Hi Mandeep, I haven't had any similar problem yet but what I have understood till now how these work may be you can create a corpus of these csvs like iterate through each rows add "Question : {your Excel cell data} Answer : { Yes / No + para/s from pdf that you said}" append like this and write some text file use that to feed the model this way similarity concept will help and may be you can tweak with the code and get top n chunks instead of 1. & u can try Vicuna that works better than llama. Hope that helps