Version 2.3.7 | July 28, 2025

Overview

ZBrain Builder 2.3.7 introduces significant enhancements to evaluation workflows, model integrations, and agent orchestration. Key highlights include prompt-level monitoring, integration with Gemini 2.5 Pro and support for Microsoft Semantic Kernel, enhanced guardrails for safer outputs, and streamlined RAG setup with actionable guidance. These updates boost reliability, flexibility, and user control across the platform.

ZBrain Builder 2.3.7 release overview

Component

Capability

What it delivers

Evaluation framework

Prompt-level monitoring and in-context event creation from prompt logs.

Enables users to evaluate prompt outputs directly, create monitoring events without leaving the logs view, and configure performance tracking without context switching.

Custom notification logic and event-driven automation based on evaluation outcomes.

Flexible control over when and how to respond to events under user-defined success or failure conditions.

Integrations

Integration of Gemini 2.5 Pro in the Ask AI Model step within Flow

Enables users to leverage Gemini 2.5 Pro for improved output quality and multimodal support.

Guardrails

Guardrail enforcement with output checking and content filtering on App-generated outputs

Improves safety and compliance by automatically scanning App responses for harmful, misleading, or policy-violating content and filters applicable for output checking

RAG

In-context error messaging with actionable guidance for indexing failures during knowledge base creation.

Faster issue resolution and a smoother user experience through clear, actionable guidance shown exactly where failures occur

Inline guidance on chunk length impact during knowledge base setup.

Helps users choose optimal chunk settings by clearly explaining trade-offs between speed and accuracy.

Improved UI clarity for segment configuration during custom chunk settings

Enhances user understanding and usability by explaining segment identifiers

Agents

Integration of Knowledge Base Search as a default tool within Crew Agent configuration

Empowers agents to retrieve and reason over relevant enterprise knowledge using structured instructions

Agent type filtering on the Agents page with options for All, Crew, and Flow Agents

Enables users to quickly navigate and manage a large agent library by isolating specific agent types

Integration of Microsoft Semantic Kernel as a selectable Crew Framework for multi-agent orchestration

Expands orchestration flexibility by enabling users to build Crew Agents using Microsoft’s open-source framework

Apps

Toaster notification on successful save of system instructions in prompt configuration.

Instant feedback that confirms changes are saved, boosting user confidence and reducing confusion.

New features

Evaluation framework

In-context monitoring from prompt logs

Users can now initiate event monitoring directly from prompt logs, enabling seamless evaluation of model performance (e.g., accuracy, creativity, and hallucination detection) without needing to navigate away from the Prompts section. This feature bridges the gap between prompt experimentation and operational observability.

Navigation: Login → Prompts → Select Prompt → Prompt Overview-> Go to Prompt logs → Click Any Log → Monitor Button

Clicking this opens the Event Monitoring configuration modal, pre-filled with the monitored input (from system message or user prompt) and output (from the assistant’s response). Users can define evaluation conditions (e.g., Similarity < 0.4, Hallucination = true), set notification preferences, choose evaluation frequency (e.g., hourly, daily), and newly created events will appear under Monitor.

Key outcomes

  • Create monitoring events directly from prompt logs without needing to switch to the Apps section.

  • Centralized visibility, as all created events automatically appear under the Monitor section for tracking.

Conditional notifications for event monitoring

Users can now configure precise notification triggers for evaluation events, based on success, failure, or both, and attach custom flows to handle those events. Notification triggers only for the selected outcomes based on evaluation results. This enhancement allows proactive and automated response behavior without manual oversight for every evaluation outcome.

Navigation: Login → Monitor → Event → Event Settings → Enable Notification Toggle → Checkboxes

Key outcomes

  • Targeted alerts based on success, failure, or both, reducing notification noise.

  • Automated actions through custom flows for each event outcome.

  • Built-in validation ensures only properly configured flows are used.

  • Quick testing with in-context simulation of flow execution.

  • Reliable persistence of all settings with a single update.

Integrations

Gemini 2.5 Pro integration in the Ask AI model step in Flow

ZBrain Builder now supports Gemini 2.5 Pro as a selectable model in the Ask AI Model step within Flow. This integration expands the model roster available to users, allowing selection of Gemini 2.5 Pro for workflows that benefit from its enhanced language performance and multimodal capabilities.

Navigation: Login → Flows → Open any Flow → ZBrain → Ask AI Model → Model

Key outcomes

  • Expanded model flexibility with enhanced content quality and multimodal support

  • Session persistence with the selected model without requiring re-selection

  • The model works with performance stability

Guardrails

Output guardrail & jailbreak filtering

ZBrain Builder now supports guardrails, such as output checking and Jailbreak filtering, on App-generated outputs, ensuring app responses remain safe, compliant, and aligned with enterprise policies. This enhancement includes built-in detection for high-risk categories such as misinformation and system prompt leakage, further strengthening response integrity across all apps.

Navigation path: App → Select any app→ Configure your app→ Add Guardrail → Select Output option

Key outcomes

  • Policy-aligned automatic output control

  • Expanded safety coverage with detection of critical risk types such as misinformation and system prompt exposure

  • Separate toggles for input, output, and jailbreak filtering offer granular control, all accessible within a unified Guardrail panel.

  • Guardrail settings persist across sessions, ensuring that monitoring behavior remains stable and predictable when apps are reconfigured or reused.

