A Platform Administrator is adding a new topic to an agent. What is best practice for topic instructions? 70

The best practice for training a conversational agent on a new topic is to provide a comprehensive and varied list of example user questions, often called "utterances." Natural Language Understanding (NLU) models learn to recognize user intent by identifying patterns within this training data. A larger, more diverse set of examples helps the agent accurately map a wide range of user phrasing to the correct topic, improving its overall effectiveness and reducing misunderstandings. This example-based approach is more robust than providing abstract or generic instructions.

A. Write the instructions generically to allow the agent maximum flexibility in its reasoning.

Generic instructions lead to ambiguity and poor performance. The agent requires specific examples to accurately classify user intent, not abstract rules that can be misinterpreted.

C. Keep instructions as minimal as possible to reduce agent latency.

While performance is important, insufficient training data (minimal instructions/examples) results in an inaccurate and ineffective agent, which is a more critical failure than minor latency.

D. Add business-specific language rather than plain language.

The training data must reflect the language customers actually use (plain language), not internal business jargon, to ensure the agent can understand real-world user requests.

1. Official Vendor Documentation (Salesforce):

Source: Salesforce Help, "Einstein Bots Best Practices"

Reference: In the "Build a Great Intent Model" section, the documentation states, "To build a robust intent model, enter at least 20 different utterances for each intent... The more utterances you provide, the better the bot understands your customers." This directly supports providing an extended list of examples.

2. Official Vendor Documentation (Salesforce):

Source: Salesforce Help, "Gather Bot Conversation Data"

Reference: Under the subsection "Use the Right Amount of Data," it is recommended to "Aim for at least 50-100 utterances per intent for your initial bot version." This emphasizes the need for an "extended list" as mentioned in the correct answer.

3. University Courseware (Stanford):

Source: Stanford University, CS224N: Natural Language Processing with Deep Learning, Winter 2021 Lecture 15 Slides.

Reference: Slide 23 ("Data Augmentation for Text") discusses techniques for increasing the size and diversity of training datasets for text classification tasks (like intent recognition). This academic principle confirms that more high-quality, varied examples lead to better model performance, aligning with the practice of providing an extended list of user questions.