Vanč Levstik

This talk addresses the challenge of "semantic blindness" in Large Language Models (LLMs) when dealing with vast, complex, and inconsistently named datasets, such as sensor names in large-scale industrial environments. The core thesis is that LLMs struggle with sheer scale and naming ambiguity, leading to errors and hallucinations. The proposed solution involves a hybrid approach that leverages the LLM for planning and decision-making while offloading structured data processing, retrieval, and set operations to deterministic code.