Fetching and inspecting context generated by Assistants API implementation

API
1

The Assistants API, by design, abstracts away a lot of details of the underlying RAG.

When debugging an undesired behavior of an assistant, it would be helpful to inspect the exact prompt used by the assistant to generate the answer.

For instance, if I am doing a fact lookup in a document, but the returned answer is missing this fact, it would be helpful to know if the fact was simply not included by the underlying RAG implementation, or if there is an issue somewhere else in my prompt engineering.

Is this currently supported by the Assistants API? I did not see anything in the API documentation.

1 Like
2

Related questions:

Unfortunately, the best solution contained here is prompting the Assistant to regurgitate it’s own context, which is inefficient and inelegant.

1 Like