text-embedding-3-small for business

text-embedding-3-small is OpenAI's efficient embedding model. It produces 1,536-dimension vectors from inputs of up to 8,192 tokens, and it is the lowest-cost option in the current OpenAI embedding line.

For many business knowledge assistants, it is a sensible starting point. It is cheap enough to embed large collections, fast enough for live search and small enough to keep storage costs down. If testing shows that it misses passages a stronger model would find, moving up to text-embedding-3-large or another provider is straightforward, as long as you plan for re-embedding.

We usually combine embeddings with keyword search, because some questions depend on exact words such as product codes or names. Hybrid search, plus a rerank step such as Cohere Rerank, often matters more than the choice of embedding model.

The quality of chunks matters too. Splitting documents at sensible boundaries, keeping titles with their sections and storing the source for every chunk all help the assistant find and cite the right text. See RAG knowledge assistants.

Cost is rarely the problem with embeddings, since even large collections are cheap to embed once. The real costs are storage, the time to re-embed when you change models and the quality of what the assistant finds. We keep a small set of test questions for every project and run it after any change to chunking, models or search settings, so problems show up before users see them.