Executive Summary

AI Core is an agentic AI platform powering most of the AI capabilities within M42 Professional. AI Core 26.1 expands the platform from a RAG-focused service into a broader agentic AI platform. The main addition is the new Agents subsystem: tool-using agents with chat, scheduled and agent-to-agent triggers, long-context handling, dedicated metrics, source/tool filtering and admin UI support.

The release also strengthens governance, observability and operational control. Content-safety checks now cover more interaction points, RBAC support has been extended for Confluence and Matrix42 Enterprise, citation extraction improves source attribution, and new telemetry gives teams a clearer view of LLM, embedding, tool and agent activity.

The tools ecosystem has been modernized around JavaScript skills, MCP servers and internal Go tools, supported by a dynamic Tools Repository UI and an updated MCP integration. AI Core 26.1 also introduces crawler-only mode for externally orchestrated ingestion, broader automatic testing, UI/UX improvements, dependency updates and bundled stability fixes across the stack.

Key Features

Agents and Multi-Agent Orchestration

• New agentic pipeline with OpenAI-compatible chat completion endpoint, streaming support, session handling, per-agent rate limits and integrated platform controls.
• Support for chat, scheduled and agent-to-agent triggers, including overlap protection, graceful shutdown and internal-service authentication.
• Agent-to-agent orchestration with recursion-depth controls for safer multi-agent workflows.
• Native Go-based agent tools, including keyword search and image-analysis capabilities.
• Configurable source and tool filtering per agent.
• Agent support for client-provided message history, extended response metadata and context-status reporting.
• Context summarization for longer conversations, so older turns can be summarized instead of simply removed.
• Dedicated agent logs, trigger analytics and token-usage views.

Governance, Access Control and Content Safety

• Content-safety checks expanded to images, tool outputs and ingestion workflows.
• Configurable regex-based pre-filters for fast handling of known unsafe or malformed prompt patterns.
• Confluence RBAC support with end-user permission checks and safer default behavior when authorization context is incomplete.
• Matrix Enterprise ESM ESM RBAC handling tightened for scenarios where end-user context is unavailable.
• Secret handling standardized through struct tags, improving consistent redaction and preservation of sensitive fields.
• Additional hardening around logging, configuration constraints, authentication fallback and admin actions.

RAG, Knowledge Quality and Attribution

• Document quality checks during indexing, including text-length, word-count, character-ratio and LLM-based assessment rules.
• Structured quality-check results surfaced in admin UI with progress reporting.
• Citation extraction to identify which retrieved documents were actually used in a response, improving verifiable source attribution.
• RAG Index upload flow expanded for multi-file and folder uploads, with managed upload identifiers instead of exposed server paths.
• Improvements to source matching, result ordering, FTS synchronization and document filtering.

Tools, Skills and MCP

• New Tools Repository replacing the static tools configuration UI, with editor support, hot reload and safer validation.
• Tools reorganized into three primary categories: JavaScript skills, MCP servers and internal Go tools.
• Migration to the official MCP Go SDK, including improved SSE transport and session handling.
• Centralized ToolSet abstraction for per-request and per-agent tool execution.
• Independent enablement controls for Passthrough and RAG usage across tool types.
• Per-agent tool timeout configuration for RAG, FTS and vision tools.
• Expanded Matrix42 tool coverage for announcements, orders and ticket workflows.
• Shared ESM authentication package to centralize end-user identity handling and retry behavior.

Telemetry, Observability and Analytics

• New LLM telemetry capturing provider, model, task type, token usage, duration, errors, streaming mode and transaction IDs.
• Embedding telemetry with daily aggregation by date, model and backend.
• OpenTelemetry span export for LLM API calls.
• Token analytics split by input/output and task type, with dedicated Tools and Embeddings views.
• Tool success/error analytics and date-range filtering for security analytics.
• Reusable feedback/rating system across RAG, Passthrough and Agent chat.
• Master API v2 response and endpoint cleanup, including improved range support for security statistics.

Automatic Testing

• New Automatic Tests module for RAG quality evaluation, multi-turn conversations, CSV import/export and dedicated test storage.
• Additional test suites for embeddings, firewall, FTS and pre-LLM behavior.
• Agent workflow tests with tool-trace capture and LLM-based evaluation.
• Broader unit and integration coverage across admin UI, API, crypto, dbtools, metrics, tools and user-doc modules.

Operational Modes and Crawlers

• New crawler-only mode exposing a stateless REST API for Confluence, Web and SharePoint crawls.
• Direct write support to Azure Blob Storage for externally orchestrated ingestion scenarios.
• Crawler/plugin stability improvements, including progress reporting, HTTP error handling, pagination handling and CSV edge cases.

Admin UI and UX

• Unified admin layout with shared shell/sidebar/main patterns.
• Improved responsive and mobile experience.
• Better agent chat UX, expandable rows, source lists, scroll behavior and chat-box sizing.
• Refined logs, search, ratings and analytics views.
• Improved configuration validation, field bounds, tab highlighting, notifications and documentation links.
• Updated architecture diagrams and agent documentation.

Improvements and Maintenance

• Agent configuration and lifecycle refinements, including UUID-based identifiers, improved context behavior, cleaner tool-reference handling and updated docs.
• RAG and search refinements, including BM25/FTS behavior, source ordering, filtering and indexing synchronization.
• Tools/MCP cleanup, including modal validation, clearer tool names, structured logging, JWT helper extraction and deprecation cleanup for legacy tools.
• Platform maintenance, including dependency updates, removal of deprecated imports, Ubuntu-based Dockerfile support, tini init process, Chromium process handling and Consul-token path updates.
• Admin UI cleanup, including reduced duplicated log/search code, better sidebar reuse, expanded documentation links and refined test-tab workflows.

Bundled Fixes

Agents and Tool Execution

• Improved usage accounting, tool/source access checks, FTS source handling, rate-limiter behavior and save-failure handling.
• Fixed context propagation and several edge cases in agent tool execution.

RAG, Search and Embeddings

• Corrected source ordering, source matching, document ordering and BM25/FTS behavior.
• Improved embedding token accounting, URL matching in tests and handling around zero-vector normalization.
• Removed redundant SQLite dependency that could unnecessarily pull in cgo.

Crawlers, Plugins and Ingestion

• Improved crawler behavior around HTTP errors, pagination, CSV trailing rows and progress reporting.
• Added resilience improvements for plugin failure modes and ingestion workflows.

Tools and MCP

• Improved passthrough tool identity propagation.
• Addressed MCP test-connection and long-lived SSE stream handling.
• Corrected internal-tool provider classification and related metrics categorization.
• Consolidated smaller tools-package fixes.

Admin UI and API

• Fixed tab highlighting, test-tab flows, long-sidebar chat sizing, modal submit behavior, clipboard handling and loading-state edge cases.
• Improved Azure OpenAI deployment-name handling for names containing dots.
• Fixed admin UI tests affected by agent-list behavior.

Access Control, Secrets and Platform Stability

• Strengthened RBAC handling for requests without complete end-user context.
• Improved secret-field redaction and preservation when editing sources.
• Added process-management improvements for Chromium-based workloads.
• Bundled minor typo, cleanup and small-fix changes.