palimpsest

A palimpsest is a manuscript page scraped clean so the surface can be reused — the
original artwork and impressions intact, new text written over it. That's what this
tool does to a document: same pages, same artwork, same typography — only the words change.

Layout-preserving Spanish→English document translation for PDF and Office formats (.docx / .pptx / .xlsx) — as a CLI, a library, or a local web UI.

OCR when there's no text layer, styled-run extraction, in-place redraw with real embedded fonts, entity/glossary-aware machine translation (Gemini, Claude, or Google Translate), and an honest translation cache that never lets a failed translation masquerade as a successful one.

• Replicas, not rebuilds — same pages, same fonts, same artwork; only the prose changes.
• Protected entities — company names, personal names, and amounts survive verbatim, guaranteed.
• Three backends, one protocol — Gemini (free tier, default), Claude (paid, highest quality), or Google Translate (free, no key).
• An honest report, always — a paragraph that fails to translate is never silently dropped or rendered as if it succeeded.
• A local web UI, in addition to the CLI and the library — drop a file, watch it translate, compare pages side by side.

Contents: Install · Quickstart · Web UI · Why this exists · Architecture · Contributing · License

Install

pip install palimpsest-translate            # Google Translate backend only, free, no key
pip install palimpsest-translate[gemini]    # + Gemini backend (default, free API key)
pip install palimpsest-translate[anthropic] # + Claude backend (paid, highest quality)
pip install palimpsest-translate[ocr]       # + OCR for scanned PDFs (needs Tesseract on PATH)
pip install palimpsest-translate[all]       # everything

Requires Python 3.11+. To work on palimpsest itself, or to install from source:

git clone https://github.com/ianperaltahirujo/palimpsest.git
cd palimpsest
pip install -e ".[all,dev]"

Quickstart

palimpsest translate deed.pdf                    # -> deed.en.pdf, Gemini backend (needs GEMINI_API_KEY, free)
palimpsest translate deed.pdf --backend anthropic # needs ANTHROPIC_API_KEY in the environment
palimpsest translate deed.pdf --backend google    # Google Translate, no key at all
palimpsest translate deed.pdf --dry-run           # classify, count paragraphs, estimate cost -- no translation

That's it for a single file — no config required. palimpsest translate works out of the box because a missing palimpsest.toml just means "no protected entities, no glossary, default thresholds," not an error.

A whole corpus

palimpsest config init            # scaffolds private/ (gitignored) for your entity list and document map
palimpsest batch                  # translates everything named in private/documents.toml, resumes on re-run
palimpsest batch --force          # rebuild everything
palimpsest batch --only pdf --jobs 4

See examples/palimpsest.toml, examples/entities.example.toml, and examples/documents.example.toml for the file shapes, and docs/configuration.md for the full reference.

Web UI

pip install "palimpsest-translate[server]"
python tools/build_frontend.py   # builds web/prototype and stages it into the package
palimpsest serve                 # http://127.0.0.1:8765, opens your browser

Local-first, single-user, no accounts: everything runs on your own machine and binds to loopback by default (--host requires an explicit --i-know to bind wider). API keys are read from the environment exactly like the CLI, and never stored in a job record or echoed back in a response — you can also type one into the page itself instead of a shell (PUT /api/keys, applied immediately, saved to .env). See web/prototype/README.md for the frontend itself, including --dev mode for working on the UI against a live API without rebuilding on every change, its "Standalone" mode for publishing a build somewhere other than this server (e.g. GitHub Pages, via --allow-origin), and its "Office files" section for what previewing a .docx/.pptx/.xlsx in the browser needs (LibreOffice, optional and only for that preview — translation itself never needs it).

A standalone frontend build's default expectation is that each visitor runs their own palimpsest serve locally. For a zero-command deployment instead — a visitor opens the page, types an API key, and translates — see docs/deployment.md for running the backend itself somewhere persistent (a Dockerfile and a Render walkthrough), plus what per-visitor isolation that requires and doesn't guarantee.

