📺Monitor

Overview

The Monitor feature in ZBrain delivers continuous, end-to-end oversight of every deployed agent and application. It automatically captures each input and output, evaluates the output against a comprehensive set of metrics, and surfaces real-time performance trends, so you can detect issues early, correct them quickly, and maintain consistently high-quality AI interactions.

Monitor schedule evaluations at defined intervals, log both successes and failures, and present the results in an intuitive console. Built-in notifications keep you updated on whether the flow is successful or fails.

Metric categories

• LLM-based -Utilizes a language model to evaluate answers for relevance and factual accuracy.

  • Response relevancy - Measures how accurately the response answers the user’s query.

  • Faithfulness - Evaluates factual alignment with context to minimize hallucinations.

• Non-LLM - Relies on deterministic checks (health, exact match, similarity) without invoking an LLM.

  • Health check - Confirms the app/agent can return a valid response; halts further checks if an invalid response is received.

  • Exact match- Compares the app/agent agent response character-by-character with the expected output.

  • F1 score: Balances precision and recall to evaluate content overlap.

  • Levenshtein similarity: Measures similarity based on edit distance between two strings.

  • ROUGE-L score: Detects the longest common sequence between the response and reference text.

• LLM-as-a-judge - Have an LLM emulate human reviewers on traits like creativity, clarity, and helpfulness.

  • Creativity: Rates originality in response generation.

  • Helpfulness: Evaluates how effectively the response assists in resolving the user’s query.

  • Clarity: Measures how clearly the message is conveyed.

• Performance - Measures the total time (in milliseconds) taken by the LLM to return a response after receiving a query. It provides the Response Latency metric. Using this metric, users can monitor and enforce execution time thresholds for AI Agents, Flows, or Apps.

  • Set thresholds in either seconds or minutes.

  • When a threshold is breached or satisfaction is achieved, the system triggers Success or Failure evaluations and sends relevant notifications.

Key capabilities of Monitor

• Automated evaluation: Assess responses using LLM-based and non-LLM-based metrics.

• Performance tracking: Track success/failure trends.

• Query-level monitoring: Configure evaluations at the individual query level within a session.

• Agent and app support: Monitor both AI apps and AI agents.

• Input flexibility: Monitor responses for .txt, PDF, image, and other file types.

• Notification alerts: Enable real-time notifications for event status updates when an event succeeds or fails.

Monitor interface navigation

The monitor module consists of four main sections, accessible from the left navigation panel:

• Events: View and manage all configured monitoring events

• Monitor logs: Review detailed execution results and metrics

• Event settings: Configure evaluation metrics and parameters

• User management: Configure role-based user permissions

Together, these capabilities provide a single dashboard for validating fixes, identifying quality drift, and ensuring that every user interaction meets your organization’s standards.

Accessing Monitor documentation

ZBrain Builder includes a context-aware documentation panel within the Monitor module.

• Click the '?' icon in the top-right corner of the Monitor interface.

• When clicked, it opens a side panel that displays documentation specifically tailored to the Monitor module, all without requiring you to navigate away from your current screen.