GenSearch settings and Persona/Prompt overview
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 Mpower Expert. Steps and recommendations for Expert may vary from those for other generative AI platforms or tools.
Why is a persona important?
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.
Since AI does not have a personality of its own, it is important for you to set one. The persona is the prompt engineering (directions) you give GenSearch to form what we consider its "personality." The persona tells Expert Generative Search who it is and how to interact with your customers. It makes GenSearch more engaging and personable, and helps ensure its responses are in line with your brand and business standards.
How to set a GenSearch persona in the UI
To set up an Expert GenSearch persona:
- Navigate to Dashboard > Site Administration > Generative Search.
- Fill out the fields according to the prompt / persona you created.
- Click Submit.
GenSearch settings
The settings help you control GenSearch to fit with your brand identity, and stop it from hallucinating or answering queries that are outside its intended scope.
- Name: The name will appear next to the search bar. For example, Expert GenSearch is called "ExpertAI," reflected in the following screenshot.
- Default state: Determines if generative search is automatically enabled upon user login (on) or requires a manual toggle-on (off)
- Assistant: Defines the GenSearch personality and its role. This informs GenSearch how to act. There are four components to the assistant persona:
- Your brand
- Tone and voice
- Identity: context about who GenSearch is
- GenSearch's objective / role
- User: A description of your audience / end user and their objectives
- Output instructions: Additional formatting instructions or guidance on how to answer queries. Without these guardrails, GenSearch will default to try to be helpful, which can lead to hallucinations and negative customer sentiment. The output instructions prevent GenSearch from going off script, and typically include additional formatting instructions or guidance on how to answer queries.
- Missing information response: This response will appear if a query does not yield any results.
- Target output word length: How many words at most will appear in the GenResponse. Longer responses take more time to generate, which can cause frustration.
- Relevance threshold: Relevance measures how closely the retrieved kernels match the user query on a scale of 0 (low) to 1 (high). The higher the relevance, the more likely the information is useful for generating an accurate response. Threshold allows you to control how strict the system is when retrieving information.
- Setting the threshold high (.8 - 1) reduces the chances of incorrect or irrelevant information but may miss useful content.
- Setting the threshold low (.3 - .5) increases coverage but may introduce less relevant or distracting information.
- Kernel limit: When a user enters a query in GenSearch, the Kernels that are most relevant to the query are retrieved. These Kernels are sent to the LLM to generate a response.
- Lower limit (1-3): The LLM receives only the most relevant information. This means responses are more concise, but may be missing additional useful context.
- Higher limit (6-10): More content is provided, but too much information can cause the LLM to exclude critical details.
We recommend you start with the default / out-of-box settings:
- Target Output Word Length: 250 characters
- Threshold: 0.7
- Kernel Limit: 5