Sentiment Tracking

Sentiment tracking analyses user messages in AI agent conversations to measure satisfaction and detect negative trends. Use it to monitor conversational quality at scale and catch issues before they impact user experience.

Metrics

NodeLoom tracks the following sentiment metrics for each AI agent workflow:

MetricTypeDescription
Total messagesCounterTotal number of user messages analysed across all conversations.
Positive countCounterMessages classified as positive (satisfaction, gratitude, approval).
Negative countCounterMessages classified as negative (frustration, complaints, dissatisfaction).
Neutral countCounterMessages classified as neutral (factual questions, instructions, acknowledgements).
Average scoreFloat (0--1)Running average sentiment score across all messages. 0 = fully negative, 0.5 = neutral, 1 = fully positive.

Classification model

Sentiment is classified using a lightweight, on-device model that runs in the backend. No user messages are sent to external services for sentiment analysis.

Alert Types

Sentiment alerts are triggered when metrics cross configurable thresholds:

Alert TypeDescription
Negative spikeA high proportion of messages in the evaluation window were classified as negative. Indicates a sudden surge in user dissatisfaction.
Score dropThe average sentiment score dropped significantly compared to the previous period. Indicates a gradual decline in conversation quality.
High volumeThe total number of messages exceeds a threshold, which may indicate an agent loop, spam, or a misconfigured workflow generating excessive messages.

Configurable Thresholds

Each team can customise sentiment thresholds from the workspace monitoring settings:

SettingDescription
Negative spike thresholdPercentage of negative messages in the evaluation window that triggers an alert.
Score drop thresholdPercentage decrease in average score compared to the previous period.
Evaluation windowThe rolling time window used to calculate alert conditions.
High volume thresholdMessage count that triggers a high-volume alert.
Minimum sample sizeMinimum messages in the window before alerts can fire. Prevents false positives from small samples.

Tuning for your use case

Customer support agents may benefit from a lower negative spike threshold (e.g., 30%) to catch issues early. Internal tool agents with terse user messages may need a higher threshold to avoid false positives.

The sentiment dashboard provides trend views at multiple granularities:

ViewPeriodUse Case
DailyLast 30 daysSpot day-to-day fluctuations and correlate with deployments or incidents.
WeeklyLast 12 weeksIdentify weekly patterns (e.g., Monday spikes) and measure improvement over time.

Each trend view shows the sentiment score line chart, positive/negative message counts as stacked bars, and any triggered alerts overlaid as markers.

Per-Workflow Breakdown

Sentiment metrics are tracked per AI agent workflow, so you can compare performance across different agents. The monitoring dashboard shows a sortable table of all agent workflows with their current average score, message counts, and alert status. Use this to identify which agents need prompt improvements or additional tools.

Notifications

Sentiment alerts use the same notification channels as anomaly detection and drift alerts -- email and webhook. Configure preferences per team from the workspace monitoring settings.

Next Steps