Enterprise RAG developers now see a new anchor‑detection strategy that stitches keyword, embedding, and one final LLM step into a lean pipeline. The paper, part of a four‑brick series, lays out how Anchor detection works on two structured tables—line_df and toc_df—so that a single GPT call can rank candidate sections with audit‑ready reasoning.

First stage runs keyword matching on both tables, a zero‑cost baseline that flags lines containing the user’s query terms. Parallel to it, optional Embeddings capture semantic drift, catching matches that miss surface vocabulary. When the keyword hit is clean, the system can skip embeddings, saving microseconds per query while still covering edge cases within tight latency budgets.

Stage two aggregates the hits into structural units, matching keyword and embedding results to sections in the TOC or falling back to page chunks. This consolidation lets the final arbiter see each candidate’s context, its structural attachment, and any overlapping signals in one shot, dramatically reducing the number of LLM calls needed for a document‑level answer within a single inference.

Finally, a single LLM call performs the TOC reasoning and ranking, producing a JSON that lists chosen sections with human‑readable justifications. The approach keeps retrieval logic deterministic, limits expensive inference to the last step, and delivers explanations auditors can audit months later. The method proves that careful signal layering can slash LLM usage without sacrificing accuracy in production deployments.