Evaluate the Quality of a Response
When you preview a generated response, evaluate how well the response meets the goals of the prompt. Make sure that the response is factually accurate and doesn’t contain harmful content or bias.
Required Editions
| Available in: Lightning Experience |
| Available in: Enterprise, Performance, and Unlimited Editions with the Einstein for Platform, or Einstein or Agentforce for Sales or Service add-on, or Agentforce Foundations |
To assess a response’s effectiveness, it’s helpful to ask yourself questions about these topics.
| Topic | Questions |
|---|---|
| Relevance | Is the response relevant in this context? Does the response fit in with the conversation or content that surrounds it? |
| Goal Completion | Does the response fulfill the goals of the prompt? Does it address everything that the prompt requests? |
| Style and Tone | Is the style, voice, and tone of the response appropriate? Is the response’s vocabulary and punctuation correct? |
| Factual Accuracy | Does the response correctly incorporate the grounded data? Is the information in the response complete and accurate? Does the response contain redundant, excess, or erroneous information? |
| Consistency | How varied is the response? When you regenerate the response without changing the prompt template, how does the response change? How does the response change when you ground the prompt with different data? |
| Toxicity | Is the response safe? Does it avoid potentially harmful content, such as offensive, disrespectful, or abusive language? LLMs are trained on huge amounts of data, which puts the model at risk for producing toxic verbiage that leaks into your responses. |
| Bias | Does the response reflect fairness and inclusivity? Does it assume the gender identity of a person based on their name alone, sideline participants with disabilities, or displace assumptions about race or socioeconomic status? LLMs are trained on huge amounts of data, which puts the model at risk for producing biased verbiage that leaks into your responses. |
Did this article solve your issue?
Let us know so we can improve!
