Gemini Structured Output

"Classify the following review. Return valid JSON only: { 'sentiment': 'POSITIVE' | 'NEGATIVE', 'score': number }".

2. Working with Schemas (Input & Output)

• Input Blueprints: You can provide a JSON Schema as input to give the model a "blueprint" of the expected data.
• Mechanism: This defines expected fields (e.g., release_date), data types (e.g., string, float), and descriptions, helping the model understand relationships between data points.
• Example Schema Definition:

  {
    "type": "object",
    "properties": {
      "name": { "type": "string", "description": "Product name" },
      "price": { "type": "number", "format": "float" }
    }
  }
  

3. Handling Truncation (JSON Repair)

• The Problem: JSON is verbose. If the output hits the token limit, the JSON object may be cut off (missing closing braces), making it invalid.
• The Solution:
  • Increase Token Limit: Ensure the max_tokens setting is high enough for the data structure.
  • Use Tools: Implement a library like json-repair in your application pipeline to automatically close broken tags and fix malformed objects.

Decision Matrix: Text vs. Structured