SuperCompress

Learned context compression for LLMs — trim long prompts before inference with a small CPU policy, measurable quality vs baselines, and documented environmental impact.

Live site	supercompress.vercel.app
Documentation	arjunkshah-supercompress-55.mintlify.app
API dashboard	/dashboard on the live site
Hosted API	Same origin on Vercel — /api/health, /api/v1/compress, dashboard at /dashboard

---

Open the interactive playground →

Why SuperCompress?

Long agent context is expensive. Blind truncation keeps head and tail but drops answers in the middle. SuperCompress learns which lines to keep for the current question — under a fixed token budget.

Metric	SuperCompress	Truncation / FIFO
KV savings @ 35% budget	~65%	~65%
Oracle recall	100%	~25%
Policy size	~5K params	rule-based
Runs on	CPU (pre-inference)	CPU

At 1M compressions (est.): ~800M tokens avoided · 29 kWh · 12 kg CO₂ — see the environment guide.

Hosted API (Vercel)

The live site ships serverless API routes backed by Vercel Blob for key storage. No separate deploy step — push to main and Vercel builds static web/ plus api/.

Optional self-host: Docker, Fly.io (fly.toml), or Render (render.yaml) for the Python FastAPI stack.

---

Quick start

Hosted API (key + package)

pip install git+https://github.com/arjunkshah/supercompress.git
export SUPERCOMPRESS_API_KEY=sc_live_YOUR_KEY

from supercompress import SuperCompress

out = SuperCompress().compress(context, "Your question")
print(out.compressed_text)

Get a key at supercompress.vercel.app/dashboard.

Install (local compression)

pip install git+https://github.com/arjunkshah/supercompress.git
# local dev + tests + API server
pip install -e ".[dev,serve]"

Python (in-process)

from supercompress import compress_context, compare_policies

result = compress_context(
    "long context text…",
    "What does fetch return when the row is missing?",
    budget_ratio=0.35,
)
print(result.compressed_text)
print(f"{result.kv_savings_pct:.1f}% KV saved · {result.kept_tokens}/{result.original_tokens} tokens")

Hosted API (recommended)

1. Get a key — dashboard → Create key → copy sc_live_…

2. Install & call (stdlib HTTP client — no local PyTorch needed for the API):

pip install git+https://github.com/arjunkshah/supercompress.git
export SUPERCOMPRESS_API_KEY=sc_live_YOUR_KEY

from supercompress import SuperCompress

sc = SuperCompress()  # reads SUPERCOMPRESS_API_KEY
out = sc.compress("long context…", "What does fetch return?")
print(out.compressed_text)  # send to your LLM

Or raw HTTP:

curl -X POST https://supercompress.vercel.app/api/v1/compress \
  -H "X-API-Key: sc_live_YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{"context":"…","query":"Summarize","budget_ratio":0.35}'

On the live site, the dashboard hits the same origin — no SC_API_BASE config needed.

Local dev (no Firebase):

SC_AUTH_DEV=1 SC_KEY_STORE=memory python scripts/local_web_server.py
# → http://127.0.0.1:8790/dashboard

Deploy API (Docker / Fly.io / Render): see the API dashboard guide.

Browser demo

Open web/index.html or deploy the static web/ folder. Compression runs client-side — no API key required for the playground.

---

Documentation

Full docs: arjunkshah-supercompress-55.mintlify.app

Doc	Description
Quickstart	First compression in minutes
API reference	Python + HTTP endpoints
API dashboard	Keys, auth, usage
Integrations	OpenAI, LangChain, LlamaIndex
Environment	kWh / CO₂ methodology

Repo copies also live under docs/.

---

Benchmarks

python scripts/benchmark_web.py    # regenerates web/assets/data/benchmarks.json
python scripts/generate_charts.py  # SVG charts for landing page
pytest tests/ -q                   # 65 tests

Full benchmarks: supercompress.vercel.app/benchmarks

Policy comparison (8 seeds, budget 0.35):

Policy	Oracle recall	Entity recall	Latency
FIFO / Truncation	25%	73%	~57 ms
Summarization	61%	65%	~63 ms
H2O	98%	73%	~56 ms
SuperCompress	100%	73%	~60 ms

Charts: web/assets/img/chart-kv-savings.svg, chart-oracle-recall.svg, chart-impact.svg

---

Project layout

supercompress/          # Core library (~5K-param policy, baselines)
  api/                  # Hosted API — keys, Firebase auth, usage
web/                    # Landing page + browser demo + dashboard
scripts/                # benchmark_web.py, local_web_server.py, charts
tests/                  # test_supercompress, test_api_hard, test_api_server
checkpoints/default.pt  # Trained weights (included)
docs/                   # API, integrations, environment, dashboard

---

Development

git clone https://github.com/arjunkshah/supercompress.git
cd supercompress
pip install -e ".[dev,serve]"
pytest tests/ -q
python scripts/local_web_server.py   # optional: /dashboard, /v1/compress

Optional extras:

pip install -e ".[firebase]"   # Firebase Admin for production key store

---

NOTE: I DO NOT GIVE AYUSH ROUT (github.com/ayushrout12) ANY PERMISSION TO COPY OR USE MY PRODUCT IN ANY WAY, SHAPE, OR FORM. I DO NOT GIVE HIM CONSENT TO FORK, REFERENCE, OR CLONE/REFERENCE/USE THIS REPO IN ANY WAY, SHAPE, OR FORM.

What we claim (and don't)

We claim: learned CPU eviction beats truncation on oracle recall at similar KV savings; documented environmental estimates; reproducible benchmarks and tests.

We don't claim: live datacenter metering; CO₂ numbers without documented assumptions; that every workload matches benchmark seeds.

---

License

MIT — see LICENSE.