@ramenm/soft-llm-stream

Small headless smoothing for bursty LLM text streams.

soft-llm-stream sits between an incoming stream and the UI. It normalizes real provider output, keeps reveal speed perceptually steadier, preserves grapheme boundaries, stays soft after long gaps, and finishes without ugly tail snaps.

Requirements

Node.js 22+.

Install

npm install @ramenm/soft-llm-stream

Basic usage

import { createSoftLlmChatStream } from '@ramenm/soft-llm-stream';

const store = createSoftLlmChatStream({
  source: fetch('/api/chat'),
  adapter: 'sse',
  pacing: 'auto',
});

store.subscribe(() => {
  const snapshot = store.getSnapshot();
  render(snapshot.text, snapshot.meta.phase);
});

await store.start();

Published package: core runtime plus an optional AI SDK transform wrapper. It never parses Markdown and never writes HTML. Full demo, docs, tests, and validation reports live in the GitHub repo.