OpenAI text-embedding-3-large
FreeHigh retrieval nDCG@10
FreeFree tier
Inputs: text
About OpenAI text-embedding-3-large
OpenAI's text-embedding-3-large is a high-performance text embedding model that converts text into dense vector representations. With 3072 output dimensions (controllable via the dimensions parameter), it achieves state-of-the-art results on retrieval tasks, including a strong nDCG@10 score on the MTEB benchmark. It supports both semantic search and clustering, and is available through OpenAI's API for use in retrieval-augmented generation (RAG), classification, and other NLP pipelines.
Key Features
3072-dimension output (adjustable with dimensions parameter)
Top performance on MTEB retrieval benchmark (nDCG@10)
Supports semantic search, clustering, and classification
Simple API integration via OpenAI platform
Compatible with vector databases for RAG workflows
Pros & Cons
Pros
- Highest retrieval accuracy among OpenAI embedding models
- Flexible dimensionality reduction without retraining
- Easy to use with OpenAI SDK and API
- Consistent and reliable infrastructure from OpenAI
Cons
- Higher cost per token compared to text-embedding-3-small
- Requires API call latency for each embedding
- Not available for on-premises deployment
- Larger storage footprint due to high dimensions
Best For
Semantic search and information retrievalClustering and topic modelingDocument classification and sentiment analysisRetrieval-augmented generation (RAG) pipelinesDeduplication and similarity detection
FAQ
What is the dimension size of text-embedding-3-large?
The default output dimension is 3072, but it can be reduced by setting the dimensions parameter in the API request.
How does it compare to text-embedding-ada-002?
Text-embedding-3-large significantly outperforms ada-002 on retrieval metrics while offering more flexible dimension sizes.
Is text-embedding-3-large suitable for RAG?
Yes, it is one of the top choices for retrieval-augmented generation due to its high nDCG@10 score and compatibility with vector databases.