DynaContext Improves LLM Parameter Extraction with Dynamic Context.

A new framework called DynaContext addresses the limitations of static prompt optimization in large language models (LLMs) when dealing with varied data. Traditional methods use a single instruction, which struggles with heterogeneous information like extracting parameters from diverse electronic component descriptions, where different items require unique fields, constraints, and evidence. DynaContext overcomes this by dynamically generating item-specific prompts. It integrates an offline-optimized extraction core with real-time contextual adaptation and a self-improvement mechanism. Each input is routed through internal, external, or fallback evidence paths, and a tailored prompt is composed using the core, schema, evidence, and validated demonstrations. Outputs are rigorously checked by deterministic validation and an LLM judge, with uncertain cases sent for human review, and only human-verified corrections are added to the demonstration memory. Evaluations show substantial improvements: on a single-category benchmark, accuracy rose from 86.6% to 98.6% with DynaContext. For 850 heterogeneous parameter facts, the framework increased average field-level F1 from 51.8% for an unoptimized control to 71.0% with the full configuration, outperforming existing static-prompting pipelines by 17.3 F1 points.