In 2022, I built my first AI agent with what I now think of as “pure RAG.” Documents were chopped into small, disconnected pieces and retrieval returned whichever chunks looked most similar to the question.
It was like looking at a map through a straw. When a sentence was cut off, our first fix was simply to retrieve more chunks and hope the missing continuation appeared somewhere in the larger pile.
Humans do not usually find a fact by reading an entire 400-page book. We navigate: table of contents, chapter summary, then the relevant page. A retrieval system can follow the same shape.
A two-stage navigation system
The first stage is a librarian: a fast model that identifies the right area without deeply reading everything.
The second stage uses breadcrumbs. Every chunk carries useful structure—chapter, jurisdiction, page, product, customer, or another domain-specific hierarchy. A stronger research model can then reason over the right material instead of a random neighbourhood of similar text.
A self-healing loop
Retrieval should not be a single irreversible call. If the research model detects that its context is incomplete, it can send a specific request back to the librarian for the missing section.
That feedback loop separates finding from reasoning. Teams can build it directly or use managed search, parsing, and orchestration tools that support parent-child retrieval and agent loops.
The important change is conceptual: retrieval stops being a search for isolated fragments and becomes a navigation strategy.