Why Book Translation Needs a Second Pass
Most LLM translation demos stop after a single generation pass. That is enough to preserve rough meaning, but not enough to preserve rhythm, tone, and narrative continuity across long chapters.
Book Translator uses a two-step workflow:
- Draft translation for semantic fidelity.
- Self-reflection pass for style, flow, and readability.
That extra pass matters because long-form translation quality breaks down in subtle ways. Literal phrasing accumulates. Transitional sentences become stiff. Paragraph rhythm starts sounding machine-generated even when each sentence is technically correct.
The project treats translation less like one-shot prompting and more like an editorial pipeline. It runs locally with Ollama, which keeps sensitive manuscripts off third-party APIs while still giving you a repeatable CLI workflow.
Key design choices:
- chunking for long documents
- local-first inference via Ollama
- explicit self-reflection stage for refinement
- CLI-first workflow for repeatable runs
If you are building long-form AI writing systems, the main lesson is simple: generation quality is often a workflow problem, not just a model problem.
Top comments (0)