🤖 Two-stage AI calls for structured output

Unless latency is critical, you should use 2 AI calls to generate structured output.

1. First, generate the raw text or reasoning.
2. Then, structure or format it in a second call.

The benefit is that you can let the AI generate step-by-step reasoning and analysis to solve the problem without forcing the solution into a strict JSON schema. It can focus on the reasoning itself, not the format.

If you add a key like reasoning_steps to your JSON, you're still confining the reasoning to a rigid structure. That structure doesn't improve the quality of the reasoning, and focusing on it might get in the way.

Example Implementation

// Step 1: Generate raw analysis with detailed reasoning
const analysis = await generateText({
  model: ANALYSIS_MODEL,
  providerOptions: {
    openai: {
      reasoningEffort: 'high',
    }
  },
  system: analysisSystemMessage,
  prompt: analysisPrompt,
});

// Step 2: Structure the raw analysis into desired format
const structuredResponse = await convertTextToJson(
  analysis.content,
  ResponseSchema
);

The key difference is in the prompt structure:

Analysis Prompt (Step 1): Focuses purely on thinking and reasoning without format constraints. Encourages comprehensive analysis, exploration of multiple angles, and thorough examination of the problem.

Structuring Prompt (Step 2): Takes the raw analysis and formats it according to specific schema requirements, extracting key insights and organizing them into the desired structure.