How to train QnA Maker using the portal if the correct answer is not among the top 3 ranked answers?

We have a knowledge base with near 1000 answers and a average of 15 alternate questions for each one of them. We are training the base both adding alternate questions as trying to indicate the correct answer using the portal.

We have struggling to improve the precision but without relevant success yet. An important problem is that base has a lot of similar answers between themselves, many times the difference among them is only an year (e.g. 2014 or 2018).

Some times when we use the GenerateAnswer API we need increase the parameter top a lot to receive the correct answer. For some questions we need assign the value 500 for top parameter, otherwise we get other answers with high scores, but still incorrect.

We thought that we could increase that results using the portal to train the knowledge base indicating the correct answer, but when we are trying it, we get just the first three top ranked answers, all incorrect.

*Note: We already migrated a test base to new approach (Language services/Question Answering) hoping that could be help us, but the problem stays, actually a bit worst.

How can we deal with this kind of situation? Can we change the top parameter value used by the portal? Are there another ways we can use to improve the rank?

Start with the QnA Maker portal training workflow and the GenerateAnswer API's top parameter described in the issue. Investigate whether the portal's ranked-answer limit is configurable and which documented ranking or knowledge-base practices address similar answers; done means providing a clear, supported way to improve ranking.

Written by the indexing model from the issue text.