Executive Summary

Matrix42 is pleased to announce the general availability of AI Core 1.3.0. This major milestone release marks the evolution of AI Core from a specialized Retrieval-Augmented Generation (RAG) engine into a comprehensive, enterprise-grade Agent Platform. 

Version 1.3.0 introduces significant architecture and capability enhancements, enabling autonomous agents to execute complex file operations, retain operational memory, enforce human approval workflows for critical actions, and seamlessly integrate into enterprise service management ecosystems.

• Human-in-the-Loop (HITL) Approvals: Durable, auditable approval workflows that ensure agents pause for operator sign-off before taking consequential actions.
• Agent Workspace Filesystem: Dedicated scratch space and knowledge mounts for agents to process, filter, and inspect large data files without context bloat.
• Agent Procedural Memory: Intelligent self-learning allowing agents to remember efficient tool-calling recipes from past interactions.
• In-Product Configuration Assistant: A conversational admin assistant that helps manage RAG sources and settings, with full administrator review and approval.
• Next-Level Retrieval Quality (Chunk Enrichment & Question Index): Automated semantic enrichment (titles, keywords, contextual prefixes) and a dedicated answerable-questions vector index for unmatched search precision.
• Expanded Enterprise Connectivity: Role-Based Access Control (RBAC) for knowledge bases, authenticated web crawling, and automated attachment indexing.
• Enterprise Security & Network Guardrails: Granular token verification, default-deny outbound network policies, and role-based admin controls.

Technical Summary

Human-in-the-Loop (HITL) Approval Framework

AI Core 1.3.0 introduces a robust governance layer for autonomous agents. Agents can now execute complex tasks while ensuring that high-impact actions (such as updating database records or triggering external workflows) require explicit human approval.

• Durable Approval Requests: Approvals persist across system restarts and survive session interruptions.
  Multi-Channel Decision Making: Administrators and operators can review and approve actions via interactive cards in the UI or directly within API channels.
• Scheduled Task Governance: Automated background agent triggers pause safely at gated actions, creating auditable records instead of failing silently.

Use Case: An IT automation agent identifies an outdated user privilege. Before revoking access or modifying directory settings, the agent generates an approval card in the management interface. The system administrator reviews the proposed change and approves it with a single click before the agent proceeds.

Agent Workspace & File Processing

Every file-enabled agent now has access to an isolated, per-session Scratch Workspace and an optional persistent knowledge base mount.

• Offloading Large Tool Outputs: Instead of flooding the chat context with massive API payloads or document text, agents offload raw data to their workspace and use built-in tools (grep, jq, cat) to inspect only what they need.
• Interactive Tooling: Equipped with specialized file manipulation tools, agents can process complex JSON outputs, aggregate multi-page results, and manage temporary work files.

Use Case: An agent queries a system logging API that returns 5,000 lines of data. Rather than exceeding model limits or slowing down responses, the agent saves the raw log to its workspace, uses built-in filtering tools to extract the relevant error codes, and returns a concise summary to the user.

Agent Procedural Memory

Agents now learn from experience. AI Core 1.3.0 introduces a background curation process that analyzes successful task executions and distils working tool-call patterns into reusable procedures.

• Faster & More Reliable Execution: On future requests, agents recall proven tool-calling strategies, reducing trial-and-error reasoning.
• Enterprise Guardrails: Memory writes are continuously validated and filtered to prevent instruction drift or unauthorized prompt modifications.

Use Case: An agent learns the exact sequence of parameters needed to query a complex internal inventory API. On subsequent user requests for inventory checks, the agent instantly recalls the exact procedure, dramatically speeding up response times.

Interactive Configuration Assistant

To simplify platform administration, AI Core 1.3.0 features Apuri, an embedded conversational configuration assistant within the Admin UI.

• Guided Administration: Helps admins configure RAG data sources, document-processing settings, and indexing parameters through natural conversation.
• Strict Admin Control: The assistant proposes changes as structured patches, which take effect only after explicit administrator review and approval.
• Security First: Built-in guardrails ensure sensitive secrets, API keys, and passwords are never exposed or handled by the assistant.

Use Case: A system admin wants to set up a new web crawler for internal documentation. Instead of manually navigating multiple settings menus, they tell the Configuration Assistant what site to index. The assistant prepares the configuration parameters and presents a clean summary card for the admin to approve.

Advanced Retrieval Quality (Chunk Metadata & Question Index)

AI Core 1.3.0 significantly elevates RAG precision through automated semantic enrichment during content ingestion:

• Semantic Chunk Enrichment: Automated background processors generate document titles, contextual prefixes, and key concepts that are prepended to text chunks, maximizing vector recall and accuracy.
• Answerable Questions Index: The system automatically extracts candidate questions answered by each chunk and indexes them into a secondary vector index, ensuring direct hits when users ask similar questions.
• Rate-Controlled Ingestion: Smart pacing protects self-hosted or rate-limited model backends during large reindexing jobs.

Use Case: A user asks, "How do I submit an expense report on mobile?" The system matches the query not only against raw policy text, but also against the specialized Answerable Questions Index, delivering an exact, highly relevant answer on the first attempt.

Enterprise Data Sources & Access Control

• Knowledge Base Role-Based Access Control (RBAC): Native support for granular access permissions ensuring users only retrieve information they are explicitly authorized to view.
• Authenticated Web Crawling: The crawler can authenticate against protected internal portals using secure login flows, indexing gated internal knowledge safely.
• Attachment Indexing: Automated extraction and indexing of file attachments (PDFs, DOCX, CSV, Excel workbooks) connected to service records and portal items.

Security, Resilience & Operational Maturity

• End-User Token Verification: Optional verification of end-user identity tokens across RAG and agent endpoints.
• Outbound Network Policy: Strict default-deny outbound network rules for custom scripts and tools to prevent unauthorized internal network access or data exfiltration.
• Audit History for Security Filters: A dedicated audit view logs content blocked during data ingestion, categorized by source and document.
• Test Suites & CSV Export: Independent question test suites for quality assurance, complete with CSV export/import capabilities.

Improvements & Bug Fixes