Proxy-Pointer RAG introduces a 'Graphability Index' to reduce LLM token costs during Knowledge Graph ingestion from enterprise documents. Instead of sending entire documents to an LLM for NER and relation extraction, the approach classifies document sections by their predicted relational density before extraction begins. Sections rated Low or Very Low are bypassed entirely. Validated against three real-world corporate credit agreements (Emerson, AT&T, Texas Roadhouse), the index stabilized quickly and achieved payload reductions of 16%, 34%, and 38% respectively, with zero mismatches by the third document. The method leverages structural predictability of legal documents — treating them as semantic trees rather than flat text — and is available as an open-source MIT-licensed Python package.
