Analytics pipelines

This document covers the pipeline framework in Analytics and Insights: the built-in pipelines that power Agent Performance, Quality Monitor, Customer Insights, and parts of At a Glance, and the visual node-based editor for building custom evaluation logic. Use the Pipelines page to check that the data behind the dashboards is being produced. Navigation: Project → Insights → Pipelines Each pipeline card displays its name, description, enabled or disabled status (green Enabled badge or gray Disabled badge), trigger count, health dot, and last processed timestamp. Use the search bar at the top to find pipelines by name. Select a card to configure the pipeline.

Card element	Description
Enabled	Whether the pipeline is on for this project.
Triggers	How many triggers start the pipeline, such as a session ending.
Health dot	The success rate of runs in the last 24 hours: completed runs ÷ all runs, running ones included. Green above 95%, amber from 50% to 95%, red below 50%, grey when there are no runs.
Last processed	When the pipeline last ran. Never processed means the pipeline has no recorded runs. On an enabled pipeline, it means the dashboards it feeds get no new data.

Sections

Section	Purpose
Built-in	Pre-configured pipelines that ship with the platform, ready to enable with a single toggle.
Custom	User-defined processing workflows created using the visual pipeline editor.
Recent Runs	Pipeline execution history with timestamps, durations, status (success or failure), and links to output data. See Recent runs.
Data	Pipeline output data available for dashboard integration, export, and downstream consumption.

Built-in pipelines

The platform ships with eleven pre-built pipelines covering evaluation, classification, detection, and monitoring needs. Enable each pipeline with a single toggle. The defaults need no configuration.

Pipeline	Description
Sentiment Analysis	Per-message sentiment scoring with conversation-level trajectory analysis. Feeds the Sentiment Trajectory chart in Customer Insights.
Intent Classification	Classifies conversation intent using LLM analysis with customer-defined taxonomy or auto-discovery. Feeds the Intent Distribution chart in Customer Insights.
Quality Evaluation	LLM-as-judge quality evaluation with configurable rubric dimensions. Feeds the Quality Score in Agent Performance and Quality Monitor.
Hallucination Detection	Detects unsupported claims, self-contradictions, and factual accuracy issues in agent responses.
Knowledge Gap Analysis	Identifies gaps in knowledge base coverage by analyzing retrieval precision and uncovered topics.
Guardrail Analysis	Evaluates guardrail effectiveness. Detects false positives, false negatives, and bypass attempts.
Context Preservation	Evaluates whether agents preserve relevant user and workflow context through the conversation.
Friction Detection	Detects user frustration signals: rephrased questions, message escalation, caps, and exclamation patterns.
Anomaly Detection	Monitors analytics metrics for statistical anomalies using z-score and SPC (Statistical Process Control) charts.
Drift Detection	Monitors analytics metrics for gradual performance drift by comparing baseline and current windows.
Evaluation Run	Executes persona × scenario × evaluator matrix evaluation with bias mitigation and trajectory scoring. Powers eval matrices.

Custom pipelines

Custom pipelines let you define your own analytics logic using a visual, drag-and-drop node-based editor. Use custom pipelines to build organization-specific evaluation criteria that go beyond the built-in set, for example regulatory compliance checks, brand voice adherence, or questions such as “was the customer’s goal met?” or “what is the retention risk?”. A typical custom pipeline is a small flow: a trigger, a step that reads the conversation, an analysis step (for example an LLM prompt that returns the fields you ask for), and a step that stores the results.

Option	Description
When it runs	When a conversation ends, on each user or agent message, on a schedule, or by hand.
Ready-made steps	Get data, prepare data, analyze (LLM evaluate, intent, quality, goal completion, and more), add logic, call out to email, Slack, HTTP, or Kafka, and store results.
Results	Every field the pipeline stores becomes a column you can use in Custom dashboards.

You can also tune a built-in pipeline instead of building a new one. For example, give Intent Classification your own intent list, or give Quality Evaluation your own dimensions and weights. On the Custom tab, select Create Pipeline to open the editor. Each custom pipeline card shows its description, node count, owner, last change, and status.

How the editor works

The editor presents a visual canvas where you construct a pipeline as a directed graph of connected nodes. A Node Palette on the left side provides draggable node types organized into categories:

Category	Node types
Get Data	Read Conversation, Read Message Window, Database Query, Inspect Output
Prepare Data	Aggregate, Filter, Transform Data
Analyze	LLM evaluate, intent, quality, goal completion, and more
Logic	Sub-Pipeline (reuse an existing pipeline as a step), Wait for Event (pause execution until a named event fires)
Act & Store	Call out to email, Slack, HTTP, or Kafka, and store results

The core pattern: define a trigger, connect one or more processing nodes, then wire the output to a storage or metrics node.

Pipeline structure

Component	Description
Trigger	Defines when the pipeline runs. Trigger types include Kafka events, scheduled intervals, in-platform events (for example, conversation completed), or filtered subsets of sessions.
Processing nodes	The evaluation and transformation logic applied to each triggered item. Nodes can read conversation data, query databases, aggregate metrics, filter records, transform data shapes, or call LLMs for evaluation.
Metrics output	The results the pipeline produces. Metrics take the form of named, typed values (counts, percentages, scores) that appear in the Data section and that you can wire into dashboards.

Pipeline lifecycle

Pipelines follow a Draft, Validate, Test, Activate lifecycle. The editor toolbar shows the current state (Draft badge, “Validation passed” indicator) and provides buttons for Test, Validate, Save, and Activate. You can iterate on draft pipelines without disrupting live ones, save drafts as you go, validate to catch configuration errors before deployment, and run test executions to verify output before activating. Keep these rules in mind:

Test a pipeline on one conversation before you activate it.
Drafts don’t run.
Pipelines don’t backfill older conversations. They process only conversations that end after activation.
Pipelines skip eval and synthetic sessions.

Sub-pipelines

You can reference one pipeline as a step inside another, enabling reuse and composability. For example, a “Compliance Evaluation” pipeline could call both the built-in “Guardrail Analysis” and a custom “Regulatory Wording Check” as sub-steps.

Attach to dashboards

Once a custom pipeline produces data, every field it stores becomes available to Custom dashboards, where you can chart it next to built-in metrics.

Start with the built-in pipelines to establish baselines, then create custom pipelines for organization-specific quality dimensions. The validate-before-activate workflow catches configuration errors before pipelines go live.

Recent runs

The Recent Runs tab shows the run history for the selected window.

Element	Description
Runs by trigger	Runs started by a session ending, by an event, or by hand.
Run outcomes	Completed, failed, and still running. Failed runs mean gaps in the quality and sentiment dashboards.
Average duration	Mean run time.
Run filters	Filter by pipeline type, status, trigger, and window.
Run list	Each run’s status (a tick for completed, a cross for failed), trigger, start time, and duration.

Run history is in early access and may be incomplete. It covers runs on this platform only, from when the feature was turned on.

Get started

Build

Test and evaluate

Deploy

Analytics and insights

References

Sections

Built-in pipelines

Custom pipelines

How the editor works

Pipeline structure

Pipeline lifecycle

Sub-pipelines

Attach to dashboards

Recent runs

​Sections

​Built-in pipelines

​Custom pipelines

​How the editor works

​Pipeline structure

​Pipeline lifecycle

​Sub-pipelines

​Attach to dashboards

​Recent runs

Sections

Built-in pipelines

Custom pipelines

How the editor works

Pipeline structure

Pipeline lifecycle

Sub-pipelines

Attach to dashboards

Recent runs