Topic discovery
Smart Topic Builder topic discovery generates new topics and topic definitions from untagged comment data
Processing time: Around 15-30 minutes; sends an email notification when complete
- Number of existing topics in the tag pool
- LLM errors and error correction
- Rate limits
Smart Topic Builder analyzes comment data and existing topics to generate proposed topic names and hierarchy based on required date, topic set, and character minimum parameters, and on optional description, comment, and segment field parameters.
The topic set must have at least 1 active topic, at least 200 untagged phrases, and a topic set description for topic discovery. Up to 10,000 untagged phrases are processed for discovery. Only one discovery job per topic set can run on a single instance at one time. To initiate a new discovery while one is running, the existing discovery job must be completed or failed. A warning displays when attempting to overwrite previous results. Discovery cannot be canceled.
-
Initiating topic discovery removes all results from previous discovery jobs.
-
Only the user that initiated topic discovery is notified of completion or failure.
-
The topic discovery prompt is not retained between discovery jobs, and the parameters that defined the discovery job are not stored.
To begin a new topic discovery, use the Discover Topics with AI button on the Topics screen.
The following images show the Discover Topics with AI button and the dialog that displays.
The Discover Topics with AI dialog includes these options:
- Topic Set
- Required. Specify the topic set to be processed for topic discovery. Note: If a topic set has no comments, it is not shown in the dropdown.
- What would you like AI to discover?
- Optional. Provide instructions or a description to give the AI more context about the discovery and desired results.
For example, the instructions to Identify actionable topics from customer feedback related to product quality ask the language model to limit results to topics that are actionable and related to quality.
Another description might ask the language model to limit results further to premium products, or to broaden results to include feedback about retail experience in addition to product quality.
- Date range
- Required. Specify the start and end date for comment data to be processed for discovery.
- Comment Fields
- Optional. Limit analysis to specified comment fields.
- Segmentation
- Optional. Limit analysis to specified segments, like units or demographics.
- Comment character minimum
- Required. Default 20. Comments with fewer characters than the specified minimum are ignored.
Discovery processing typically takes no more than 15 minutes. When discovery is complete, admins are notified in Experience Cloud and via email. Existing results are cleared if a new discovery is initiated.
When discovery is complete, generated topic suggestions appear on the Topics screen's Topic Discovery tab. Select a topic in the list of suggestions to review the AI's rationale and supporting examples in the side panel.
Use the Edit Definition button to manually modify the generated topic definition results.
Use the Accept or Reject button to accept or reject the topic suggestion and complete review. Accepting a topic suggestion removes it from the Topic Discovery tab and adds it to the list in the Topic Hierarchy tab as an unpublished draft. Rejecting a topic suggestion removes it from the list; this action cannot be undone.
If a new parent topic and a new child topic are proposed, accept the parent topic first before accepting the child topic to maintain the hierarchy. If the parent topic does not exist in the topic set, the accepted child topic will be brought in as a new parent topic. Accepted topics can be moved in the hierarchy as needed.
The following image shows the Topics screen's Topic Discovery tab, with the rationale and definition displayed for the Customer Service - Question Handling topic suggestion.
Suggested topics are shown with their hierarchy level; the number and percentage of comments in which the topic was detected; and, if one exists, the existing topic that mostly closely resembles the suggested topic.
Quality guardrails — Smart Topic Builder evaluates potential topics against strict default criteria, which include a directive to prioritize topics that demonstrate actionability, strategic value, and novelty. If a topic matches your prompt, but the LLM determines that topic is too vague or not actionable, it may filter that topic out in favor of others it deems more useful.
Data-driven reality — Smart Topic Builder grounds its suggestions strictly in the designated comment data. If the discovery prompt specifies a specific trend, but the volume of relevant, untagged comments doesn't meet the necessary minimum thresholds to form a statistically significant cluster, the LLM instead suggests the most prevalent themes it found in the text.
The nature of generative AI — Processing open-ended text alongside rigorous system rules requires the AI to balance dozens of competing constraints simultaneously. Because generative AI models are inherently non-deterministic, highly complex or overly broad free-text requests can occasionally cause the model to prioritize its default structural filters over the specific nuances of the discovery prompt.
