GenAI control options

If the generative AI does not understand queries, assigns them to the wrong intent or gives the wrong answers, optimization is required. The different ways of optimization, depending on the specific case, are explained in this article.

1. Unidentified query

2. Wrongly identified intent

3. Incomplete response

4. Wrong response is displayed

1) A query is not understood correctly

If the AI does not understand a query, it makes sense - as with editorial intents - to enter this specific query via the AI feedback tool.
The feedback entered takes effect immediately. 

A special case may arise here: The AI feedback tool displays the correct intent, i.e. the AI detection is correct, yet the "not understood message" is displayed in the chat widget. 

The reason is that no suitable information is stored for a reply. The solution in this case is to add suitable resources, on the basis of which the AI builds the appropriate response.

 

2) Wrongly identified intent

The AI recognizes a request but assigns it to the wrong intent. In this case, the AI Feedback Tool enables one to submit the necessary corrections. 
The correct intent is selected in the AI feedback tool and the correction takes effect immediately. 
If it is not possible to select the correct intent because there is not yet a suitable intent, there are two options: 

Add a new intent
Adding a new intent is a good idea if the chatbot receives enough real user inquiries.
If it can be seen that many user inquiries are coming in on a intent that is not yet covered by a specific intent, the AI may have already recognized this need and an AI suggestion exists. If a intent suggestion exists, it can be added with just a few clicks. 



Assign sources to the general knowledge base
If it can be seen that only a few user inquiries are received, connected to a specific intent, it is sufficient to assign the resource to the general knowledge base. This makes it possible to respond to inquiries without a specific intent to bee identified beforehand. 

It is possible to assign the added resources to a specific intent later on. 

This article describes when it makes sense to use general generative AI and the advantages of specific intents, particularly in the area of analysis and intent management. 

 

3) The response is incomplete or insufficiently given

In most cases, the reason for an inadequate or incomplete answer is the incomplete documentation in the sources themselves. Adding further, suitable resources (PDF, CSV, question/answer pair, document) often solves the problem. 


However, if the resources are suitable and the AI still does not generate a correct answer, the “Add further information” field, which is available in the individual intents in the hub, can help. 
Prompts (instructions) entered in this field instruct the AI to output certain information so that the generated answers are complete.

 
It is only possible to add further information in individual intents. It is not possible to influence the general generative AI by adding further information. The addition of further resources is necessary here.


If the content of the answers is correct, but the presentation (superfluous punctuation marks, other special characters, inappropriate formulations) is not correct, the AI communication guidelines allow conditional influence. 

These guidelines define the formulation of the created texts. AI-communication guide-lines take effect on every component using AI and are added only in consultation with the customer success team

 

4) Incorrect answer gets displayed

In some cases the response given by the AI is simply wrong. Often times the reason is a wrongfully detected intent or the incorrect usage of resources by the AI. 

If the detection of the correct intent is not given, part 2. of this article provides an answer. 

If the AI uses incorrect resources, the resources itself have to be optimized or an enlargement of the data-base is necessary. If a specific intent has yet to be created, it is useful to do so and to add the fitting resources to it. 

The intent works the same as a drawer in an archive, that enables the AI to directly access the correct place and search for the correct content to provide an answer. 

If the correct intent exists, part 3. of this article provides the fitting solution.