Proxy-Pointer RAG: Eliminating Wasteful Entity & Relations Extraction in Knowledge Graphs

A novel approach called Proxy-Pointer RAG is proposed to optimize the creation of Knowledge Graphs from large, structured documents like legal contracts. This method reduces the high costs of using Large Language Models (LLMs) for entity and relation extraction by leveraging the documents' predictable structure. A 'Graphability Index' is created to score document sections based on their predicted relational density, allowing the system to selectively process only high-value content. High-yield sections are sent to an LLM for deep analysis, while low-yield sections are bypassed, drastically cutting token consumption and cost without sacrificing the integrity of the resulting graph.