Azure-docs: Is metadata Editorial:chitchat required always or only in mixed KBs?

[hansmbakker]

The documentation says:

To ensure the ranker understands that this is chit-chat, add the metadata key/value pair "Editorial: chit-chat"

I'm building a bot using Enterprise Bot Template. There, chitchat is added as a separate KB. In my understanding, rankers only have the scope of a KB, so I assume there it would not make a difference.

Could you please explain this behavior better, and in general, could you please document the ranking and scoring rules of QnA Maker better?

Is it possible to configure or tweak this somehow?

---

Document Details

⚠ Do not edit this section. It is required for docs.microsoft.com ➟ GitHub issue linking.

• ID: f9d0609c-6004-24ad-5446-88bdddd41b70
• Version Independent ID: 94baa464-93b0-efa8-1d6a-8d09c29dbd87
• Content: Adding chit-chat to a QnA Maker knowledge base - Azure Cognitive Services
• Content Source: articles/cognitive-services/QnAMaker/How-To/chit-chat-knowledge-base.md
• Service: cognitive-services
• Sub-service: qna-maker
• GitHub Login: @tulasim88
• Microsoft Alias: tulasim

[YASWANTHM-MSFT]

@hansmbakker, Thanks for the Comment. We are actively investigating and will get back to you soon.

[RohitMungi-MSFT]

@hansmbakker QnA Maker depends on language analyzers in Azure search for providing results. Special re-ranking features are available for En-* languages that enable better relevance.

While the Azure Search capabilities are on par for supported languages, QnA Maker has an additional ranker that sits above the Azure search results. In this ranker model, we use some special semantic and word-based features in en-*, that are not yet available for other languages. We do not make these features available, as they are part of the internal working of the QnA Maker's ranker.

You can learn more about language analyzers from Azure Search from here.

There are some best practices for Ranking/Scoring that are documented here that can help you train/create a better application.

[RohitMungi-MSFT]

@Hansmbakker If you do not have any more questions we will now proceed to close this thread. If there are further questions regarding this matter, please tag @RohitMungi-MSFT in your reply. We will gladly continue the discussion and we will reopen the issue.

[tulasim88]

@hansmbakker for your question about whether the metadata matters when you have a separate KB for chit-chat- the answer is Yes.
The ranking model chooses the best answer for a given user question, AND determines an appropriate confidence score. For chit-chat, the ranking model uses slightly different features and weights than for "normal" QnA pairs. So we recommend using this metadata

@diberry as FYI.

[hansmbakker]

Thank you for the explanation