Type-Aware Retrieval-Augmented Generation with Dependency Closure for Solver-Executable Industrial Optimization Modeling

A research team including Y. Zhong, R. Huang, and M. Wang has published a breakthrough paper introducing a novel type-aware retrieval-augmented generation (RAG) method specifically designed for industrial optimization modeling. The system addresses a critical limitation of current large language models (LLMs) in engineering contexts: they frequently generate non-compilable, non-executable code due to missing declarations, type inconsistencies, and incomplete dependency contexts. This new approach promises to reliably translate natural-language requirements—like 'optimize battery production for demand response'—directly into solver-executable mathematical models, a task where conventional RAG and LLM approaches have consistently failed.

The technical innovation lies in moving beyond unstructured text indexing. The method parses heterogeneous sources like academic papers and existing solver code into typed units (variables, constraints, objectives) and encodes their mathematical relationships into a domain-specific knowledge graph. When given an instruction, it performs hybrid retrieval and computes a 'minimal dependency-closed context'—the smallest set of typed symbols required for a complete, executable model—via graph propagation. Validated on two complex industrial cases (demand response in battery production and flexible job shop scheduling), the method achieved 100% compilability and reached known optimal solutions, while all baseline methods failed entirely. This demonstrates a path to deploy LLMs in high-stakes optimization tasks like supply chain logistics and energy grid management, where executable accuracy is non-negotiable.