Skip to main content
NICE CXone Expert
Expert Success Center

Generative Persona/Prompt setup

Create a persona to make GenSearch more personable and engaging while adhering to your brand standards and identity.

Personas differ based on the task that is being performed. These best practices are intended for a generative search experience that is designed to assist with Q&A tasks to help your customers self-serve.

This page covers information tailored specifically to CXone Expert. Steps and recommendations for Expert may vary from those for other generative AI platforms or tools. 

Why create a persona?

Language is fluid and nuanced, and interactions benefit from added context. GenSearch is like an actor, who understands the world but is not a subject matter expert on your brand. It needs direction from you to play its role successfully.

The persona is the prompt engineering (directions) you give GenSearch to form what we consider its "personality." Since AI does not have a personality of its own, it is important for you to set one. A persona makes GenSearch more engaging and personable, and helps ensure its responses are in line with brand and business standards.

The persona tells Expert Generative Search:

  • Who it is and what it is meant to do
  • Where to pull content from (your site, the Web, affiliated sites)
  • How to interact with your customers

We recommend creating a persona document to outline the personality traits you value. You can then use this document to construct persona prompts.

Before using GenSearch, you should carefully create and test your prompts and personas.

By default, GenSearch is equipped with a basic prompt / persona that is not intended to provide refined customer-specific answers. This base level of moderation is useful for "wild" queries like "How do I build a bomb?" but you should not expect this to prevent queries like "Who is the best movie princess?" 

Persona pieces

The 4 persona pieces help you control GenSearch to fit with your brand identity and stop it from hallucinating or answering questions that are outside its intended scope.

  1. Assistant: There are four components to the assistant persona:
    1. Your brand
    2. Tone and voice
    3. Identity: context about who GenSearch is
    4. GenSearch's objective / role
  2. User: A description of your audience and their objectives.
  3. Output constraints: This can include additional formatting instructions or guidance on how to answer questions. Without these constraints, GenSearch will default to try to be helpful, which can lead to hallucinations and negative customer sentiment.
  4. Target response length: The number of words you expect the response to be. The default response length is 50 words, and this is a guideline so response length will vary slightly.
  5. Kernel Limit: The number of Kernels (or text blocks) that will be sent to the foundation model for obtaining a Generative Search Response (also called a "completion"). Five is a good starting number. As you test, you can adjust the Kernel limit from 1-10.

Prompt / persona examples

Assistant

  1. You are a polite and knowledgeable subject matter expert trained to summarize articles found on a website.
  2. Never forget you are a helpful and legally responsible agent who wants to answer questions related to filing taxes in the United States.
  3. You are a polite and knowledgeable subject matter expert on Product X. Product X is the new brand name for Product A. Product A is a SAAS XXX product.
  4. You are a legal advisor working for Company Y. You provide positive and negative statements on questions related to an account holder.

User 

  1. I am a customer of Company Y that needs help resolving my problem.
  2. I am trying to understand and customize the Software Z platform.
  3. I am a brand new customer looking to learn and looking for support.

Output Instructions

  1. Be brief, use bullet points and step by step instructions when necessary respond with language a 6th grader will understand.
  2. For questions asking for instructions, provide a numbered list of steps to achieve the intended outcome. In your response, it is important to avoid mentioning which articles were summarized. Do not explain your answer or mention what it is based on. Only answer the question if it is related to the information provided. If the question can be answered, start your reply by answering the question immediately without any introduction.

Missing context

  1. I could not find that in the knowledge base.
  2. I could not find an answer to your question in the Help Center. I can only answer questions related to [specified topics].

Target output word length: 50

Score threshold: 0.75
The threshold is the relevancy distance between kernels in the vector database. The more closely related kernels are, the smaller their threshold will be.

Kernel Limit: 5

Persona prompt best practices

Once you have a persona document, create persona prompts so GenSearch knows how you want it to behave. Although you could paste the entire document if it is short enough, taking time to write thoughtful prompts will help you maximize how well GenSearch understands the personality you are assigning to it, and how well it interacts with your customers.

We recommend that you:

  • Use do language instead of do not.
  • Limit your persona prompt to 250 words. Longer prompts lead to slower processing time and attention issues, so you want to establish GenSearch's personality in the fewest words possible.
  • Expand GenSearch's understanding of your terminology and the kinds of site visitors you engage with.
    • Example: We are a storage facility and most of our site traffic is prospective customers looking for a unit to rent. When we say "unit" we mean a storage unit for personal belongings.
  • Provide context about the right language to use. GenSearch on its own has no context, so clarifying helps it construct useful responses.
    • Example, "issue" can have several meanings. You might tell GenSearch that "Issue" refers to a magazine publication that customers could have read, and is usually followed by a number (Issue 23) or a time-based clarifier (the winter issue, last week's issue).
  • Try to describe the expected subject matter expertise in 3-5 topics. If you have too many topics, your guidance about which to include or exclude may be ignored.
  • Your target response length should be about 50 words. Longer responses take more time to generate, which can leave your customers frustrated.
  • For Expert Kernels customers: You can choose the number of Kernels that are retrieved. Keep the number small to reduce token usage and ensure that GenSearch has enough "attention" to process all of the kernels it retrieves. Asking GenSearch to focus on the top 3 or 5 most relevant Kernels helps ensure it is looking at the information that is most needed by the user without being hampered by irrelevant or extraneous content.

 

 

  • Was this article helpful?