context-fission-sdk (npm)

TypeScript SDK for ContextFission.

This package provides:

• Native TypeScript algorithms (compression, retrieval, scoring helpers, scheduling, knapsack, fusion)
• A Python runtime bridge to execute the full ContextFission pipeline behavior from Node.js

Install

npm install context-fission-sdk

For full runtime parity (pipeline + local model workflows), also install Python package:

pip install "context-fission[local-models]"

If your system Python command is not python3, pass the executable explicitly.

Quick Start (TypeScript-only features)

import { ContextFissionCompressor, BM25Retriever } from 'context-fission-sdk';

const compressor = new ContextFissionCompressor();
const compressed = compressor.compressText('In order to improve quality, we need better context.', 'medium');
console.log(compressed.savingsPct);

const retriever = new BM25Retriever();
retriever.index([
  { id: 'a', text: 'Auth tokens rotate every 15 minutes.' },
  { id: 'b', text: 'DB pool max 10 per worker.' },
]);
console.log(retriever.retrieve('auth token rotation', 1));

Full Runtime Parity from Node.js (Python Bridge)

import { ContextFissionPythonRuntime } from 'context-fission-sdk';

const runtime = new ContextFissionPythonRuntime({ pythonExecutable: 'python3' });

const install = runtime.checkInstallation();
if (!install.ok) {
  throw new Error(`context-fission is not available: ${install.error}`);
}

const model = runtime.downloadModel('Qwen/Qwen2.5-0.5B-Instruct', {
  ignorePatterns: ['flax*', 'tf_*'],
});
console.log(model.path);

const result = runtime.runPipelineFromPath('./docs/sample.md', {
  query: 'authentication flow',
  budget: 4096,
  mode: 'compact',
  compress: true,
  compressionIntensity: 'adaptive',
});

console.log(result.totalTokens, result.selectedBlocks);

Running on User Systems

1. Install Node.js 18+ and Python 3.11-3.13.
2. Install npm package: npm install context-fission-sdk
3. Install Python runtime package: pip install "context-fission[local-models]"
4. Set optional env vars for local decoder model:

export CONTEXTFISSION_LOCAL_DECODER_MODEL=Qwen/Qwen2.5-0.5B-Instruct
export CONTEXTFISSION_CAVEMAN_IO=true

5. Use the Python runtime bridge APIs in Node.js for full pipeline behavior.

Notes

• Local model inference depends on transformers + torch availability.
• On Python 3.14, torch wheels may not yet be available.
• TypeScript-only APIs still work without Python installed.