Do retrievers assume all correct information can be found in one document?

[superchargez]

It is possible that one document may not have all the correct information regarding a particular query, for example a large paragraph that contains chunks of information that is related to query might get separated in various documents (due to limits of input words of transformer models). For example querying "mention three prophecies regarding red comet" the generator should list all three prophecies instead of giving out one.

(I have tried different queries, not this particular one, and retriever did not find all correct answers). So, I think this can be improved, but how though I have idea (except for manual labour of training/fine tuning, which is wrong way to do it, because there are always going to be queries which we would not have thought of previously).

Have a good day.

[bogdankostic]

Hi @superchargez! This problem seems to be an instance of MultihopQA, wehre you need to aggregate information from different documents. This topic is already on our roadmap and we plan to work on this from March 2022.

[superchargez]

Hi @superchargez! This problem seems to be an instance of MultihopQA, wehre you need to aggregate information from different documents. This topic is already on our roadmap and we plan to work on this from March 2022.

Much appreciated I hope a nice tutorial will be available for this as well soon.

[superchargez]

I'll mark it 'answered' when this gets done.