I’d creating a jsonlines file based on a bunch of documentation that I’d like GPT3 to search and answer questions about.
How big should each of the documents be to get the best results out of GPT3? Should each text document be a single sentance? A paragraph? An entire topic?
Also, if I have simple tables in my documentation, has anyone found a good way to include that information into customer files? For example a 2x3 table of definitions:
foo A placeholder word used in coding often used to represent a key.
bar A placeholder word used in coding often used to represent a value.