Apps

Confirmation for system instructions saving

This feature adds a toaster notification when users save system instructions within the prompt configuration modal. Upon clicking the ‘Save’ button, a brief confirmation message (e.g., “Settings saved successfully”) now appears, assuring users that their changes have been recorded.

Navigation path: Login → Apps → Select any App → Configure Your Bot → System Instruction → Click on Library → Create a New Prompt

Key outcomes

  • Users receive immediate feedback that their input has been saved.

  • Reduced uncertainty by eliminating confusion over whether changes were applied.

  • Improved UX consistency that aligns with common save-feedback patterns across the platform.

  • Faster task completion as Users can confidently move to the next step without hesitation.

Agents

Knowledge base integration for the agent crew

This feature introduces Knowledge Base(KB) Search as a default tool available when configuring an Agent Crew. Users can now enhance agent capabilities by attaching a Retrieval-Augmented Generation (RAG)-powered knowledge base, enabling the agent to access domain-specific context during execution.

Configuration options include:

  • Select a knowledge base from a dropdown list of available KBs.

  • Add a description to define the purpose and scope of the KB for the agent.

  • Set max token limit to control how much retrieved context is passed to the agent.

  • Define agent instructions that clearly explain:

    • How the agent should use the KB.

    • The nature of the tool (e.g., RAG-powered retrieval).

Navigation path: Agent crew → Define Crew Structure → Add Agent → Create New → Agent Tools → Default Tools → Knowledge Base Search

Key outcomes

  • Agents now operate with access to verified knowledge base, improving factual accuracy.

  • Custom agent behavior with explicit usage instructions guides how the agent queries and interprets the KB.

  • Enhances task performance by anchoring responses in structured, contextual data.

  • Scalable agent design that simplifies building specialized agents without embedding static context in prompts.

Agent type filters on the agents interface

This feature introduces a filter dropdown on the Agents interface, enabling users to refine the list of agents by type. With clearly defined filter options, users can now switch between viewing Crew Agents, Flow Agents, or All Agents in a single unified interface, making it easier to locate and manage agents based on structure and function.

  • All: Displays both Crew and Flow Agents in a combined view.

  • Crew: Displays only Crew Agents.

  • Agent: Displays only Flow Agents.

Navigation path: Login → Agents → Filter (top-right dropdown)

Key outcomes

  • Improved discoverability of relevant agents by type.

  • Organizational clarity that helps users distinguish between orchestration-level and task-specific agents.

Microsoft Semantic Kernel integration for crew agents

ZBrain Builder now supports Microsoft Semantic Kernel as a new framework option for configuring Crew Agents, alongside existing options such as Google ADK and LangGraph. This integration enables users to orchestrate multi-agent workflows using Microsoft Semantic Kernel, while maintaining full compatibility with ZBrain Builder's existing capabilities.

Navigation path: Create Agent Crew → Crew Overview → Crew Framework → Microsoft Semantic Kernel

Supported functionalities:

  • Flow agent execution – End-to-end coordination of sub-agents using the Semantic Kernel Action Planner.

  • Tool compatibility – Supports both default and custom tools across agents.

  • Memory management – Works with both crew-level and tenant-level memory contexts.

  • Model interoperability – Enables smooth model selection, switching, and invocation within the new framework.

  • Thought trace – Captures and visualizes reasoning steps during execution.

  • Streaming support – Delivers real-time agent responses without disruption.

Key outcomes

  • Greater orchestration flexibility with the most suitable agent framework for a specific use case.

  • Full support for existing features, including tools, memory, and streaming, ensures a consistent user experience across frameworks.

  • Improved interoperability as it seamlessly integrates with ZBrain Builder’s multi-model ecosystem and agent architecture.

  • Broader ecosystem support that opens the door to using Microsoft-native planning logic within ZBrain Builder’s low-code interface.

Improvements

RAG

Actionable error messaging for indexing failures

ZBrain Builder now provides clear, contextual error feedback when indexing fails during Knowledge Base creation. Instead of a generic failure message, users will see detailed error explanations, such as unsupported formats, along with specific corrective actions. This error message is surfaced directly on the creation page and follows a consistent, user-friendly UI design, ensuring users know exactly how to fix the issue without unnecessary backtracking or guesswork.

Key outcomes

  • Clarity at the point of failure during indexing.

  • Actionable guidance helps users correct issues.

  • By removing ambiguity, users can fix and retry without having to restart the KB creation process.

Chunk length guidance in Knowledge Base(KB) setup

Users configuring chunk length during knowledge base creation will now see clear, contextual explanations of how high or low values impact processing. The interface now provides a simple, user-friendly tool tip that describes the trade-offs between accuracy, speed, and indexing quality, helping users make informed decisions during setup.

This guidance appears inline or as a tooltip next to the chunk length field for immediate visibility.

Key outcomes

  • Users understand the performance trade-offs before selecting a chunk length.

  • Reduces misconfiguration and reprocessing due to poorly chosen chunk values.

  • Improved KB quality as it helps optimize for use case specific goals (e.g., accuracy vs. speed).

Enhanced Segment Identifier tooltip in custom chunk settings

This update introduces improvements to the segment identifier configuration text box to enhance clarity and user experience:

  • Updated help text: The Segment Identifier field now includes clear explanations of characters like \n (newline) and \t (tab), helping users understand how content will be segmented.

These changes ensure a more intuitive setup process and better-informed user actions during knowledge base creation.

Key outcomes

  • Improved clarity as users clearly understand how segment identifiers like \n and \t affect chunking.

Last updated