BookForge
BookForge is the EPUB translation engine that keeps the LLM away from your document structure. It parses EPUBs into validated JSON payloads, checkpoints every segment, preserves markup/footnotes/links, and rebuilds valid EPUBs.
I built this to translate books for my partner. It's MIT-licensed in case it's useful to you.
Why BookForge
EPUB structure is program-owned. Models receive prose-only JSON payloads and never see or regenerate raw XHTML. Inline markers, protected spans, package metadata, resources, and document ordering are validated and reassembled by deterministic Rust code.
That boundary is the point of the project: malformed model output can be retried without asking another model to repair the book. The result is a checkpointed translation workflow whose structure can be tested independently of translation quality.
Status
BookForge v1.8 is usable for EPUB translation and PDF-to-EPUB ingestion:
- EPUB inspect, parse, segment, and rebuild
- EPUBCheck-backed standalone and post-translation validation
- EPUBCheck-clean structural regression against a pinned nine-book Standard Ebooks corpus; see docs/corpus.md
- Plain, marker-safe, and run-preserving translation contracts
- Mock provider for deterministic tests
- OpenAI-compatible provider
- DeepSeek and OpenRouter presets
- Ollama and llama.cpp local-model presets
- Bounded parallel segment translation with
--concurrency - SQLite checkpoint store
- Resume and retry commands
- Status and tail commands for persisted jobs
- Segment-level cache reuse for compatible prior translations
- Static side-by-side review HTML with flag export/import
- QA reports in JSON and Markdown
- Optional LLM QA review pass
- Cost estimates for known provider/model pairs
- Externalized, overridable provider pricing
Install
Install from crates.io:
Or build the current checkout:
The binary is:
For development, use:
Quick start
Use cargo run -p bookforge-cli -- in front of commands when running from a
source checkout. Provider preset names are shown by bookforge translate --help.
Commands
Convert a PDF to a translatable EPUB (requires poppler
command-line tools on PATH, or POPPLER_PATH pointing at their bin
directory; on Windows use the poppler-windows release zip):
The converter detects two-column layouts per page (scientific papers),
repairs hyphenated line breaks, joins paragraphs across pages, and maps
oversized fonts to headings. It prints a fidelity report comparing the
reconstructed text against the raw pdftotext baseline — check that
coverage number (and inspect on the result) before spending tokens.
Figures, tables-as-images, and low-confidence page fallbacks are
roadmap items (ROADMAP §9b, phases P2–P4); for image-heavy PDFs expect
text-only output for now.
Inspect an EPUB:
The inspect output includes a text-coverage metric: the percentage of
visible body text that lands in translatable blocks. Files with low
coverage (text in unsupported markup such as bare <div>s) are listed
individually — that text would ship untranslated, so check coverage
before spending tokens on a full run.
Estimate tokens and approximate cost:
Pricing is loaded from the bundled pricing/providers.json. Override it with
--pricing custom.json or BOOKFORGE_PRICING_PATH.
Translate with OpenRouter:
Translate with the default fast profile:
Translate with a glossary:
Check provider and storage health:
Translate with DeepSeek:
Use any OpenAI-compatible endpoint:
Local Ollama and llama.cpp recipes are documented in docs/local-models.md.
Resume a job:
Generate a side-by-side review page:
Ingest exported review flags and mark bad translations for retry:
Manage glossary terms:
Inspect persisted job state and recent events:
Retry failed or review-needed segments:
Validate a translated EPUB and report:
BookForge invokes EPUBCheck when it is available. Set
BOOKFORGE_EPUBCHECK to an executable, its containing directory, or an
epubcheck.jar. Missing EPUBCheck is reported as status: unavailable and is
non-fatal. Use --strict-epubcheck to make warnings fail validation.
QA Modes
Translation always runs hard validators before committing a segment. The optional LLM QA pass is controlled with:
off is the default. Reports still include deterministic soft warnings such as changed URLs, changed numbers, suspicious length ratios, model commentary, and repeated text.
Two structural defaults to know about:
pre/codeblocks are never sent to the model. They are copied through to the output byte-for-byte, preserving internal whitespace.- The sliding context window (
--context-window, default 3) is best-effort: a segment uses whichever predecessors have already finished and never waits for them. Pass--context-strictto restore the v1.3 fence behavior, which guarantees a complete context block but serializes segments within the context scope.
Checkpoints And Cache
Runtime state is stored in:
.bookforge/jobs.sqlite
That path is ignored by git. Segment translations are persisted as each segment completes. New jobs reuse compatible cached translations when the source hash, prompt version, provider, model, source language, and target language match.
Progress events can be written in every UI mode:
Review artifacts contain the full source and translated text of the book. They are written locally under .bookforge/runs/<job-id>/review/; treat them as private user data.
Known limitations: provider API keys are read from environment variables. PDF ingestion currently prioritizes text reconstruction; complex figures and tables may require review of the conversion report.
Benchmarks
Run the mock release smoke benchmark with:
See docs/benchmarks.md for metrics to capture in real-provider runs.
Run the pinned structural corpus with:
Secrets And Local Tests
Do not commit API keys or ad hoc test books. The repository ignores:
test/
.bookforge/
.claude/
.codex
*.env
*.key
key.txt
For local OpenRouter testing, place the key outside tracked paths or export it directly:
Development Checks
See CONTRIBUTING.md for what's expected in issues
and pull requests, and the architectural invariants any change has to
respect.
Repository Layout
crates/bookforge-core IR, segmentation, shared config
crates/bookforge-epub EPUB inspect/read/rebuild
crates/bookforge-llm prompts, providers, scheduler, validators
crates/bookforge-llm/prompts Versioned prompt templates
crates/bookforge-store SQLite checkpoint store
crates/bookforge-cli CLI commands and reports
docs/ Architecture notes
pricing/ bundled provider/model pricing
tests/corpus/ pinned Standard Ebooks corpus manifest
BookForge remains a tool built for one reader and shared under MIT. Bug reports should include the BookForge version, operating system, provider/model, and a redacted validation or QA report where possible. The sequenced project plan is in docs/ROADMAP.md.