VitePress

Appearance

Welcome to @selfagency/llm-stream-parser ​

Composable parsers and stream processing utilities for LLM responses.

Overview ​

@selfagency/llm-stream-parser provides production-ready parsing utilities for handling LLM output streams with support for:

  • Thinking extraction - Extract and process reasoning sections (e.g., <think>...</think>)
  • XML filtering - Context block deduplication and filtering
  • Tool-call extraction - Parse and validate structured tool invocations
  • Structured output - JSON parsing with schema validation and repair
  • Stream processing - Flexible processor for orchestrating multiple parsers
  • Adapters - Generic and specialized adapters for different integration patterns

Quick Start ​

Installation ​

bash
npm install @selfagency/llm-stream-parser
pnpm add @selfagency/llm-stream-parser
yarn add @selfagency/llm-stream-parser

Basic Usage ​

typescript
import { parseJson, extractXmlToolCalls } from '@selfagency/llm-stream-parser';

// Parse JSON from streaming LLM output
const result = parseJson(streamData);

// Extract tool calls from XML format
const toolCalls = extractXmlToolCalls(response, new Set(['send_message', 'search']));

Features ​

  • Streaming-friendly - Designed for chunk-by-chunk processing
  • Composable - Mix and match parsers with the pipe operator
  • Type-safe - Full TypeScript support
  • Limit enforcement - Built-in safety rails for JSON depth, key counts, and tool calls
  • Privacy-aware - Configurable privacy scrubbing for sensitive context blocks

Use Cases ​

  • Agent frameworks - Process Claude or other LLM outputs in agents
  • RAG systems - Parse structured output from retrieval-augmented generation
  • Tool use - Extract and validate tool calls from model responses
  • Real-time applications - Stream processing without buffering entire responses

Documentation ​

License ​

MIT