Structured output from the LLM Agent...

[rajib76]

I want to create a structured output from the agent. Basically I would like to give it a pydantic schema and expect the output to be mapped to that schema. How do I do that with ADK. For example in the below example, I want each of the discussion items as a str in a List.

meeting_notes_agent = LlmAgent(
    name="NotesCreator",
    description="Creates notes from an meeting transcript",
    instruction="""
    You are an expert notes extractor. You will extract meeting notes from a given meeting transcript. 
    Here is the transcript for you.
    {meeting_transcript}
    """,
    output_key="notes_output"

)

I will create a pydantic object as below

class Note:
   note_item: str = Field(description="a specific discussion point in the transcript")

class Notes
   note_items: list[Note]

[hangfei]

Use output schema: https://google.github.io/adk-docs/agents/llm-agents/#structuring-data-input_schema-output_schema-output_key

[BrownsugarZeer]

Hi @hangfei, regarding the output_schema mentioned in the docs:

Your instructions must guide the LLM to produce JSON matching the schema directly.

Does this mean I still need an after_model_callback to ensure the JSON format is correct, or will the output_schema guarantee a valid output structure?

FYI, trustcall is a useful tool built on LangGraph. I continue to use it because output_schema can sometimes produce incorrectly formatted JSON, while trustcall guarantees the correct format.

from pydantic import BaseModel
from trustcall import create_extractor

class Foo(BaseModel):
    x: int

bound = create_extractor(llm, tools=[Foo], tool_choice="Foo")
result = bound.invoke("Give me a number.")
result["responses"][0]
# return in pydantic instance: Foo(x=42)

Does ADK have the same plan in the future?

Thanks!

[rajib76]

I figured it out, the name and model was missing. But I may have bumped into another potential defect. Will do a little more investigation and share

[rajib76]

Below is the error

2025-05-10 10:00:02,285 - INFO - cost_calculator.py:655 - selected model name for cost calculation: openai/gpt-4o-2024-08-06
2025-05-10 10:00:02,290 - ERROR - fast_api.py:637 - Error in event_generator: 1 validation error for ChatTranscript
  Invalid JSON: expected value at line 1 column 1 [type=json_invalid, input_value='```json\n{\n  "meeting_t...ty"\n    ]\n  }\n}\n```', input_type=str]
    For further information visit https://errors.pydantic.dev/2.11/v/json_invalid
Traceback (most recent call last):

It looks like the framework does not ensure to create a valid JSON model. If I do the same with Open AI SDK's structured output, it works perfectly fine.

The code used is as below

from dotenv import load_dotenv
from google.adk.agents import LlmAgent
from google.adk.models.lite_llm import LiteLlm
from pydantic import BaseModel, Field


class ChatTranscript(BaseModel):
    meeting_notes: str = Field(description="Important notes from the transcript")
    action_items: str = Field(description="All the action items from the transcript")


load_dotenv()
model = "openai/gpt-4o"
model_abstraction = LiteLlm(model=model)

note_extractor_agent = LlmAgent(
    name="notes_extractor",
    model=model_abstraction,
    instruction="""You are an excellent meeting note extractor. Please extract the meeting minutes  from provided 
    chat transcript.Given a transcript, respond ONLY with a JSON object """ ,
    output_schema=ChatTranscript, # Enforce JSON output
    output_key="meeting_minutes"  # Store result in state['found_capital']
    # Cannot use tools=[get_capital_city] effectively here
)

root_agent = note_extractor_agent

[bakeryproducts]

It looks like the framework does not ensure to create a valid JSON model. If I do the same with Open AI SDK's structured output, it works perfectly fine

My experience as well but with a gemini 2.5pro. openai-api structured output is (feeling) guaranteed
Not sure if its a gemini problem, litellm or ADK problem?

My typical error is with LLM output "" instead of something like thought=""

pydantic_core._pydantic_core.ValidationError: 1 validation error for Thinker
  Invalid JSON: EOF while parsing a value at line 1 column 0 [type=json_invalid, input_value='', input_type=str]
    For further information visit https://errors.pydantic.dev/2.11/v/json_invalid

in llm_agent.py

imo there should be a callback right after model.
Current after_model_callback is at least after pydantic scheme check (with write-to-state key output_key set in LlmAgent)

[BrownsugarZeer]

Not really, the produced JSON format will directly pass into model_validate_json without error catching. Once received invalid format, it raises an error instead of going to after_model_callback.

[101scholar]

It seems that there is a bug as for output_schema. It works with gemini-2.5-flash, but no structured output as for LiteLlm("openai/gpt-4.1").

[syangx39]

I followed the sample as in https://github.com/google/adk-python/blob/main/contributing/samples/fields_output_schema/agent.py, with model changed to model="hosted_vllm/meta-llama/Llama-3.1-8B-Instruct" via LiteLlm -- the output_schema did not work for me.