QnAMaker giving incorrect results for same text with different casing or punctuation

Question

I have a (translated to Dutch) version of the chitchat knowledgebase that includes a Q&A pair for 'who are you?' with the following phrasings, amongst many others:

• "Wie ben jij"
• "Wie ben je"

Depending on the casing of the text I get different results:

• "wie ben je" -> Correct
• "wie ben je?" -> Correct
• "Wie ben je" -> Incorrect!

What's strange is that when I use the same casing as the phrasing it produces a completely different answer. Even more strange is that when I click inspect in the test window, the correct answer isn't even listed.

I've confirmed that:

• the phrase "Wie ben je" isn't used for any of the other Q&A pairs except this one
• I made sure to Save and train the knowledge base
• I'm using a Qna service dedicated for the Dutch language and I didn't mix languages

What should I do to fix the recognition?

Answer 1

@Sander Schutten I have tried to export your file and created a new KB and ran the same test. Here is the result.

I think both the scenarios provided the expected results. Could you please check if you can create a new QnA resource and create a new KB to test the same file? My best guess is the resource may be having issues or the app service tied to this QnA resource needs a restart. This can be done from Azure portal using the restart option in overview page of the resource.