Prompting best practices when using structured output (output_schema)

[sigott]

Just wondering if there are any prompting guidelines when using output_schema. Is it still useful to use few shot prompting (give examples of output JSON) when using output_schema or is that unnecessary. My rudimentary testing seems to suggest few shot prompting isn't needed but wouldn't mind hearing from others.

[adk-bot]

Response from ADK Answering Agent (experimental, answer may be inaccurate)

TLDR: Your intuition is correct—output_schema is powerful and often removes the need for few-shot examples for structure alone. However, providing examples can still be very useful for guiding the semantic content and handling complex logic, especially now that output_schema can be used in combination with other tools.

---

Hello! That's a great question, and your initial testing is on the right track.

Here’s a more detailed breakdown of the best practices:

1. output_schema is often sufficient: For many use cases, defining a Pydantic BaseModel in output_schema is enough. The model is highly capable of understanding the desired structure and generating a valid JSON response that conforms to your schema. This is because output_schema acts as a strong constraint on the model's final output. [1]

2. When to use Few-Shot Examples: While not always necessary for the structure, few-shot examples remain a valuable tool for guiding the content and logic of the output. Consider using them when:

   • The logic is complex: If the mapping from input to the structured output is nuanced or requires specific reasoning, examples can demonstrate this logic more clearly than instructions alone.
   • Specific formatting is required: If you need a field to have a specific format that isn't captured by the data type (e.g., a string that must be a formatted date, or a name in ALL_CAPS), providing an example is an excellent way to show the model what you want.
   • Guiding a sequence of actions: A recent ADK update allows output_schema to be used in conjunction with other tools. In this scenario, few-shot examples can demonstrate a full workflow: the user asks a question, the agent uses a tool to find information, and then it uses the special set_model_response tool to format the final answer according to your schema. [2]

3. Alternative: Detailed Instructions: Don't forget the power of well-written instructions. Instead of (or in addition to) few-shot examples, you can provide a detailed description of the desired output, explaining what each field represents and how to populate it. This can be very effective for guiding the model's behavior. [3]

In summary, you can rely on output_schema for structural enforcement, but keep few-shot examples in your toolkit to refine the model's output for more complex or specific requirements.

[1] https://google.github.io/adk-docs/agents/llm-agents/
[2] https://github.com/google/adk-python/blob/main/contributing/samples/output_schema_with_tools/README.md
[3] https://github.com/google/adk-python/blob/main/contributing/samples/multi_agent_seq_config/sub_agents/code_reviewer_agent.yaml