Add more robust intermediate retrieval step

[ryderwishart]

If we add an intermediate agent, we can better handle edge cases, for example when the model answers based on training data rather than context documents, etc.