In the Tool Retrieval section of the course, it is not made clear how employing the object retriever makes any difference to the processing.
As an example, when we check tools[1].metadata, there is a tool selected that makes no sense (it is even said so in the video). But if you sent the same query (“Tell me about the eval dataset used in MetaGPT and SWE-Bench”) to the agent, that supperfluous tool is not called.
So there doesn’t appear to be any use of the insights of the object retriever on a future query, meaning it’s not clear how does it allow for a 2-step response that minimizes the number of tools being considered.
Any insights on this?