Why summaries make routing cheap

The agent has to decide at every node: descend into a child, or stop and retrieve? The naive way is to paste the full section text into the prompt and ask. This works, but the bill adds up fast on a long paper. Summaries change the math.

Token cost per navigation decision

Full section text versus summary-based routing. Same decision, very different bill.

The principle is simple: pay once to summarize at ingest time, reuse the summary on every query. Full section text is only loaded when the agent commits to retrieval, not during navigation. Routing becomes a cheap constant-cost operation no matter how long the section is.

Quiz: Quiz

Loading practice…