01 · pruning
Less context. More velocity.
Trim repetitive history and irrelevant retrieval before requests reach your model.
12,840 tok in3,920 tok out
Make every
token count.
Lumi0 shapes raw AI context into the exact signal your model needs. Lower cost, faster answers, and less noise in every request.
context.pipeline
Drag to compare.
69.5% fewer tokens · same answer
before12,840 tokens
1const messages = [
2 { role: "system", content: SYSTEM_PROMPT }, // 1,240 tok
3 ...conversation.history, // 9,870 tok
4 ...retrieval.documents, // 1,730 tok
5 { role: "user", content: question },
6];
7
8const response = await model.chat({ messages });
9// -> 12,840 prompt tokens · 2.1s TTFT
Capabilities
Compression you can actually see.
every request is ranked, pruned, and auditable
02 · ranking
Context that follows the task.
Keep the decisions, constraints, and facts that make each answer grounded.
03 · audit
Inspect every reduction.
Understand what stayed, what left, and why with a transparent pipeline.
How it works
Three steps between you and a cheaper prompt.
One POST in — a compressed, budget-fit prompt out. No agent rewrites, no retrieval surgery.
in · 12,840 tok
Send us the mess.
Pass conversation history, retrieved documents, and system prompts in one request. No reformatting required.
score → prune → rebuild
We keep the signal.
Lumi0 scores every fragment against the live query, drops redundancy, and rebuilds a prompt that fits your budget.
out · 3,920 tok · 200 OK
Your model answers.
The compressed payload drops straight into your existing chat call. Same API, same model — a fraction of the tokens.
Measured in production
rolling 30d · all regions69%average tokens removed
3.4×faster time to first token
98.2%answer quality retained
<40mscompression overhead
Drop-in SDK
Two lines. Then it disappears.
Wrap your existing messages with compress() and keep the rest of your stack exactly as it is. Works with any model provider.
- TypeScript and Python SDKs
- Provider-agnostic — OpenAI, Anthropic, local
- Deterministic budgets per request
$ npm i lumi0
npm i lumi0 · pip install lumi0
agent.ts~3,920 tokens
1import { Lumi0 } from "lumi0";
2
3const messages = await lumi0.compress(rawMessages, {
4 query: question,
5 budget: 4096,
6});
7
8const response = await model.chat({ messages });
9// -> same model, a fraction of the tokensBuilt for your AI stack
Your context layer,
without the overhead.
Start with a single API call. Lumi0 works alongside the models, frameworks, and data sources you already use.