Dynamic electronic data generation through human-guided multi-agent collaboration

Abstract

A multi-agent framework for generating synthetic data includes a human-in-the-loop agent that receives schema definitions, optional seed data, and domain-specific context, and transforms such inputs into a structured schema JSON; and an orchestrator agent that decomposes electronic data generation into subtasks and invokes autonomous agents including a topic modelling agent, an input enrichment agent, a variability injection agent, and a data generation agent. The topic modelling agent employs a large language model with an extended context window to extract hierarchical categories and salient keywords. The input enrichment agent expands categories using retrieval, keyword expansion, and iterative refinement, while the variability injection agent produces structured variation artifacts simulating real-world irregularities. The data generation agent synthesizes multimodal datasets that are persisted by an artifact serializer/loader and validated by a validation tool for coverage, schema conformity, and distributional balance. The synthetic dataset is suitable for training and validating machine learning models.