EPRAG: An epistemic policy framework for action selection in multi-source enterprise retrieval-augmented generation

Background and Objective: Enterprise retrieval-augmented generation (RAG) systems must select an appropriate action when retrieved sources differ in freshness, authority, access, or evidential compatibility. Methods: We propose EPRAG (Epistemic Policy RAG), an executable framework that separates evidence-state diagnosis from response selection through a six-state taxonomy and a deterministic five-action policy. Counterfactual answerability is implemented as a practical LLM/NLI-based approximation, while provenance is treated primarily as an auditability feature. Evaluation: EPRAG is evaluated on Spider2-derived (n=504) and BIRD-derived (n=720) benchmark adaptations, a manual adjudication study (n=400), a single-site enterprise helpdesk cohort (n=400), and a taxonomy-blind public audit (n=200). A structured GPT-5 judge is also evaluated on BIRD (n=240) and helpdesk (n=100) subsets using the same inputs and five-action space. Results: EPRAG achieves 91.7%–96.8% action accuracy on the public benchmarks and 76.0% against operationally derived helpdesk labels. In the taxonomy-blind audit, annotator agreement was 93.5% (κ=0.916), and EPRAG achieved 85.0% agreement with the trusted action (macro-F1 =0.849). EPRAG also outperformed the structured GPT-5 judge on BIRD (97.9% vs. 40.8%) and helpdesk (80.0% vs. 23.0%). Conclusions: EPRAG provides a bounded, auditable decision-policy baseline for multi-source enterprise RAG. The helpdesk evidence is single-site and label-dependent, the GPT-5 comparison is not an upper bound, and broader validation remains necessary.