Retail supply‑chain operations consist of tightly coupled decision modules that must be re‑configured as business requirements evolve. Large language models (LLMs) enable a natural‑language interface for such re‑configuration, yet most existing approaches target a single optimization model and struggle to scale to heterogeneous decision pipelines. A single requirement can admit multiple intervention routes, each producing distinct downstream effects.
We formalize requirement‑driven adaptation as the joint selection of an intervention route and an admissible module‑level change, and introduce a graph‑constrained agentic framework. In this framework, domain agents expose admissible reformulation interfaces, while a central processor searches bounded paths on a predefined graph. Candidate solutions are validated and compared using downstream key‑performance indicators (KPIs), ensuring the chosen path yields the best operational outcome.
In collaboration with a major retail partner, we gathered 100 warehouse‑level requirements from practitioner interviews and evaluated three base LLMs—GPT, Qwen, and DeepSeek. Compared with direct LLM‑only reformulation, our framework consistently improves correctness and end‑to‑end success across all models, raising success rates from 72%‑76% to 79%‑83%.
Core technical contributions
- Modeling requirement adaptation as a graph‑based path‑search problem;
- Leveraging domain agents to provide controllable, module‑level reformulation interfaces;
- Closing the loop with downstream KPI evaluation of candidate interventions.
Review