Hey Michael, thanks for sharing the link. Super useful! As for the incoherent response quality, it would be good to break it down and see where it went wrong e.g., was it the vector retrieval step, or the summarization (aka generation) step of the retrieved chunks. From there, it would be easier to understand what didnt go well. In terms of the prompting, yeah indeed it is an important effort to embark on in order to maximize its abilities. I did not delve into that component for this project, but will certainly explore Llama 2's prompting intricacies in subsequent ones. Cheers!