OpenAI text-embedding-3-small

אודות OpenAI

הוקם בשנת 2015 · San Francisco, California, USA

OpenAI was founded in December 2015 by Sam Altman, Elon Musk, Greg Brockman, Ilya Sutskever, Wojciech Zaremba and John Schulman and restructured to a capped-profit company in 2019. The embedding line started with text-embedding-ada-001 (2021), was unified into ada-002 (December 2022) and replaced in January 2024 by the v3 generation. text-embedding-3-small is the smaller, cheaper sibling of text-embedding-3-large: it beats ada-002 by ~5 points on MTEB at one fifth the price ($0.00002 per 1k tokens). Together with the large variant it was the first OpenAI embedding model to support Matryoshka-style dimension shortening, allowing callers to choose any vector size between 256 and 1,536 without retraining or quality drop.

ארכיטקטורה

Transformer bi-encoder with Matryoshka representation learning

OpenAI text-embedding-3-small is the entry-level embedding model in the v3 generation. It produces 1,536-dimensional vectors by default and supports the same Matryoshka shortening as the large variant, so callers can request any dimension between 256 and 1,536 in the API and OpenAI will truncate-and-renormalise on the fly. Maximum input length is 8,191 tokens. Training used a contrastive retrieval objective on a curated multilingual web corpus including search-query/document pairs. The model scores around 62.3% MTEB and 44.0% MIRACL multilingual retrieval, a clear step up from 61.0% / 31.4% for ada-002, while being roughly five times cheaper and faster. It is positioned as the default embedding model for cost-sensitive production RAG and large-scale search. Like all OpenAI models the system is closed-source and hosted only.

פרמטרים

Undisclosed (smaller than text-embedding-3-large)

הקשר

8,191 טוקנים

יכולות

• 1,536-dim vectors with Matryoshka truncation to any 256-1,536 size
• 8,191-token context window for long-document embedding
• Multilingual coverage across 100+ languages
• $0.00002 per 1k tokens (~5x cheaper than text-embedding-3-large)
• Drop-in upgrade from ada-002 via the same /embeddings endpoint
• L2-normalised vectors with cosine similarity
• Best for: cost-sensitive RAG, large-scale search, embeddings at high QPS

הדרכה ורישיון

Not disclosed. OpenAI describes a 'curated multilingual web corpus' with search-query/document pairs.

רישיון: Proprietary commercial API. Generated embeddings may be stored and used commercially under the OpenAI Usage Policy.

בדיקות בטיחות: Embeddings are not subject to content-filter constraints; OpenAI publishes general bias and fairness statements.