Anchor Detection for RAG: Parallel Detectors, Then One LLM Call at the End

A three-stage pipeline is proposed for improving document retrieval in enterprise RAG systems. This method first runs keyword and embedding detectors in parallel on structured tables representing the document's text lines and table of contents. The resulting hits are then aggregated into structural units, such as document sections. In the final stage, a single LLM call acts as an arbiter to rank these aggregated candidates and provide its reasoning. This architecture efficiently combines the speed of deterministic search methods with the advanced reasoning capabilities of a large language model.