Year: 2025
Type: article
Abstract: Enterprises do not choose among prompt engineering, retrieval-augmented generation (RAG), fine-tuning, and hybrids in the abstract. They choose deployable combinations of model, context, behaviour control, orchestration, and serving under task-specific constraints. This paper presents E-LADF 2.0, a design-science decision artefact for configuration-level feasibility, value functions, and robustness. It begins with a registered acceptance test, diagnoses the locked baseline, freezes a causally plausible configuration shortlist, applies noncompensatory feasibility gates, and then combines anchored 0-to-100 value functions with analytic hierarchy process weights. A bounded uncertainty analysis perturbs criterion weights and performance values. A configuration is recommended only when its worst-case advantage exceeds a predeclared practical difference; otherwise the framework returns ‘not separated’ and prescribes a targeted pilot. Five constructed scenarios directly supply every partial value, criterion weight, and uncertainty input used in the arithmetic and, by design, exercise prompt, RAG, PEFT, hybrid, and abstention outcomes. They demonstrate execution, traceability, and edge-case handling, not empirical effectiveness. The contribution is a configuration-level procedure that separates constraints from preferences, replaces universal archetype scores with case-specific partial values, and treats uncertainty as a decision outcome.
Author: Karan Nisar
Language: English
Field: Decision Sciences
Domain: Social Sciences
Open Access status: hybrid
OpenAlex ID: W7203945641