🗜️ llm-zip

Context compression sidecar for LLM applications.
Reduce token costs 3–5× before calling any AI API.

Why · How it works · Quickstart · CLI · API response · File support · Benchmarks · Limitations · Contributing

Why llm-zip exists

AI inference is getting expensive. As context windows grow and agentic workflows multiply, the tokens you send to GPT, Claude, or Gemini compound fast — and the "free ride" subsidy era is ending.

llm-zip was built out of a simple need: compress context before it reaches the model, without changing anything else in the stack. No proxy. No middleware. No API keys. Just a sidecar you call over HTTP, that hands back a smaller text and tells you exactly how much you saved.

It started as an internal tool. It's open source because the problem is universal.

Curious about how this started? Read STORY.md.

How it works

Your app  ──→  POST /v1/estimate  ──→  llm-zip
                                           │
                              counts tokens, estimates savings
                              (no compression — CPU-free)
                                           │
              token counts + savings     ◄─┘

Your app  ──→  POST /v1/compress  ──→  llm-zip
                                           │
                              compresses with LLMLingua-2
                              scores semantic preservation
                              calculates USD savings
                                           │
              compressed text + metrics  ◄─┘

Your app  ──→  calls OpenAI / Anthropic / Gemini
               with the smaller context

llm-zip never touches your API keys or your model calls. It only compresses.

Quickstart

Requires Docker. First-time model download is ~700MB and takes 2–5 min.

Step 1 — Configure

Windows (CMD / PowerShell):

copy .llmzip.config.example .llmzip.config
notepad .llmzip.config

Linux / macOS:

cp .llmzip.config.example .llmzip.config
nano .llmzip.config

Step 2 — Download models (only on first run)

docker-compose run llmzip download-models

Step 3 — Start

docker-compose up -d

→ API: http://localhost:8000 · Docs: http://localhost:8000/docs

Running in split mode? See v0.2.0 release notes for docker-compose.split.yml and separate Dockerfile.api / Dockerfile.models — lets you scale the API layer without duplicating the ~700MB model weights.

The CLI Experience

llm-zip isn't just an API — it comes with a UNIX-friendly CLI powered by Typer and Rich. It supports stdin pipes, outputs clean JSON, and tracks live model prices.

# Check current prices (fetches live from LiteLLM)
$ llmzip prices -p anthropic

Rates from LiteLLM as of 2026-06-05
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━━┓
┃ Model                                     ┃ Input $/M ┃ Output $/M ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━━┩
│ anthropic.claude-3-5-sonnet-20241022-v2:0 │    3.0000 │    15.0000 │
│ anthropic.claude-3-haiku-20240307-v1:0    │    0.2500 │     1.2500 │
└───────────────────────────────────────────┴───────────┴────────────┘

Model names come directly from LiteLLM and may use provider-specific naming (e.g. Bedrock-style). The exact list depends on live data at execution time.

# Compress directly from the terminal via pipes
$ cat huge_context.txt | llmzip compress --ratio 0.4 --model gpt-4o-mini > compressed.txt
✓ 24000 → 9600 tokens (2.5×) | score: 0.92 | saved: ~$0.0021 (gpt-4o-mini)

What you get back

{
  "compressed": "...",
  "original_tokens": 8400,
  "compressed_tokens": 1680,
  "compression_ratio": 5.0,
  "preservation_score": 0.91,
  "estimated_savings": {
    "gpt-4o-mini":           "$0.0010",
    "claude-haiku-4-5":      "$0.0013",
    "gemini-2.5-flash-lite": "$0.0006"
  },
  "skipped": false,
  "warning": null
}

Prices are fetched live from LiteLLM with local caching — no manual database updates needed.

MarkItDown Integration

One of the practical problems that motivated llm-zip was RAG over internal documents. In private systems — invoices, reports, manuals, internal specs — the source content lives in PDFs, Word files, and spreadsheets.

llm-zip integrates MarkItDown as an optional preprocessing layer. When FILE_CONVERSION=true, you can send a file directly to /v1/compress/file and get back compressed text ready for your RAG pipeline.

curl -X POST http://localhost:8000/v1/compress/file \
  -F "file=@invoice.pdf" \
  -F "ratio=0.5"

Supported formats: PDF, Word (.docx), Excel (.xlsx), PowerPoint (.pptx), and more.

Key Features

Environment Variables

For CI or Docker deployments, you can override defaults with environment variables:

📊 The Science & Benchmarks

llm-zip is powered by Microsoft's LLMLingua-2. According to their research paper evaluated on the LongBench dataset:

• Compression sweet spot: optimal performance between 2× and 5× compression
• Quality retention: 90–98% Exact Match performance vs. uncompressed prompts
• Latency reduction: end-to-end latency improvement of 1.6× to 2.9× (task-agnostic, no prompt needed to start)

🧪 Internal Test Results

The following results were obtained by running llm-zip against real documents during development. They are not synthetic benchmarks — they reflect actual compression output on representative material.

How savings are calculated: tokens_saved × input_price_per_token, using live prices fetched from LiteLLM at run time. For OpenAI models, tokenization is exact (tiktoken). For Anthropic, Google and DeepSeek, a character-ratio heuristic is used (±10% margin). Only input token savings are counted — output token reduction, which also occurs in many workflows, is not included. The figures below are approximate lower-bound estimates: real savings may be higher or lower depending on your actual usage pattern and the model you call.

¹ Estimated saving for claude-sonnet-4-6 input tokens only. See methodology note above. ² Synthetic test dataset modeled after Argentine customs documentation (Código Aduanero, NCM, ARCA/DGA). Not a real operational document.

Limitations & Caveats

• Language support — LLMLingua-2 was trained on English meeting transcripts. Spanish and other languages work but may see 10–15% lower compression quality. Community benchmarks are welcome.
• Extractive compression — tokens are removed, not rewritten. Some nuance may be lost at aggressive ratios (above 0.7).
• Pricing accuracy — savings are estimates. Non-OpenAI tokenizers use a character-ratio heuristic with a ±10% margin of error.
• Format loss — file conversion extracts plain text only. Layout, tables, and complex formatting are not preserved.
• Production testing — this is v0.2.0. Concurrent batch processing should be monitored closely in high-throughput environments.

Contributing

The most valuable contributions right now are benchmark results across different languages, domains, and models. If you use llm-zip in a real project, a PR with your numbers helps everyone calibrate expectations.

Bug reports, feature requests, and code contributions are also welcome — see CONTRIBUTING.md.

git clone https://github.com/finktech-dev/llm-zip.git
cd llm-zip
pip install -e ".[dev]"
pytest

License & Disclaimer

MIT © FinkTech

Built utilizing LLMLingua-2 and MarkItDown open-source libraries.

This project is an independent, community-driven open-source tool. It is not affiliated with, endorsed by, or sponsored by Microsoft Corporation, OpenAI, Anthropic, or any other LLM provider. All trademarks and registered trademarks are the property of their respective owners.