or a list of strings for batch processing, improving efficiency.
• Optimize with Task Types: To improve performance for a specific application, you can specify a task_type in your request. Supported types include SEMANTIC_SIMILARITY, CLASSIFICATION, RETRIEVAL_DOCUMENT (for documents being searched), and RETRIEVAL_QUERY (for the search query itself).
• Controllable Embedding Size: The gemini-embedding-001 model uses Matryoshka Representation Learning (MRL), which allows you to truncate the default 3072-dimension vector to smaller sizes (e.g., 768, 1536) using the output_dimensionality parameter. This reduces storage and computational costs with minimal loss in quality.
• Normalization for Quality: While the default 3072-dimension embedding is pre-normalized, smaller dimension embeddings must be manually normalized to ensure accurate semantic similarity calculations.
• Storage Solutions: For production use, storing embeddings in a vector database is recommended. The document mentions Google Cloud services (BigQuery, AlloyDB) and provides links to tutorials for third-party databases like ChromaDB, QDrant, Weaviate, and Pinecone.
• Model Availability: gemini-embedding-001 is the current, recommended model. Legacy models (embedding-001 and text-embedding-004) are being deprecated and should be avoided for new projects.