This article, the second in a series on adaptive parsing for enterprise RAG systems, details how Large Language Models (LLMs) act as a final defense against parsing errors. It showcases two real-world escalations: parsing flat tables with Azure and extracting information from figures using a vision LLM.

The system prioritizes cost-effectiveness by starting with cheaper parsers like Py Mu PDF. Only when deterministic checks or LLM self-evaluation flags potential issues does it escalate to more powerful, expensive parsers. This "loop engineering" ensures efficiency while maintaining accuracy. The companion code demonstrates parsing a table from the "Attention" paper with Py Mu PDF and then escalating to Azure Layout when the initial parse is insufficient.

Another case involves extracting data from a figure, where a vision LLM is employed. The article highlights the importance of LLM generation-time checks, specifically self-evaluation and groundedness, in catching parsing failures that simpler methods miss. These checks trigger the escalation to deeper parsers, ensuring the LLM answers from accurate information, even from complex document structures. The Transformer architecture paper is used as a case study.