LILT — LaTeX Intelligent Localization Tool

Continuous LaTeX localization engine (CLI) — local-first, structure-preserving, TM-backed, LLM-assisted.

LILT turns .tex prose into localized .tex without pasting raw structure into a chat model. It parses into segments, masks macros/math into placeholders, stores state in a Translation Memory, runs Draft → Critique → Refine, validates structure, then rebuilds the document. You compile the PDF yourself.

Not affiliated with Lilt Inc. / lilt.com.
The name is the acronym LaTeX Intelligent Localization Tool.
Not on PyPI as lilt (that name is another project). Install from Git; distribution package: latex-lilt; CLI: lilt.

Who it is for

Good fit	Not a fit
Authors, translators, and maintainers of academic/technical LaTeX	Markdown / UI / app string i18n
People who run OpenAI-compatible LLMs (local or cloud)	CAT / WYSIWYG seekers; drag-and-drop SaaS
Incremental TM workflows with human review and Git-friendly JSONL	“Fire and forget” with no validation
Compile-minded integrity over blind fluency	Users of commercial Lilt or pip install lilt

Why LILT

• Integrity first — placeholders and syntax validation before a translation is accepted.
• Human priority — reviewed / approved / locked are never auto-overwritten.
• TM as source of truth — append-oriented JSONL under .lilt/tm/.
• Local-first LLM — OpenAI-compatible endpoints (LM Studio, Ollama-style, or cloud) per stage.
• Continuous — re-sync when sources change; resume interrupted translates; checkpoints.

Not a generic CAT tool, not gettext/po4a, not a PDF compiler.

Key features

• Multi-file LaTeX via \input / \include discovery on sync
• Placeholder masking so the model sees prose, not raw structure
• Reflection stages: Draft → Critique → Refine (or draft-only cost profiles)
• Human Review queue and CSV/JSON export–import
• Fail-closed build of localized .tex (optional --allow-partial)
• TM inspect / budget / status tooling

How it works

.tex  →  Sync (AST → segments + TM)
      →  Translate (LLM Draft → Critique → Refine + validation)
      →  Build (localized .tex)
      →  PDF (your TeX toolchain — not a lilt command)

More: Concepts. Runtime detail: Architecture.

Quick start

Requires Python 3.13+, uv or pipx, and an OpenAI-compatible LLM before translate.

Prefer the official sample: examples/quickstart/.

# Install (Git only — do not use `pip install lilt`)
# Tool install puts `lilt` on your PATH. From a clone, use `uv run lilt` instead.
uv tool install git+https://github.com/aleaz/lilt

cd your-latex-project
lilt project init
lilt project configure .

# Edit .lilt/lilt.yaml — set source_lang, target_lang, llm.base_url, llm.model
# Cloud keys: .lilt/.env (git-ignored)

lilt pipeline sync main.tex
lilt pipeline translate --all
mkdir -p i18n/build
lilt pipeline build main main.tex i18n/build/main.tex
# Success: i18n/build/main.tex exists. Compile PDF yourself if needed.

Full walkthrough: Getting started · First translation · Quick Start example.

Common use cases

Scenario	Start here
Academic paper or book	Getting started
Multi-file project / resume after interrupt	Workflows
Local vs cloud LLM / stages / automation	Advanced usage · Configuration
Human review & TM export	Human review
Stuck on errors	Troubleshooting · FAQ

Supported capabilities

Area	Support
Input	LaTeX projects (multi-file, macros, math, citations — LaTeX-aware parsing)
LLM	OpenAI-compatible HTTP (provider: openai); per-stage models/endpoints
Workflows	Sync → translate → build; workflow or sequential modes; TM + review
Integrations	CLI + shell automation you own; no product SaaS orchestrator
Limits	No OCR/diagrams; no PDF CLI; no corpus/eval in this repo; Windows not first-class tested; Release Candidate (1.0.0rc2) — not GA; CLI/config may still change

Deferred ideas: appendix-deferred.

Documentation

Audience	Go to
Users	Docs hub · Getting started · Quick Start · Guides · FAQ
Reference	CLI · Config
Architecture	Architecture · Glossary
Developers	Developer Guide
Contributors	CONTRIBUTING.md · Contributor Guide
Problems	Troubleshooting · SUPPORT.md

Development

git clone https://github.com/aleaz/lilt && cd lilt
uv sync
make ci    # matches GitHub Actions

Details: Development overview. Process: CONTRIBUTING.md.

Project status

Release Candidate 1.0.0rc2 (git tag v1.0.0-rc.2 — see CHANGELOG). Core pipeline (sync, translate, build, review, TM, telemetry) is implemented and tested. Not General Availability. Treat the RC line as SemVer-unstable until a later stable 1.0.0: CLI and config may still change. See maturity.

Community and contribution

• Contribute: CONTRIBUTING.md · Code of Conduct
• Questions vs bugs: SUPPORT.md
• Security: SECURITY.md

License

LILT is released under the MIT License.

Copyright (c) 2026 Alejandro Azario