Migrating NLU/keyword Virtual Agent topics to LLM topics

  • Release version: Xanadu
  • Updated August 1, 2024
  • 5 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 Migrating NLU/keyword Virtual Agent topics to LLM topics

    This workflow enables ServiceNow customers to migrate existing Virtual Agent Natural Language Understanding (NLU) and keyword topics to new Large Language Model (LLM) topics. It leverages generative AI to enhance Virtual Agent capabilities without requiring manual recreation of topics. The migration process creates copies of existing topics, converting them into LLM-compatible formats while preserving the original NLU/keyword topics.

    Show full answer Show less

    Key Features

    • Topic Migration Scope: All types of NLU/keyword topics can be migrated, including topics, topic blocks, setup topics, small talk, custom input controls, and custom response controls.
    • Roles and Access: Users with virtualagentadmin or snvadgenai.topicmigrationadmin roles can perform migrations via the Virtual Agent Designer interface.
    • Generative AI for Descriptions: During migration, generative AI can automatically create detailed LLM topic descriptions based on NLU utterances and keywords. Alternatively, existing descriptions can be retained, but empty descriptions must be completed before publishing.
    • Node Migration: NLU node names migrate as-is. The Prompt field from NLU nodes is transferred to the LLM Detail description field, with special handling when scripts or data pills are present—prompting manual review and editing for accuracy.
    • Entity and Vocabulary Migration: Entities mapped in NLU nodes and vocabulary sources (list and table) migrate into corresponding LLM nodes, with vocabulary lists converted to static choice nodes and tables to dynamic choice nodes.
    • System Properties: The system property snvadgenai.utterances.count.for.topic.description controls the maximum number of utterances used by generative AI to construct topic descriptions.
    • Migration Logging: Migration status and issues are tracked in the topicmigrationexecutionitem.list table, accessible via Virtual Agent Designer or downloadable as a CSV for review.

    Practical Considerations

    • Migration does not alter original NLU/keyword topics; it creates LLM-based copies.
    • Publishing of migrated topics is optional and can be controlled post-migration.
    • Empty LLM topic descriptions must be filled before publishing to ensure proper Virtual Agent operation.
    • Manual review and enhancement of migrated node descriptions are recommended, especially when scripts or data pills are involved, to improve user experience.
    • Testing and refining topic descriptions against original utterances can improve relevance and interaction quality.
    • Migration issues can be identified and addressed by reviewing logs and reports provided in the platform.

    The topic migration workflow enables you to migrate your existing Natural Language Understanding (NLU) and keyword topics into new large language model (LLM) topics.

    Now Assist capabilities bring generative AI to Virtual Agent using LLM topics. With topic migration, there’s no need to manually recreate all your NLU and keyword topics to be LLM topics. You can select the topics that you want to migrate into LLM topics from your existing NLU and keyword topics. Migrating NLU and keyword topics doesn’t change the original NLU or keyword topics. A copy of the existing topic is created during topic migration, but the new topic's Model Type field is set to LLM and includes LLM-compatible nodes and descriptions.

    All types of topics can be migrated from the NLU/Keyword model type to the LLM model type during topic migration, including the following asset types:

    • Topic
    • Topic block
    • Setup topic
    • Small talk
    • Custom input control
    • Custom response control

    If no migration issues occur, the preceding asset types can all optionally be published except the topic block. Although dynamic topic blocks can be optionally published, the topic block type is automatically published after it's been migrated.

    Roles and accessibility

    Users with the virtual_agent_admin role or sn_vad_genai.topic_migration_admin role can work with topic migration. Topic migration is accessible through Virtual Agent Designer in the Migrate Topics to LLM option.

    Figure 1. Migrate Topics to LLM option in Virtual Agent Designer
    Migrate Topics to LLM is an option in the Virtual Agent Designer header.

    System properties

    Use the system property sn_vad_genai.utterances.count.for.topic.description to set the maximum number of NLU utterances that can be given to the LLM to generate the topic description for an NLU topic migrated to an LLM topic.

