Skip to main content

​
Introduction

Nebuly provides an advanced “user intelligence layer” that automatically analyzes your LLM interactions to reveal user intent, friction points, and satisfaction levels out of the box. However, while the platform can semantically interpret conversation transcripts, it lacks the organizational context required to drive strategic business decisions. With external metadata, distinguishing if a “frustrated” user is a high-value premium customer or if a specific drop-off pattern is localized to a certain department or geographic region, becomes very easy. This is why utilizing Nebuly’s tagging capabilities is essential for moving from individual session tracking to organizational intelligence.
Tagging and enrichment overview in Nebuly

​
Where tags appear in Nebuly

Once tags are ingested, they populate several key areas within the Nebuly platform to help you “see the forest for the trees”:

​
1. Advanced filtering and segmentation in reports and in the Nebuly UI

Tags become immediately available as filters throughout the platform. This allows you to slice your data to see, for example, only Finance department interactions where users expressed Frustration. For details see our documentation on filtering and searching.

​
2. ROI & “hours-saved” reports

By tagging interactions with a department or role, Nebuly can calculate the specific Business Case for AI. For instance, you can view a report showing that the Engineering department saved 450 hours this month, whereas Legal saved only 20, signaling a need for better legal-specific training.

​
3. Risk & compliance alerts

Tags allow security teams to identify Departmental Risk Variance. If the platform detects a spike in PII exposure, tags help pinpoint if the issue is restricted to a specific team or seasonal period (e.g., hiring season), enabling targeted intervention. For details see Alerting.

​
Technical implementation

Tagging is handled via the Interaction API, specifically the trace_interaction endpoint. Metadata is sent as a tags object consisting of key-value pairs associated with a specific interaction: 
{
  "tags": {
    "department": "Engineering",
    "role": "senior_engineer",
    "country": "DE",
    "customer_tier": "enterprise"
  }
}
You can find more details in our Interaction API documentation.

​
RAG sources

If your agent retrieves context (a RAG pipeline, a database lookup, a search tool), you can send the data sources it used alongside each interaction. Like tags, this adds context Nebuly cannot infer from the transcript alone, this time about where the answer came from. RAG sources are sent through the Interaction API as retrieval steps in the interaction’s traces. Each retrieval records the source that was queried, the query sent to it, and the results returned: 
{
  "source": "vector_db",
  "input": "What is the refund policy?",
  "outputs": ["Refunds are accepted within 30 days of purchase...", "..."]
}
Once ingested, the data source becomes a filter and group-by dimension (shown as RAG Source), and Nebuly surfaces the most-used RAG sources across your users. Why this matters: with the data source recorded on each interaction, you can correlate user frustration and behavior with the specific sources that were queried and the data retrieved. For example, you can check whether a spike in failures or frustration is concentrated on one knowledge base, which lets you tell a model problem apart from a weak, outdated, or missing data source and fix the right thing. See filtering and searching to slice by data source. By adding specific tags, you can create Behavioral Archetypes (e.g., Power Users vs. Skeptics) and track the Adoption Lifecycle across your organization or customer base.
Tag CategoryKey ExamplesStrategic Value
Organizationaldepartment, role, seniorityMeasure Departmental ROI and identify which roles drive adoption.
Geographiccountry, region, office_locationDetect Geographic Friction, such as poor model performance for specific languages.
User Statuscustomer_tier, subscription_planPrioritize improvements for high-value segments and track Silent Churn.
Technicalmodel_version, prompt_template_idConduct A/B tests to see which model version yields higher Goal Completion Rates.