How the reranking works in the real usecases?

[qaiwiz]

Hi, I noticed that you are using some random mock up and I wonder how practically one has to think of reranking in real usecases. Is it intended to generate a regression algo that takes a query and bunch of responses and score them accordingly (I mean based some supervised data)?

[noodlehaus]

The reranking will all depend on your use case, and what (meta) data you have attached to your documents. You could augment the score in terms of the relevance of the metadata (programmatically), or you could also pass the initial results through another LLM call to do the sorting/reranking. You could also use Cohere.

[qaiwiz]

Do you know any real world use-case anyone currently implementing?