PARSING transforms PDFs into DataFrames; line_df holds every line, toc_df maps sections. Retrieval, the third brick in enterprise RAG, reframes itself as filtering tables instead of free‑text search. By targeting small toc_df rows and expanding line_df context, the system mirrors a seasoned HR worker’s Ctrl+F routine, but with algorithmic precision in 45 seconds and 10% faster than manual lookup today.

Unlike vector similarity, filtering leverages structured columns: regex on line_df, section titles on toc_df, and lightweight LLM passes over the 20‑row toc_df. Joining the two tables via section_id lets the model first narrow to a chapter, then scan only relevant lines. This dual‑granularity approach cuts token costs and boosts relevance for long contracts in enterprise legal teams worldwide daily operations.

By codifying the expert’s Ctrl+F workflow, the retrieval system eliminates blind spots: OCR converts scanned pages, automated TOC joins replace manual clicks, and multi‑keyword detection surfaces hidden clauses. The approach delivers answers in seconds, scales to thousands of pages, and aligns with existing document infrastructures, proving that filtering, not search, is the right mental model for enterprise RAG in practice.