OpenAI Introduces Next-Gen Embedding Models and Additional API Enhancements

Mon Jan 29, 2024 - 5:02am GMT+0000

On Thursday, OpenAI, a leading AI research firm, unveiled a novel series of embedding models. These models transform text into numerical formats, facilitating a variety of machine learning applications. Additionally, OpenAI rolled out upgraded versions of its GPT-4 Turbo and moderation models, along with new API management tools, and announced a price reduction for its GPT-3.5 Turbo model.

Embeddings are numerical sequences that encapsulate the essence of content, be it text or code. These sequences enable machine learning models and algorithms to comprehend content relationships and execute tasks such as clustering or retrieval. Embeddings are crucial in powering features like knowledge retrieval in ChatGPT, the Assistants API, and numerous RAG (retrieval augmented generation) developer tools.

OpenAI’s latest embedding models, named text-embedding-3-small and text-embedding-3-large, boast enhanced performance and cost-effectiveness compared to the previous text-embedding-ada-002 model. These new models can generate embeddings with up to 3072 dimensions, capturing more semantic details and boosting task accuracy.

The company reports significant performance improvements with these models. On the MIRACL (multi-language retrieval) benchmark, the average score rose from 31.4% to 54.9%, and on the MTEB (English task) benchmark, it increased from 61.0% to 64.6%. Furthermore, the text-embedding-3-small model’s pricing has been slashed by five times relative to text-embedding-ada-002, making it more budget-friendly for developers.

OpenAI also updated its GPT-4 Turbo and GPT-3.5 Turbo models. These advanced multimodal models, adept in processing natural language or code, now feature improved instruction adherence, JSON mode, more consistent outputs, and parallel function capabilities. Additionally, a new 16k context variant of the GPT-3.5 Turbo model has been introduced, capable of handling longer inputs and outputs than the standard 4k version.

The company has also refined its text moderation model, which identifies potentially sensitive or unsafe text. The updated version supports additional languages and domains and provides justifications for its predictions.

For API management, developers now have enhanced capabilities. They can create multiple API keys with distinct permissions and scopes and monitor their usage and billing via the OpenAI Dashboard. OpenAI also plans to reduce the price of its GPT-3.5 Turbo model by 25%, further broadening its accessibility for application development.

OpenAI asserts that these enhancements align with its ongoing commitment to refine its models and services, making them more practical and economical for developers and clients. The firm encourages developers to contribute evaluations to enhance the model’s performance across various applications. OpenAI remains dedicated to continually introducing new models, features, and tools in the future.