Fable
Excerpt cards and full-chapter bookshelf
The North Street library opened a new inquiry window. Readers only had to state one question, and the clerk had to find the relevant materials. The young clerk cut every book into small scrolls by meaning: one scroll on pricing, one on returns, one on warranty. The small scrolls read smoothly and were easy to place on the counter.
Another old scribe hung a thin string after each excerpt, connecting back to the original chapter. When a reader asked for detail, he first handed over the small excerpt; when the reader needed to make a decision, he took down the whole chapter so the reader could see the surrounding constraints, exceptions, and definitions.
At first, the young scribe was more popular. His answers were short, and customers didn't need to flip through thick books. When someone asked how many days the warranty lasted, he just needed one sentence; when someone asked where a certain button was, he handed over a small scroll of instructions, and the queue at the window quickly grew shorter.
Later a procurement manager came and asked whether a service could be used overseas. The small scroll said it could be activated, and the young clerk answered accordingly. Just as the manager was about to sign, the old clerk followed the string back to the full chapter and found that the next page said only customers that had passed data approval could use it. Once the small scroll left its parent chapter, the answer nearly went astray.
The old scribe's method also had problems. Every time he brought the whole chapter, the reader would be surrounded by irrelevant paragraphs. The question asked about only one field, but a dozen pages of the whole chapter were stuffed in, and the window slowed down again. Readers complained that they just wanted to confirm one sentence.
The director finally combined both approaches in the protocol. The index uses small excerpts for precise hits; when the answer involves commitments, restrictions, compliance, or multi-step processes, it must bring back the parent section. As for how to split the small excerpts, still divide by meaning as much as possible, so the hit point itself isn't incomplete.
The library grew much quieter. The scribes no longer conflated "cutting naturally" with "bringing back context." One decides the cut; the other decides how far up to pull after a hit. The closer the question is to a real decision, the more valuable this distinction becomes. The curator also had the scribes record every wrong answer: was the small scroll itself cut badly, or was it found correctly but not pulled back to the upper level. After a month of records, fixes no longer relied on arguments. Later, for pricing, liability, and compliance questions, the window by default presented the parent level as well; for term definitions and button locations, it handed over only the small scroll. Relationships were kept too.