I’m building a document search app using the OpenAI API for embeddings and answers, with the original files in object storage and the embeddings in our own vector database.

The upload flow is straightforward. The deletion flow is where I’m less confident: removing the original file does not automatically remove its chunks, cached answers, or references in existing conversations.

What should count as deleted?

I’m trying to define a document deletion workflow that covers:

• The original upload and any extracted text.
• Chunks and embeddings in our retrieval index.
• Cached responses that quote the document.
• Copies in application logs, backups, or provider-side storage, where applicable.

How are people separating “no longer available for retrieval” from “removed from every retained copy”? I also need to understand which API data retention details depend on the endpoints and account settings we use.

What would a useful deletion test look like without treating one unsuccessful search as proof that everything is gone?