Retrieval Is Filtering, Not Search: A Mental Model for Enterprise RAG

Retrieval for enterprise RAG systems should be reframed as a filtering problem on structured tables, not a simple search problem. This mental model uses a `line_df` for the document's text and a `toc_df` for its structural map, mimicking how human experts find information. This approach allows the system to programmatically perform tasks like keyword co-occurrence detection and joining content with the table of contents. The model also introduces the concept of a small "anchor" for matching and a larger "context" that is passed to the language model, arguing that no single retrieval method is sufficient for all types of queries.