Problem
Books in PDF form are hard for a language model to use: the text comes out unstructured, chapters are lost and the whole book is too long for one conversation.
Value
A long book becomes a set of chapters an agent can read in one go, without sending the text to an outside service.
Role
Architecture, choice of conversion engines, chapter-boundary tests and the decision records. Agents wrote most of the code.
- Set up a layered architecture in which the conversion engines are swappable and compared on the same documents.
- Used a local language model to classify documents and to recover the table of contents.
- Built a regression suite for chapter boundaries so a changed engine cannot silently split a chapter wrong.
- Recorded four architecture decisions.
Limits
- The converted books are reading material under copyright and are never shown. Only the tool is described.
Sources
- Four architecture decision records
- Private version that runs on a local machine