The chat history can sometimes reword a query to result in a less useful view.
For example, the query images with a dolphin in them turned into show me all images with a label of 'dolphin'.
On the coco_gpt_demo dataset, there are not dolphin labels, so the first one would have likely resulted in an interesting sort_by_similarity view, but the second query would (and in my opinion should) always result in a filter labels for the class dolphin, which would be an empty view. That could be informative if you ask for exactly that query, but adding that level of specificity to the general prompt I gave it initially results in a view that's not very interesting.
The chat history can sometimes reword a query to result in a less useful view.
For example, the query
images with a dolphin in them
turned intoshow me all images with a label of 'dolphin'
.On the
coco_gpt_demo
dataset, there are notdolphin
labels, so the first one would have likely resulted in an interestingsort_by_similarity
view, but the second query would (and in my opinion should) always result in a filter labels for the classdolphin
, which would be an empty view. That could be informative if you ask for exactly that query, but adding that level of specificity to the general prompt I gave it initially results in a view that's not very interesting.