When a customer asks a question, they need information. For example: 

  • What kind of mortgages do you offer? 
  • How much is an annual subscription? 
  • Are you open on Saturdays? 

To provide the appropriate information to your customer, the knowledge workbench must determine what they really ask. 

After you finish compiling frequently asked questions and corresponding answers into knowledge articles into a knowledge base, the knowledge workbench learns to analyze these questions, along with accompanying answers, and then remembers what it learned during this training. 

When you create accurate, varied, and meaningful questions and answers, the knowledge workbench is more likely to give your customers the right answers. 

Use the correct wording

The knowledge workbench uses the questions, alternative questions, and answers added by you to create a natural language understanding model that detects what your customers are likely to ask. The ability to recognize the most significant words, concepts, or phrases within your question, alternative question, and answer ensures that it correlates them with the queries it receives. 

Make sure that you have a wide enough range of vocabulary that gives the knowledge workbench the information it requires to recognize a varied set of questions with similar answers. Ensure that the questions, alternative questions, and answers are not vague; they must be distinct from one another. 

To make this work, provide at least five complete question and answer articles. Ensure that these articles include keywords that a customer is likely to use. Also, use alternative questions to provide variations of words or phrases that provide the different ways of asking questions for the same FAQ answer. 


  • What are your hours?
  • When are you open on Saturdays?
  • How late do you close today?
  • What are your hours on Wednesdays?
  • How early do you open on Tuesday?

Notice that each question uses a different set of critical words: “hours,” “open,” “close,” “Saturday,” “today.” Compare these questions to ones that do not give the AI the information it needs.

Not preferred

  • When are you?
  • How much hours?
  • Test question

For optimal search results, Genesys recommends that you limit question and answer titles to 500 words or fewer.

Building blocks

Think of the critical words used in the good questions above (“open,” “hours,” and so on) as building blocks. 

The building blocks ultimately construct the relevant question. But first the system decodes, or parses, the input and looks for those individual elements to use as building blocks. Then, the building blocks help, with a fairly high level of confidence, to form the answers your customers need. 

Your training data also includes words that are irrelevant to the intent of your questions. The system discards these words during the training, just as they are discarded when, for example, your bot fields questions from customers. 

Note: Make sure that your training question and answer articles do not just contain irrelevant words.

Best practices

As you work on your questions, keep the following best practices in mind: 

  • Focus on the most common questions. Most customer inquiries revolve around a relatively small set of questions. For example, a knowledge base with a higher top three recalls and 100 question and answer articles that correspond to 80 percent of customer questions is more beneficial than 1,000 question and answer pairs that correspond to 90 percent of customer questions. 
  • Keep your questions short. 
  • Don’t cross-reference other question and answer articles within a single answer.