Sugo

Documents, reduced. Feed your LLM the sauce, not the water.

🇮🇹 Leggi in italiano

---

Sugo turns PDFs, Office files, screenshots and images into clean, compact Markdown for RAG systems, AI agents and LLM chats. It removes what only costs tokens (repeated headers and footers, page numbers, hidden text, tracking parameters, app chrome around a screenshot) and keeps the content, the tables and the structure. Everything runs on your computer: no cloud API, no telemetry, nothing sent anywhere.

In Italian, il sugo del discorso means "the gist", and a sugo is a sauce you reduce until only the concentrated flavour remains. That is what Sugo does to your documents.

What it does

A three-page PDF, as a plain text extractor gives it to your model (pages 1 and 3 look the same):

ACME S.p.A. - Annual Report 2025                Confidential
3.2 Revenue
Revenue grew by 12% in the second half, driven by the export market.
Details: https://acme.example/report?utm_source=newsletter&utm_medium=email
Page 14 of 120

What sugo report.pdf prints:

ACME S.p.A. - Annual Report 2025

# 3.1 Costs

Costs fell by 3% thanks to the new suppliers.
Details: https://acme.example/report

# 3.2 Revenue

Revenue grew by 12% in the second half, driven by the export market.
Details: https://acme.example/report

# 3.3 Outlook

The board expects stable growth next year.
Details: https://acme.example/report

The header is kept once instead of on every page; the page numbers, the repeated "Confidential" and the tracking parameters are gone; the text is the same. sugo stats measures the difference: 180 → 97 tokens, 46% saved on this small example. Every removal is recorded with the cleaner that made it and the reason, so you can always check what was dropped. Text that cannot be read is marked as illegible, never guessed.

The same holds across a benchmark corpus in five languages, with no content lost on any document:

Document	Tokens saved
Report with a running header, footer and page numbers	33–38%
Chat or code screenshot, Markdown vs. sending the image itself	94–98% cheaper (Claude, Gemini, OpenAI)
A real EU regulation from EUR-Lex, mostly substantive legal text	2.9–3.5%

And it stays fast on an ordinary computer, not just a new one: on a 2010 desktop with 2 cores and no GPU, a 100-page PDF converts in 3.6 s and a phone screenshot in 13.7 s (full numbers in TESTING).

Status

Alpha. The first release, v0.1, is being finished: everything below works today from a source install. The PyPI package comes with v0.1.

Works today	Coming next
Native PDFs, DOCX, XLSX, PPTX, XML	Scanned-book cleaners, EPUB, email, web pages
Images and screenshots with local OCR (PNG, JPEG, WEBP)	Larger tables, normal and quality engine levels
Chat and code screenshot detection	Japanese, Chinese, Korean
Markdown, and JSON Lines chunks for vector databases	Interactive shell, guided setup
Token metrics per provider (sugo stats)	Background service and folder watcher
Command line, MCP server for agents, local web page	Work profiles and recipes
Interface and text rules in English, Italian, Spanish, French, German

See the roadmap for the full plan.

Quick start

Sugo needs Python 3.12 and uv.

uv tool install git+https://github.com/Muralnick/sugo

Convert a file and print its Markdown:

sugo report.pdf

More things you can do:

sugo convert *.pdf -o out/            # one NAME.md per file, in out/
sugo convert report.pdf --chunks      # chunks as JSON Lines, ready for RAG
sugo stats screenshot.png             # tokens saved, per LLM provider
sugo clip                             # convert the screenshot in your clipboard
sugo ui                               # open the local web page: drop a file, get Markdown
sugo doctor                           # check the installation

Add --json for output a script can read.

Use it from an AI agent (MCP)

Sugo can serve agents such as Claude Code or Codex over MCP. It reads only inside the folders you name:

claude mcp add sugo -- sugo mcp --root ~/Documents

The agent gets two tools: convert_document and estimate_savings.

On an old computer

Measured on 2026-11-03 on a 2010 desktop: Intel Core i5-650 (2 cores, 4 threads, no AVX), 8 GB RAM, SSD, no usable GPU. Weaker than the 8 GB, 4-core machine Sugo targets.

Task	Result	Target
Native PDF, 100 pages, cleaning included	3.6 s, about 28 pages/s	≥ 20 pages/s
Peak memory, that PDF	98 MB	< 1 GB
Dense phone screenshot (30 lines) → text, OCR	30 of 30 lines, 0% character errors, 411 MB	< 1 GB
Same screenshot, time	13.7 s	< 2 s

Native documents run well within the targets even on this machine. OCR is accurate but slow on a CPU without AVX; making it faster is on the roadmap. Details and scripts are in TESTING.

Supported files

Type	Formats	Size limit
Documents	PDF (with a text layer), DOCX, XLSX, PPTX	200 MB
Images	PNG, JPEG, WEBP (screenshots, photos of text)	50 MB
Data	XML	50 MB

Old .doc, .xls and .ppt files and password-protected files are not read: save a new copy first.

How it works

flowchart LR
    F[Your file] --> E[Reader<br/>PDF · Office · OCR · XML]
    E --> IR[Document → Page → Block]
    IR --> C[Cleaners<br/>propose removals]
    C --> K[Core<br/>applies and records them]
    K --> M[Markdown]
    K --> J[JSONL chunks]
    K --> S[Token metrics]

Each reader turns a file into the same structure of pages and blocks. Cleaners only propose changes; the core applies them and marks each removed block with who removed it and why, instead of deleting it. That is what makes the output checkable and reproducible. The full design is in ARCHITECTURE.

Promises

These are written down in PRINCIPLES and checked by tests:

• Local by principle. Conversion never needs the internet. No cloud API, no telemetry.
• No guessing. Unreadable text is marked as illegible, never filled in.
• Nothing silently lost. Every removal is recorded, with its reason.
• Honest metrics. Savings are measured, and Sugo also tells you when sending the original image would be cheaper.
• Safe with untrusted files. Size and time limits, XML parsed safely, models verified by checksum. See the threat model.

Documentation

Document	Content
ARCHITECTURE	Complete system design
FEATURES	Features by engine level, work profiles, supported formats
THREAT_MODEL	Security analysis
TESTING	Testing, benchmarks, performance targets
ROADMAP	Development plan and first release scope
ADRs	Architecture decision records
PRINCIPLES	The promises Sugo never breaks
AI_WORKFLOW	How this project is built with AI assistance

Contributing and security

See CONTRIBUTING. Please report vulnerabilities privately, as described in SECURITY, not in public issues.

License

AGPL-3.0-or-later (see ADR-010).