Proxy-Pointer RAG is a novel architecture that addresses entity and relationship reconciliation challenges in large enterprise knowledge graphs. Instead of relying on expensive global graph searches or traditional RAG chunking, it uses a two-step pipeline: an LLM-generated entity profile drives multi-track vector queries, and retrieved chunks serve as structural pointers to load full document sections. This gives a reconciler LLM complete semantic context to identify entity aliases, existing relationships, and graph neighborhoods. Tested against AMD 10-K filings, the approach successfully resolved entities like 'Sony' to 'Sony Interactive Entertainment, Inc.', identified new nodes like 'Pensando Systems' and localized their graph neighborhoods, and prevented duplicate edge creation — all at a fraction of the cost of full graph traversal.
