Agentic GraphRAG for sustainability requirements : quality, grounding, and energy trade-offs

Abstract

Generative AI may speed up requirements drafting, however, ungrounded large language models often produce sustainability requirements that lack measurability, traceability, and credible grounding. Thus, it may hallucinate thresholds or sources. This thesis presents EcoAnchor, a Design Science artifact that combines GraphRAG retrieval over a curated sustainability knowledge graph (i.e., SusAF, GSF SCI) with a Retriever–Drafter–Critic agents workflow. On 15 matched triggers from the public RivCoDelivery specification, six blind expert reviewers compared EcoAnchor with a human practitioner baseline (n = 6 authors) and standard single-shot GenAI on measurability, clarity, traceability, relevance, hallucination, and overall preference. Energy was measured with CodeCarbon in two phases on fixed hardware. EcoAnchor led expert ratings. In the primary energy campaign, Phase 1 EcoAnchor used about 3.6× the energy of standard GenAI (3.55 Wh vs. 1.0 Wh); Phase 2 eco-routing reduced this to 3.01 Wh (about 15% savings) while preserving 100% citation validity. Expert quality claims apply to Phase 1 outputs; Phase 2 quality was checked automatically. For small-batch, quality-critical elicitation the trade-off favours EcoAnchor; high-volume generation still needs further routing optimisation.