Domestic publishers have been betting heavily on translating foreign bestsellers in recent years, but this approach is reaching its limits. Haruki Murakami’s latest novel, Xia Fan, was first published in Japanese on July 3, and by July 6, a simplified Chinese translation based on the Claude version was already available on Z-Lib. Feedback on Xiaohongshu generally indicated that this translation was better than those by his long-time official translator, Lin Shaohua. The translator even open-sourced the entire long-form translation project on GitHub:

BigDawnGhost/wenyi

Source: https://github.com/BigDawnGhost/wenyi

WenYi

A tool focused on translating multi-language EPUB, FB2, or TXT novels into Chinese, preserving the original EPUB layout, images, table of contents, and hyperlinks as much as possible.

The main entry point for daily use is a single command: translate. It handles pre-scanning, analysis, translation, optional polishing, chapter-end review, punctuation normalization, and EPUB export; it can be resumed after interruption.

Quick Start

uv sync
export DEEPSEEK_API_KEY=sk-...
uv run trans-novel translate book.epub

After translation completes, a Chinese EPUB is generated by default in the source file directory. Running state, chapter JSON, glossary, and reports are stored in the state/ directory.

Resume after interruption:

uv run trans-novel resume book.epub

Check progress:

uv run trans-novel status book.epub

Only re-export EPUB:

uv run trans-novel tools assemble book.epub

Input and Output

• Input: EPUB, FB2, TXT.
• Default output: Chinese EPUB.
• EPUB input: the translated text is backfilled into the original XHTML template, preserving the original book style, images, table of contents, and anchors as much as possible.
• TXT input: a new EPUB is generated.
• Use --format txt when plain text output is required.

Examples:

uv run trans-novel translate book.epub
uv run trans-novel translate book.epub --format txt
uv run trans-novel translate book.epub --chapter 3

Common Flags

uv run trans-novel translate book.epub --polish
uv run trans-novel translate book.epub --no-polish
uv run trans-novel translate book.epub --qa
uv run trans-novel translate book.epub --no-qa

--polish/--no-polish override the pipeline.polish setting in config.yaml. The current repository’s config.yaml has polish: true, so the default without the flag is to polish; the code-level default is false, which only applies when the config file does not have that field.

Polishing adds one strong-tier LLM request per translation batch, which may improve quality but significantly increases time and cost. Already translated batches are skipped when resuming; toggling polishing later will not automatically retranslate old batches.

Configuration

Main settings are in config.yaml:

• language.source: auto lets the model detect the source language, or you can hardcode a language code (e.g., ja, en, ko, ru, de).
• llm.tiers: configure the strong, cheap, and fast model tiers.
• pipeline.review: chapter-end review.
• pipeline.autofix_severe: automatically re-translate severe issues and adopt passing results.
• pipeline.polish: polish Chinese after translation.
• pipeline.backtranslate_sample: back-translation sampling ratio; 0 disables it.
• pipeline.consistency_qa: cross-chapter consistency scan.
• pipeline.book_understanding: pre-scan the entire book before translation, generating a full book overview and per-chapter summaries.
• pipeline.rolling_context_segments: number of previous translated segments to carry forward for each batch.
• segment.max_chars_per_batch: size of each translation batch.
• segment.max_chars_per_segment: splitting threshold for overly long paragraphs.

For offline testing or debugging, set llm.provider to fake to avoid making network requests.

Workflow

The default continuous workflow is roughly:

Read input
→ Parse chapters, body paragraphs, and EPUB TOC
→ Model detects source language (or uses configured language)
→ Pre-scan entire book, generate full overview and per-chapter summaries
→ Analyze sample chapters, build initial glossary
→ Translate chapter by chapter, batch by batch
→ Optional polish
→ Punctuation normalization
→ Chapter-end review
→ Optional auto-retranslation of severe issues
→ Optional consistency QA
→ Backfill and export EPUB/TXT

Each batch is written to state/ after translation, so long books can be resumed after interruption. Already translated batches are skipped, and only incomplete parts are filled.

Consistency Mechanisms

• Glossary: person names, place names, proper nouns, and honorifics are stored in a SQLite glossary and injected into prompts during translation as configured.
• Book understanding: pre-scan the source text before translating, generating a full book overview and chapter summaries so early chapters can also reference the overall book direction.
• Rolling context: batches within a chapter are processed sequentially; successive batches can see the most recent translated segments.
• Segment alignment: each batch includes N input segments, and the model is required to output N JSON segments; if the count doesn’t match, a retry is made; on persistent failure, segments are processed individually as fallback.
• Chapter-end review: check each chapter for missing translations, mistranslations, terminology, pronoun issues, etc.; by default, issues are only logged; whether to auto-fix is controlled by autofix_severe.
• Punctuation normalization: translated text is standardized to Mainland Chinese simplified Chinese full-width punctuation.

Common Tools

uv run trans-novel tools glossary book.epub list
uv run trans-novel tools glossary book.epub conflicts
uv run trans-novel tools qa book.epub
uv run trans-novel tools report book.epub
uv run trans-novel tools assemble book.epub

These tools are mainly for viewing the glossary, checking consistency, generating reports, or re-exporting the final file. QA and reports by default only summarize issues; they do not automatically fix the text.

Model Tiers

The default configuration uses DeepSeek via the OpenAI SDK at https://api.deepseek.com.

• strong: translation, polishing, global analysis, title translation.
• cheap: chapter-end review, consistency QA, back-translation comparison.
• fast: full book pre-scan, chapter summaries, glossary extraction, back-translation, and other mechanical tasks.

If model IDs change, simply modify llm.tiers in config.yaml.

Project Structure

trans_novel/
  ingest/       Input parsing, EPUB/FB2/TXT splitting
  llm/          LLM abstraction interface, DeepSeek provider, FakeClient
  glossary/     SQLite glossary, extraction, conflict handling
  agents/       Analysis, translation, review, polishing, consistency, prompts
  pipeline/     Orchestrator, checkpoint state, rolling context, validation
  postprocess/  Punctuation normalization
  assemble/     EPUB/TXT backfill export, QA reports
tests/          Offline tests

Testing

UV_CACHE_DIR=/tmp/uv-cache uv run python -m unittest discover -s tests

If your local uv cache directory is writable, you can also run:

uv run python -m unittest discover -s tests

Vision and Limitations

This project was developed for the author’s personal interest, aiming to make a modest contribution to the translation of long-form books. The future goal is to make translations smoother while maintaining accuracy, moving from readability to pleasurable reading. Currently, some issues (inconsistent wording for habitual expressions and inaccurate translation of proper nouns) have been improved! If you encounter any other problems, please submit an issue. If you have any ideas, feel free to discuss them in the discussion area. If you have programming skills, I welcome your PRs to make this project even better. 👏