AI Vector Search FAQ

FAQ topics

General questions

What is a vector?

A vector is a numerical representation of text, images, audio, or video which encodes the meaning of the data, not the underlying words or pixels.

How are vectors created?

A pretrained vector embedding model does an inference on an input (text, image, audio, or video) and generates a vector as the output. The vector is the values of the last hidden layer of the neural network after an inference for the input.

How many vector dimensions are supported?

Oracle AI Vector Search supports up to 65,535 dimensions.

What number formats for the vectors are supported.

AI Vector Search supports the INT8, Float32, and Float64 formats.

How big are vectors?

  • It depends based on the formula, for example, one formula is the number of dimensions times the size of the number formats.
  • Most vectors have between 128 and 4,096 dimensions.
  • Most vectors are Float32 (4 bytes) or smaller.
  • Most vectors tend to be between 1.5 KB and 8 KB in size.

Can you create vectors inside the database?

Yes, you can create vectors inside the database via the vector_embedding() SQL function.

You choose Sentence-Transformer embedding models from Hugging Face, and then you can securely upload them to the database.

Can you create vectors outside the database?

Yes, you can create vectors outside of the database using both commercial and open source models using either REST calls or local libraries.

The embedding models that have been tested include openai.com, cohere.com, Hugging Face Transformers, Sentence-Transformers, Transformers.js, and using the ONNX Runtime.

Do you need a GPU to create vectors?

You can use create vectors using either CPUs or GPUs.

Which LLMs does AI Vector Search work with?

AI Vector Search should be able to work with any LLM.

So far, Llama2, Gemini, and PaLM 2 as well as OpenAI’s ChatGPT and LLMs developed by Cohere, Vertex AI, and Mistral AI have been tested.

Which embedding models does AI Vector Search work with?

More than 90 embedding models from OpenAI, Cohere, and ONNX Runtime have been tested, as have the Transformer, Sentence Transformer, Transformer.js, Xenova, and FastEmbed models.

Does AI Vector Search support LangChain?

Oracle AI Vector Search supports the Oracle AI Vector Search provider for LangChain.