Theme
Classifier Analytics
The Classifier Analytics report provides visibility into how the utterance classifier categorizes incoming messages. The classifier determines whether each message warrants a knowledge base search ("meaningful") or should be filtered out ("noise"). Use this report to evaluate classifier performance and tune the threshold.
Classifier analytics showing classification distribution pie chart, noise filter trend, confidence histogram, and top filtered messages
Classification Distribution
A pie chart shows the breakdown of all classified messages by category:
| Category | Description |
|---|---|
| Meaningful | Messages that triggered a knowledge base search -- the customer asked a question or raised an issue |
| Noise | Messages filtered as non-actionable -- greetings, filler, acknowledgments, data collection responses |
A healthy distribution depends on your conversation type. Voice calls typically have more noise (greetings, pauses, filler words). Chat conversations tend to have a higher proportion of meaningful messages.
Noise Filter Rate Trend
A daily line chart shows the percentage of messages filtered as noise over the selected period.
- Stable rate indicates consistent classifier behavior
- Rising rate may mean agents are handling more routine interactions, or the classifier threshold needs adjustment
- Dropping rate suggests more substantive conversations or a threshold that is too low
Confidence Scores
Each classification includes a confidence score between 0 and 1. The report shows a histogram of scores grouped into 10% buckets.
Most classifications should cluster at high confidence (80-100%). A large count in low-confidence buckets indicates the classifier is uncertain and may need tuning:
- Add domain keywords to boost confidence on industry-specific terminology
- Adjust the classifier threshold in Tenant Configuration to change the cutoff point
- Enable prompt caching to ensure consistent classifier behavior
TIP
Low-confidence classifications near the threshold boundary are the ones most likely to be wrong. Review these cases to determine if the threshold should be raised or lowered.
Top Filtered Messages
A ranked table of the most frequently filtered messages. Review these for false positives -- if meaningful customer questions appear in this list, the classifier is incorrectly filtering them.
| Column | Description |
|---|---|
| Message | The text of the filtered utterance |
| Count | Number of times this message was filtered |
WARNING
If meaningful questions like "What is the refund policy?" appear in the top filtered list, consider lowering the classifier threshold or adding relevant domain keywords. See Domain Keywords.
Classifier Cache Hit Rate
Shows the percentage of classification requests served from cache instead of calling the LLM. A higher cache hit rate means lower token costs and faster response times.
The classifier cache stores decisions for repeated utterances (e.g., "hello", "thank you") so they are not re-evaluated by the LLM. Configure the cache TTL in Tenant Configuration.
Filters
| Filter | Description |
|---|---|
| Queue | Filter metrics to a specific CCaaS queue |
| Date range | Custom start and end date pickers |