This article is contributed. See the original author and article here.
The next version of QnA Maker advances several core capabilities like better relevance and precise answering, by introducing state-of-art deep learning technologies.
Summary of new features introduced:
- Deep learnt ranker with enhanced relevance of results across all supported languages.
- Precise phrase/short answer extraction from answer passages.
- Simplified resource management by reducing the number of resources deployed.
- E2E region support for Authoring + Prediction.
Detailed description of the new features is further down in this article. Learn how to migrate to the new QnA Maker managed (Preview) knowledge base here.
QnA Maker managed (Preview) Architecture.
- As per the architecture of QnA Maker managed (Preview), there will be only two resources: QnA Maker service for authoring and computation and Azure Cognitive Search for storage and L1 ranking. This has been done with an aim of simplifying the resource creation and management process. Now, customers need to manage only 2 resources instead of 5 different resources.
- QnA Maker managed (Preview) also allows the user to do language setting specific to Knowledge Base.
- Computation has been moved out of the user subscription, so there is no dependency on the customers for scaling and availability management. This allowed us to use SOTA deep learnt model for L2 ranker which enhances the L2 ranker horizontally across all the languages, so now we support all the 50+ languages with better and enhanced precision.
- QnA Maker service will be available in multiple regions to give customers’ the flexibility to keep their end-to-end service in one region.
- For inference logs and telemetry, the latest version will be using Azure Monitoring instead of App insights. To keep the experience seamless and easy to adopt all the APIs has been kept backward compatible. There is almost zero change in the management portal experience.
New features of QnA Maker managed (Preview)
This section talks about all the distinguishing features of QnA maker managed in detail.
Simplified Create Blade
Onboarding on QnA Maker managed (Preview), and resource creation has been kept quite simple. Now, you will see a checkbox with Managed, as shown below. As soon as you select the checkbox, the form will be updated with the required resources.
Machine Reading Comprehension based answer span detection feature is most beneficial for the scenarios where the customers have big passages present as answer in their Knowledge Base. Currently, they put good amount of manual efforts in curating small/precise answers and ingest them in the Knowledge base.
The new features give them flexibility to either choose the precise answer or the answer passage, customers can take this decision based on the confidence score of the precise short answer and answer passage. Here are some examples to show how short answers can be useful:
Deep Learnt ranker
The new L2 ranker is a deep learning-based transformer model, which improves the precision of the service for all the languages. For any user query, the new L2 ranker model understands the semantics of the user query better and gives better aligned results. This model is not language specific and is targeted to improve the overall precision of all languages horizontally. Here are some examples to analyze the difference between the results of current service and QnA Maker managed (Preview) service:
Current GA results
QnA Maker managed (Preview) results
Improvements in Preview
can someone ring me
I can tell you all about Wi-Fi calling, including the devices that support Wi-Fi calling and where you can get more information yourself. Feel free to ask me a question and I’ll do what I can to answer it
Yes, you can make and receive calls using Wi-Fi calling. Pretty nifty, right?
The new L2 ranker can understand the relevance between “ring me” and “make and receive calls” and is returning more relevant result unlike the current GA, which has returned a generic answer.
can’t connect to mobile data
You’ll be connected to Wi-Fi, so it’ll only use your minutes and text allowances.
If you don’t have mobile signal, it’s no problem. With Three inTouch Wi-Fi Calling, you can call and text whenever you’re on Wi-Fi in the UK, even without mobile signal.
The new L2 ranker is again able to understand the query better as its able to understand that mobile data is somewhere connected to mobile signals and hence giving better results based on the data present in the Knowledge Base than the current GA model.
E2E region support
With QnA Maker managed (Preview) our management service is no more limited to west-US region. We are offering end to end region support for:
- South Central US
- North Europe
- Australia East.
Other hero regions will be added when we go GA.
Knowledge Base specific language setting
Now, customers can create Knowledge bases with different language setting within a service. This feature is beneficial for users who have multi-language scenarios and need to power the service for more than one language. In this case, there will be a test index specific to every Knowledge Base, so that the customer can verify how the service is performing specific to every language.
You can configure this setting only with the first Knowledge base of the service, once set the user will not be allowed to update the setting.
Public preview of QnA Maker managed (Preview) will be free in all the regions (You only pay for the Azure Cognitive Search SKU). The standard pricing will be applicable when the service goes to GA by mid-2021.
Brought to you by Dr. Ware, Microsoft Office 365 Silver Partner, Charleston SC.