Streaming

Guard text as it streams from the LLM using StreamingPipeline.

Usage

import { StreamingPipeline, promptInjection, toxicity } from 'open-guardrail';

const pipeline = new StreamingPipeline({
  guards: [
    promptInjection({ action: 'block' }),
    toxicity({ action: 'block' }),
  ],
});

// Feed chunks as they arrive
await pipeline.pushChunk('Hello, ');
await pipeline.pushChunk('how can I ');
await pipeline.pushChunk('help you?');

// Finalize and run full-text semantic checks
const result = await pipeline.finish();

How It Works

1. Chunk-level checks — guards with supportsStreaming: true run checkChunk() on each incoming chunk
2. Full-text check — on finish(), all guards run check() on the complete accumulated text
3. Early exit — if a chunk-level check blocks, you can stop streaming immediately

Guards with Streaming Support

Guards that implement checkChunk() can catch issues during streaming:

• keyword — matches on each chunk
• regex — pattern checks per chunk
• wordCount — running character/word count
• pii — detects PII patterns in chunks