09. Output Format Control

Controlling output format matters when the model output feeds into downstream systems. Unstructured text requires parsing; structured output can be consumed directly.

Format control uses three techniques:

1. Explicit structure in the prompt: "Return JSON with fields X, Y, Z"
2. Demonstrated examples: Show the exact format you want
3. Constraint statements: "Do not add explanations"

Return only the JSON object, no additional text:
{
  "review_sentiment": "positive",
  "key_phrases": ["fast delivery", "quality packaging"],
  "product_mentioned": true
}

Adding "Return only" prevents the model from adding explanatory text around the structured output.

Format control is essential for:

• JSON output consumed by applications
• Markdown tables for data transfer
• CSV-compatible structures
• Code generation with specific syntax

Generate a YAML configuration file for a web server with these requirements:
- Listen on port 8080
- Enable compression
- Set max request size to 10MB
- Log errors only

Return ONLY the YAML, no preamble or explanation.

The model may still add markdown code fences (```yaml). If you need raw output, specify:

Return raw YAML without code fences, markdown formatting, or explanatory text.

For complex formats, show a complete example:

Generate a JSON configuration for this deployment:

{
  "app": "my-service",
  "version": "1.2.3",
  "replicas": 3,
  "resources": {
    "cpu": "500m",
    "memory": "256Mi"
  },
  "env": {
    "DATABASE_URL": "postgres://db:5432/app",
    "LOG_LEVEL": "info"
  }
}

Now generate for:
app: payment-processor
version: 2.0.0
replicas: 5
cpu: 1000m, memory: 512Mi
env: DATABASE_URL, API_KEY (placeholder), LOG_LEVEL: debug

The example establishes format, indentation style, and placeholder conventions.

Local verification checkpoint

Run the smallest example from this chapter in a local workspace and record the package version, runtime, data path, and observed output. If the result depends on model size, vector count, CPU/GPU backend, or available memory, note that constraint beside the exercise so the lesson remains reproducible.