    Topic descriptions

    When migrating NLU topics to LLM topics, you can choose whether you want to have generative AI create your LLM topic descriptions. When on the Settings step of the topic migration workflow, you can choose to Keep current topic descriptions?. The Keep current topic descriptions? option is off by default. When this setting is off, generative AI pulls either NLU intent utterances or keywords from topics to create LLM topic descriptions during the migration process. If the topic is associated with an intent, a maximum of 10 utterances are pulled into the LLM topic's Detail description field under an Additional instructions value. When this setting is on, generative AI doesn’t create LLM topic descriptions and your existing NLU and keyword topic descriptions migrate into your new LLM topics. Migrating existing topic descriptions can result in issues if the existing topic doesn’t have a description.

    The Detail description field is required for LLM topics but not for NLU/keyword topics. If a new LLM topic is migrated over with an empty description, you must add a description prior to that topic being published. Regardless of how topic descriptions migrate, you can edit topic descriptions during the Review descriptions step of the topic migration workflow. You can also test your topic descriptions against the original NLU topic's utterances, if applicable, to improve relevance and topic description strength.
    Note:
    NLU utterances themselves are not effective if you use them as topic descriptions for LLM topics. Avoid using NLU utterances for LLM topic descriptions. LLM topic descriptions require more specific detailed information.
    For more information on editing topic descriptions, see Migrate NLU topics to LLM topics. For examples of strong topic descriptions, see LLM description and instruction guidelines for Virtual Agent topics.

    Node descriptions

    Although NLU topics' node names migrate exactly as-is during migration, the node descriptions can vary. NLU input nodes use the Prompt field. LLM input nodes use the Detail description field. During topic migration, the value in the Prompt field migrates into the Detail description field's value. How the value migrates can vary depending on if that value included text-only, scripts, or data pills. Refer to the following table to view how the values migrate differently and compare examples.

    Table 1. How the Prompt field's value migrates to the Detail description field's value
    NLU Prompt field value LLM Detail description field value
    Contains text Migrates text as-is.

    For example, if the NLU Prompt field's value is Enter the incident number, then the LLM Detail description field's value is also Enter the incident number.
    Contains script or data pill Migrates using the following template: Collect + Node name.

    For example, if you have an NLU node named Get incident number and the Prompt field's value contains a script or data pill, then the migrated LLM node's Detail description field's value is Collect Get incident number.

    If the topics that you plan to migrate include scripts or data pills in their existing Prompt field, review and update the LLM Detail description field after migration for each affected topic. A warning message of Add relevant detail description appears for LLM topics on the Virtual Agent Designer canvas for each node that migrated with the template of Collect + Node name. Updating the description to something more accurate and descriptive improves your users' experience of interacting with the Virtual Agent. For an example of a strong node description, see LLM description and instruction guidelines for Virtual Agent topics.

    The following entities can be migrated from NLU nodes to LLM nodes:

    • simple
    • mapped
    • pattern
    • system-derived

    If an NLU node is mapped to an entity, those utterances' entities are added to the LLM node's Detail description field under an Additional instructions: make-shift header value. The Additional instructions: content differs slightly depending on the type of entity that is migrated. The layout of the Detail description field value resembles something like the following:

    • [Prior NLU Prompt field value] if the value was text-only and not a script or data pill.
    • [Additional instructions: For this input, the data should be extracted from the user input if it has a value from this list of words: [simple entity word list].
    Figure 2. Example of an NLU node with associated entities migrating to an LLM node
    Prompt field values migrate to the Detail description field values along with NLU entities.

    If you had vocabulary sources established in your NLU topic's text nodes, the list and table vocabulary sources migrate into LLM nodes. Text nodes with list vocabulary sources are migrated into static choice nodes. The vocab source selections appear in the LLM static choice node's Detail description field under an Additional instructions: make-shift header value. Text nodes with table vocabulary sources are migrated into dynamic choice nodes. The vocabulary source's table is migrated into the LLM dynamic choice node's Table field. The vocabulary source selections are migrated into a script and appear in the LLM dynamic choice node's Detail description field under an Additional instructions: make-shift header value.

    Migration issues

    You can access migration issue data either by searching for topic_migration_execution_item.list in the Navigation pane or downloading the .CSV file version of the table. After you've migrated topics, the downloadable .CSV file is accessible through the Review migration log option in the topic migration's Migrate topics step. For more information about migration issues, see NLU to LLM migration log.