Local Book Translation Pipeline Automates PDF-to-EPUB

This is less a model demo than a systems demo: the real trick is orchestration, not raw translation quality. The “Super Bible” plus editor pass is a sensible way to reduce context drift across book-length jobs without pretending a single prompt can hold the whole novel in memory.

• –Marker preserves structure and images up front, which matters more than people think when the final deliverable is an EPUB, not just a text dump
• –Sampling excerpts across the whole book to build a shared glossary is a lightweight contextual-RAG pattern for long-form translation
• –A second-pass editor model can improve literary flow, but it also adds another place for semantic drift, so quality checks still matter
• –The folder-watcher automation is the most productizable piece here: drop in a PDF, wait, get output and runtime logs
• –The obvious ceiling is evaluation, not generation; consistency of names, pronouns, and tone across chapters will decide whether this scales beyond “impressively good”