zod-stream adds structured output validation and streaming capabilities to LLM responses. Built on top of schema-stream, it enables type-safe extraction with progressive validation.

Key Features

• 🔄 Stream structured LLM outputs with validation
• 🎯 Multiple response modes (TOOLS, FUNCTIONS, JSON, etc.)
• 📝 OpenAI client integration
• 🌳 Progressive validation with partial results
• ⚡ Built on schema-stream
• 🔍 Full TypeScript support

Why zod-stream?

zod-stream solves key challenges in handling streaming LLM responses:

• Dependency Management: Process data as soon as dependencies are met, rather than waiting for complete responses

  if (isPathComplete(['user', 'preferences'], chunk)) {
    // Start personalizing immediately, don't wait for content
    initializeUserExperience(chunk.user.preferences);
  }

• Type-Safe LLM Integration: Full TypeScript support for structured outputs from OpenAI and other providers

  const params = withResponseModel({
    response_model: { schema, name: "Extract" },
    mode: "TOOLS"  // or "FUNCTIONS", "JSON", etc.
  });

• Progressive Processing: Built on schema-stream for immediate access to partial results

  for await (const chunk of stream) {
    // Safely access partial data with full type inference
    chunk._meta._completedPaths.forEach(path => {
      processDependency(path, chunk);
    });
  }

• Provider Flexibility: Consistent interface across different LLM response formats

  // Works with various response modes
  const stream = OAIStream({ res: completion });  // OpenAI tools/functions
  const stream = JSONStream({ res: completion }); // Direct JSON

Think of it as a type-safe pipeline for handling streaming LLM data where you need to:

• Start processing before the full response arrives
• Ensure type safety throughout the stream
• Handle complex data dependencies
• Work with multiple LLM response formats