A Domain-Agnostic Agentic Architecture for Structured Extraction of Engineering Knowledge

LLM-based retrieval systems can execute queries and produce fluent answers, yet they rarely preserve a traceable record of what evidence was retrieved, how it was validated, or what reasoning led to synthesis. The lack of provenance and auditability limits debugging, compliance verification, and operational trust, particularly in engineering applications where correctness depends on interpreting structured specifications within revision-controlled documentation. This work presents the Relational Control Plane (RCP), a SQL-backed agentic architecture that persists orchestration state as queryable relational records. RCP enforces a six-stage verify-then-summarize control loop in which synthesis is permitted only after retrieved evidence satisfies validation constraints, so that unmet conditions surface as explicit, auditable failures rather than silent errors. Domain behavior is defined by reusable Strategy and Function libraries, enabling adaptation through declarative templates without modifying the orchestration logic. Evaluation across two industrial applications shows substantial gains: answer correctness increases from 45% to 85% with hallucination reduced from 55% to 5% in one case study; and from 23% to 81% with hallucination reduced from 47% to 11% in another, despite using a lightweight language model. Results show that reliable agentic AI systems emerge from architectural persistence, structured validation, and traceable workflows rather than model scale.