OpenAI Embeddings API

OpenAI currently offers text-embedding-3-small and text-embedding-3-large as their latest generation models, both released in early 2024. The 3-small model is significantly cheaper at $0.02 per 1M tokens and produces 1536-dimensional vectors, while 3-large offers higher retrieval quality at $0.13 per 1M tokens with up to 3072 dimensions. For most RAG prototypes, text-embedding-3-small provides an excellent balance of cost and performance.

After generating embeddings via the API, you store the vectors in a vector database such as Pinecone, Weaviate, Qdrant, or pgvector (an extension for PostgreSQL). To perform semantic search, you embed the user's query and run a similarity search—typically using cosine similarity—against the stored vectors to find the closest matches. Most vector databases handle this computation natively.

The text-embedding-3-small and text-embedding-3-large models both accept up to 8,191 tokens of input text per request. Documents longer than this limit must be chunked into smaller segments before embedding. For RAG pipelines, chunking strategies typically target 500–1500 tokens to balance context with retrieval precision.

Yes, the text-embedding-3-small and text-embedding-3-large models are multilingual and support a wide range of languages natively. This is a significant improvement over the older text-embedding-ada-002 model, which was primarily English-focused. Cross-lingual semantic search—matching queries in one language to documents in another—works reasonably well out of the box.