AI Knowledge Assistant for Operations & Decision Support

Operational teams had to manually search multiple systems and frequently consult subject-matter experts before determining the appropriate course of action when an incident occurred.

This resulted in long investigation times, dependency on experienced personnel, repetitive troubleshooting activities, difficulty finding similar historical incidents, inconsistent resolution procedures, slow access to operational data, and delays in decision-making.

The organization needed an AI assistant capable of combining enterprise knowledge with operational information while maintaining appropriate security and access controls.

Rootstack implemented an AI-powered operational knowledge and decision-support platform using Claude, RAG, a vector database, MCP, Python, PostgreSQL, REST APIs, Kafka, Docker, OpenTelemetry, and OAuth 2.0/OpenID Connect.

The architecture connected enterprise sources to data connectors, processing and indexing services, a vector database, a permission layer, retrieval capabilities, and controlled operational tools and APIs. This allowed the assistant to combine technical documentation, historical knowledge, operational data, and AI reasoning in a single response.

The RAG architecture indexed equipment manuals, SOPs, maintenance procedures, safety documentation, engineering documentation, internal knowledge bases, and historical incident reports. Through semantic search, users could find relevant information based on meaning rather than relying exclusively on exact keywords.

For incident investigation, the assistant analyzed the operator's description, identified relevant technical concepts, retrieved applicable procedures, and searched for historical incidents with similar characteristics. It could highlight previous symptoms, root causes, successful corrective actions, and recurring equipment problems, making previously isolated organizational knowledge available during new investigations.

The platform also integrated with authorized operational systems through REST APIs and MCP-based tools. Depending on user permissions, it could retrieve equipment status, current operational metrics, service tickets, maintenance history, production information, and incident status. The AI layer did not have unrestricted access to operational systems; available tools and APIs were explicitly defined and controlled.

For root-cause analysis, the assistant could correlate current symptoms with equipment information, historical incidents, maintenance records, and applicable SOPs. It clearly distinguished between information retrieved from systems, information contained in documentation, and AI-generated analysis. This provided teams with decision support while maintaining transparency around the origin of each recommendation.

When relevant documentation was available, the assistant could also provide step-by-step troubleshooting guidance, including initial verification, safety considerations, equipment checks, applicable parameters, recommended next steps, escalation conditions, and supporting documentation.

Source citations and auditability were incorporated into responses, allowing users to verify which procedures, historical incidents, and operational data supported a recommendation. Permission controls ensured that technical documentation, maintenance records, operational metrics, incident information, and restricted business data were only accessible to authorized users.

The project was completed in 6 months and achieved an estimated 50% reduction in incident investigation time and 65% reduction in technical information search time. Between 70% and 80% of operational questions were automated, while dependency on subject-matter experts for routine issues was reduced by an estimated 45%. More than 10,000 historical cases became searchable, over 95% of responses included supporting sources, and the platform integrated with 4+ operational systems while maintaining 24/7 availability.