LLM topic discovery in Virtual Agent

  • Release version: Yokohama
  • Updated January 30, 2025
  • 2 minutes to read
    • Summarize
    Summarized using AI
    This content was generated using new OpenAI-powered functionality. Results are provided on an as is basis and are not guaranteed to be accurate or complete.

    Summary of LLM topic discovery in Virtual Agent

    Large language models (LLMs) enhance Virtual Agent by enabling it to understand and process user statements in plain language. Using LLMs in Virtual Agent allows faster and easier setup than traditional Natural Language Understanding (NLU) models, improving conversation quality and flexibility without extensive training on intents or keywords.

    Show full answer Show less

    Key Features

    • LLM Topic Creation: When Now Assist is enabled, you can select LLM as the model type for a topic in Virtual Agent Designer. Unlike NLU, LLM topics do not require predefined models, intents, or keywords.
    • Topic Discovery: LLMs automatically match user utterances to the best topic using a clear, plain language description provided in the topic properties. If multiple topics match, users see a list of options.
    • Entity Extraction: LLMs extract necessary entities directly from user input without requiring manual entity mapping or input variable configuration as with NLU.
    • Topic Switching: LLMs enable seamless switching between conversation topics based on natural language requests without needing to cancel or restart sessions. This capability supports dynamic, multi-topic conversations and small talk integration.

    Key Outcomes

    • Reduced Setup Complexity: Eliminates the need for building and maintaining complex NLU models, intents, and keyword backups.
    • Faster Topic Deployment: Allows quicker creation, configuration, and deployment of conversational topics.
    • Improved User Experience: Supports natural, flexible dialogues that can handle topic changes mid-conversation and understand diverse user intents without rigid input requirements.

    Large language models (LLMs) enable Virtual Agent to process user statements in plain language. Conversations that use LLMs can outperform those conversations that use Natural Language Understanding (NLU), with easier setup.

    How LLMs work in Virtual Agent

    Watch this video to learn about LLM topics in Virtual Agent Designer.

    When you create a topic in Virtual Agent Designer, you can select LLM as the model type for your topic whenever Now Assist in Virtual Agent is turned on. Virtual Agent then uses LLM generative AI to discover topics that match the user's intent.

    Unlike NLU topics, LLMs don't require models, intents, or keywords to be linked to the topic. LLMs can discover topics and perform language-related tasks, such as text generation for case summaries and resolution notes, without months of training on NLU models. Overall, you can create, configure, and deploy LLM topics faster than working with NLU.

    With LLMs, Virtual Agent can do the following:

    • Perform topic discovery without needing a singular declared intent in a given topic.
    • Find intents without backup keywords as in NLU modeling.
    • Extract entity values without prior mapping as in NLU modeling.
    • Handle multiple conversation topic switches in a single conversation session.

    For more information about LLMs, see Large language models on the ServiceNow AI Platform®.

    Topic discovery

    With LLM topic discovery, topic authors no longer need to create and maintain complex NLU models and intents with backup keywords. The LLM does all of the heavy lifting for you. The only requirement is a robust, plain language topic description on the Properties tab in Virtual Agent Designer. The LLM uses this description to find the best topic match for the user utterance. If there are multiple potential matches, the user will see a list of topics to choose from.

    For example, if a user asks Virtual Agent to calculate a ride share fee, the LLM matches the user utterance with an existing topic that can calculate the ride share with a tip.

    Entity extraction

    With LLM topics, the LLM has all the information needed to determine if an utterance has the information to fulfill a request. Unlike NLU models, you don't need to associate entities with a user input node or add nodeless NLU entities as input variables to a topic. The LLM simply finds the entity that most closely fits the user intent.

    Topic switching

    Topic switching is faster and easier with LLMs compared to NLU topics. The LLM processes your requests to change intent made in natural language, and activates the appropriate topic.

    For example, if you start a conversation by asking for a mobile phone, you don't have to cancel the order first or restart the conversation. Instead, you can ask Virtual Agent to order a laptop instead. Virtual Agent immediately switches from the mobile phone topic to the laptop topic. Topic switching can be done during a query, but not within a catalog ordering flow.

    Another example is when a user might ask a casual question or engage in small talk. The question might be unrelated to the original request. You can create small talk topics or set up small talk filters to help the Virtual Agent match and launch the appropriate conversation for the switched topic. For more information, see Create a small talk topic and Configure small talk filters.

    Additional resources