Large language models can turn policy, news and industry text into hundreds of traceable parameters for supply‑chain simulation; applied to India’s lithium reserves, the method transformed 972 documents into 668 validated inputs to improve model realism and scenario specification.

Main Finding

Integrating LLMs into a Retrieval-Augmented Generation (RAG) pipeline provides a domain-agnostic method to convert unstructured narrative sources (policy texts, news, industry reports) into traceable, simulation-ready inputs. This LLM-Integrated Simulation Data Framework improves realism and contextual relevance of multi-echelon supply chain simulations in data-sparse settings, as shown by a lithium carbonate strategic-reserve case for India (972 documents → 668 validated insights).

Key Points

• Problem: Multi-echelon supply chain simulation frequently lacks timely, contextual data for realistic parameterisation and scenario design.
• Solution: A systematic framework that uses RAG-powered LLMs to extract, structure, and validate intelligence from unstructured narratives and link it to structured datasets.
• Outputs: Traceable model inputs for agent behavioural rules, numerical parameters, and scenario specifications.
• Domain-agnostic: Designed to be applied across sectors, with a demonstrated case in geopolitically sensitive, resource-constrained supply chains (lithium reserves).
• Case metrics: Processed 972 documents and produced 668 validated insights used to inform simulation agents and scenarios.
• Methodological contribution: Provides an automated collaborative interface between Operations Research simulation modelling and generative AI, enabling scalable incorporation of qualitative intelligence into empirical grounding.
• Validation: Emphasises traceability and validation steps to mitigate hallucination and ensure insights are defensible for policy analysis.

Data & Methods

• Data sources: Policy documents, news archives, industry/technical reports and likely other unstructured narrative sources relevant to the supply chain context.
• Pipeline architecture:
  • Retrieval: Select relevant documents and passages using information retrieval techniques (to create a contextual corpus).
  • Augmented generation: LLMs synthesize and extract structured insights from retrieved passages (RAG approach combines retrieval with generation to reduce hallucination).
  • Structuring: Convert extracted narrative intelligence into explicit outputs suitable for simulation—agent rules, numeric parameters (e.g., capacities, lead times, adoption rates), scenario triggers and timelines.
  • Validation & provenance: Apply validation checks and attach traceable provenance metadata to each insight to support auditability and model reproducibility.
• Case application: India’s strategic lithium carbonate reserve planning—processing 972 unstructured documents to derive 668 validated and traceable simulation inputs that informed agent behaviours and scenario specifications.
• Evaluation: Demonstrated feasibility and utility by improving contextual realism in a multi-echelon simulation where conventional structured data were limited.

Implications for AI Economics

• Improved empirical grounding: LLM-augmented pipelines allow economic and operational models to combine quantitative baselines with rich qualitative intelligence (policy intents, geopolitical signals, market narratives), improving the realism of counterfactuals and policy simulations.
• Faster, cheaper scenario generation: Automating extraction of contextual knowledge reduces time and cost of assembling simulation inputs, enabling more rapid policy experimentation and sensitivity analysis.
• Better policy relevance: Traceable, validated narrative inputs let modelled outcomes more credibly reflect policy choices and external narratives—useful for strategic resource planning and regulatory impact assessment.
• Market and strategic effects: More context-aware simulations can influence investment, strategic stockpiling, and supply-chain risk assessments—potentially affecting market behaviour and policy decisions where resources are geopolitically sensitive.
• Risks and limitations:
  • Hallucination and model risk: LLM outputs require rigorous validation, provenance tagging, and human-in-the-loop review to avoid introducing spurious or biased inputs into economic analyses.
  • Dependence on source coverage: Framework utility still depends on availability and representativeness of narrative sources; missing perspectives can bias scenarios.
  • Reproducibility & auditability: While provenance aids reproducibility, maintaining long-term audits requires preserving retrieval contexts, prompt specifications, and model versions.
  • Ethical/geopolitical sensitivity: Care needed when automating insights from politically sensitive materials; outputs may have real-world strategic consequences.
• Research and policy directions: Standardising validation protocols, integrating uncertainty quantification for narrative-derived parameters, and developing governance frameworks for LLM-assisted simulation inputs are critical next steps to safely scale the approach in AI economics and policy modelling.