Where does GPT assistant store the embeddings

Documentation says that openai automatically creates the chunks and stores the embeddings.

Retrieval augments the Assistant with knowledge from outside its model, such as proprietary product information or documents provided by your users. Once a file is uploaded and passed to the Assistant, OpenAI will automatically chunk your documents, index and store the embeddings, and implement vector search to retrieve relevant content to answer user queries.

Where does the embeddings stored. Is it stored in some Azure service? Is there a way to access that embedding store?

If you watched the Microsoft keynote today, you may have seen mention of the Azure retrieval service that is used, so yes, it is on Azure services and is part of Microsoft’s vector and embedding service. I think it’s this Vector search - Azure AI Search | Microsoft Learn


There were screenshots on twitter leaking that they use qdrant, but I don’t think it’s been officially confirmed anywhere.

It’s not been confirmed anywhere because they are quite open about using Microsoft AI search, it’s not a secret.

No, the GPT Assistant does not operate by extracting text from a PDF, embedding it into vectors, and saving them into a vector database for subsequent searches. Instead, its approach to handling user queries about a PDF is different and does not involve creating or maintaining a persistent database of embeddings. Here’s how it generally works:

this is how GPT response. I dont know what it’s truth now

@jimkeecn I thinks there’s sth. missing in your answer, the most interesting part after “Here’s how it generally works:”

Perhaps this is deliberate, and is meant to suggest that OpenAI has not disclosed how they do it – and I did not get it :wink:

1 Like