Why this exists, not just what it does

Generic "translate this PDF" tools tend to do one of two things badly: rebuild the document from extracted text (losing layout, fonts, and artwork), or leave the original untouched and bolt on a translation elsewhere. Neither produces a document you'd actually hand to someone in place of the original.

palimpsest was extracted from a real pipeline built for legal and financial documents — trust deeds, audited financial statements, environmental permits, budget workbooks — where the layout, the signatures, the letterhead, and the exact numbers all have to survive untouched, and only the prose changes language. That constraint shaped almost every design decision here:

• Replicas, not rebuilds. A digital PDF's text is extracted with its styled runs, alignment, and measured leading, cleared without touching images or vector art, and redrawn in a real embedded font at the original position. A scanned PDF gets a real OCR text layer first.
• An honest cache. Every cached translation carries an explicit status (ok / failed / identical / refused) — a failed machine translation can never be silently served as a successful one, and a document that still has untranslated paragraphs says so in its report rather than shipping quietly incomplete.
• Protected entities as a first-class concept, not a glossary workaround — company names, personal names, and amounts are guaranteed to survive verbatim, including the specific failure modes documented in docs/design/protected-entities.md: OCR fragmenting a heading into a bare GRUPO and letting it machine-translate as if it were an ordinary noun.

The four documents in docs/design/ are a genuine engineering postmortem — what broke, why, and what specifically fixed it — carried over (and de-identified) from the pipeline this project was extracted from:

• v1-postmortem.md — the first version of this pipeline rebuilt documents from scratch and lost logos, fonts, and whole spreadsheet parts; nine concrete bugs and what fixed each one.
• protected-entities.md — a company name translated as if it were a common noun, found after a full corpus run had already shipped; two independent root causes and three separate places a protective guard got bypassed before the fix became architectural instead of another patch.
• bold-calibration.md — how bold text is recovered from scanned pixels when the OCR text layer carries no weight information, and why the threshold is a corpus-fitted number, not a universal constant.
• limitations.md — an honest account of what this pipeline does not handle, from a real corpus run rather than written speculatively.
• backends.md — why the Gemini/Claude LLM backends handle entity protection completely differently from Google Translate's phrase-level API, why Gemini is the default despite Claude being the stronger model, and the pricing/quality tradeoffs between all three.

Architecture

src/palimpsest/
  cli.py                 subcommands: translate, batch, compare, cache, audit, config
  corpus.py               whole-corpus batch orchestration
  config/                 layered TOML config (packaged defaults -> project file -> private/)
  core/                   cross-platform paths, errors, logging, the IR serialization boundary
  text/                   entity protection, glossary, ordinals, post-translation fixups
  pdf/                    classify, OCR, layout extraction, font resolution, clearing, render, pipeline
  office/                 OOXML surgery (translate .xlsx/.docx/.pptx by editing the zip in place)
  translate/              backend protocol, Gemini + Claude + Google backends, cache, cost estimation
  qa/                     side-by-side comparison renders, bilingual PDF output, cache audit

Three backends behind one Backend protocol: Gemini (the google-genai SDK, free-tier, LLM — the default) and Claude (the anthropic SDK, paid, LLM) both protect entities by prompt instruction plus a post-hoc verification pass, since an LLM benefits from reading the whole sentence in a way placeholder substitution would defeat; Google Translate (deep-translator, free, phrase-level, no key at all) has no concept of "entity" at all, so its entities are protected by placeholder substitution instead. See docs/design/backends.md for why they're architected differently rather than sharing one protection scheme.

What this borrows from, and doesn't share code with

Several architectural ideas — most notably TOML-based project configuration and bilingual side-by-side PDF output — were inspired by BabelDOC, an AGPL-3.0 scientific-PDF translation tool, after comparing it against this pipeline's own approach. BabelDOC's headline feature, ML-based layout analysis (DocLayout-YOLO), was deliberately not adopted: this pipeline's own geometry-based column/paragraph detection (pdf.layout.column_bands) already solves the layout problem it exists for, for this document type, without a model download or GPU dependency. See NOTICE — no BabelDOC source is included in or derived into this project.

Contributing

See CONTRIBUTING.md. The short version: no test may make a real network call, and tools/scrub_check.py runs in CI to keep this repository free of the confidential document data it was extracted from.

License

Apache 2.0 — see LICENSE. See NOTICE for the BabelDOC design-inspiration credit above.