Controlled Hybrid Retrieval Architecture for Structured Behavioural Case Knowledge (PROTEX System Design)

This work presents PROTEX, a controlled hybrid retrieval architecture designed for the representation and analysis of structured behavioural case knowledge in AI-supported systems. The framework addresses methodological limitations of naive Retrieval-Augmented Generation (RAG) approaches when applied to complex, narrative-driven domains such as criminal behavioural analysis.

PROTEX combines semantic retrieval with deterministic metadata filtering, explicit query routing, and confidence signalling. Cases are ingested as epistemic objects composed of structured profiles and contextually anchored narrative fragments, stored in separate index namespaces. Query processing prioritises intent classification and knowledge-layer selection before semantic similarity is applied, thereby constraining inference and reducing the risk of over-interpretation.

The system is not intended as an operational or predictive tool. Instead, it provides an architectural and methodological blueprint for responsible retrieval-based analysis under conditions of epistemic uncertainty. While developed in the context of investigative psychology, the design principles outlined here are transferable to other domains involving qualitative, narrative case material.