Introducing text and code embeddings

New API capability: the embeddings endpoint

OpenAI has added the embeddings endpoint to its API—a dedicated interface for generating high-quality dense vector embeddings from raw text and code snippets, decoupled from chat/completions and optimized for representational efficiency.

Core capabilities and use cases

• Cross-modal semantic encoding: A single model (e.g., text-embedding-ada-002, followed by text-embedding-3-small and text-embedding-3-large) supports unified embedding of natural language and programming languages;
• Plug-and-play downstream tasks: Enables zero-shot or few-shot semantic search, document clustering, unsupervised topic modeling, and classification—without model fine-tuning;
• Latency-optimized serving: Designed for low-latency, high-throughput vector computation, making it suitable for retrieval layers in RAG systems or real-time recommendation engines.

Technical positioning and ecosystem role

• Complementary—not substitutive—to LLM inference: embeddings serve as foundational infrastructure for RAG, agent memory indexing, and code intelligence tools (e.g., backend of GitHub Copilot);
• Interoperable with open-source stacks including Hugging Face Sentence Transformers, LlamaIndex, and LangChain, outputting standardized float32 vectors;
• Current default models are text-embedding-3-small (512-dim) and text-embedding-3-large (3072-dim), released late 2023 to supersede text-embedding-ada-002, with configurable output dimensions and optional L2 normalization.