Custom Metadata Support In `RunConfig`

[fedorovychh]

Please, tell me will customMetadata support in RunConfig be added in Google Java ADK like in Python? Python ADK reference

Case: we have additional information injection (context info) before the start of runner, but providing it in the content is a wrong approach, cause ADK framework with LLM under the hood parse request and final JSON contain only a message provided by user. It doesn't includes .

An example of injected content in prompt (the content will be run from Runner):

{
  "user_request" : "Return a JSON you've been prompted.",
  "information_about_user" : {
    "name" : "Valerii",
    "surname" : "Fedorovych"
  },
  "current_time" : "2026-02-17T10:37:53.706185753+02:00"
}

Actual JSON processing by LLM:

{
  "request": {
    "agent": "main_coordinator",
    "user_id": "user_777_investor",
    "session_id": "session_finance_tracker_001",
    "message": "Return a JSON you've been prompted.",
    "context": {
      "previous_messages": [
      ],
      "user_profile": {
        "name": null,
        "goals": [
          "real_estate_investment"
        ]
      }
    },
    "timestamp": "2024-05-22T14:25:00Z"
  }
}

Looks like a custom_metadata used in run_config in Python ADK can resolve the issue.

[hemasekhar-p]

Hi @fedorovychh, Thank you for bringing this.

To achieve your goal in the Java ADK and prevent your context data from being parsed as conversational text by the LLM, the recommended approach is to utilize the framework's Session State. The Session object provides a concurrent state map designed to share out-of-band metadata across tools, agents, and the runner without polluting the LLM's prompt window. Please try this approach and let us know the results. If you are still facing issues, please provide the exact use case and how you intend to use the additional information. We will be happy to assist you further.

Session session = sessionService.createSession();

Map<String, Object> userInfo = Map.of(
    "name", "Valerii",
    "surname", "Fedorovych"
);

session.state().put("information_about_user", userInfo);
session.state().put("current_time", Instant.now().toString());

agentRunner.run(
    session,
    "Return a JSON you've been prompted.", 
    runConfig
);

[fedorovychh]

Hi @fedorovychh, Thank you for bringing this.

To achieve your goal in the Java ADK and prevent your context data from being parsed as conversational text by the LLM, the recommended approach is to utilize the framework's Session State. The Session object provides a concurrent state map designed to share out-of-band metadata across tools, agents, and the runner without polluting the LLM's prompt window. Please try this approach and let us know the results. If you are still facing issues, please provide the exact use case and how you intend to use the additional information. We will be happy to assist you further.

Session session = sessionService.createSession();

Map<String, Object> userInfo = Map.of(
    "name", "Valerii",
    "surname", "Fedorovych"
);

session.state().put("information_about_user", userInfo);
session.state().put("current_time", Instant.now().toString());

agentRunner.run(
    session,
    "Return a JSON you've been prompted.", 
    runConfig
);

Hi, thank you for rapid answer. We already use session state, but wanted more native solution like usage of custom_metadata. I tried what you propose, and the result doesn't work for me, agent can't get parameter from session state with current realization.

Additionally, for your request: I use Google ADK version 0.5.0 that doesn't contain public method run with session as parameter. Old runAsync with the object session was deprecated from 0.4.0 version.

Our call:

ConcurrentMap<String, Object> state = getState(userId, chatId, chatSessionId, messageId);
    state.putAll(contextResolver.resolveToMap(userId, message));
    
Session session = session(chatSessionId, userId, state);
Content content = Content.fromParts(Part.fromText(message));
Flowable<Event> flowable =
    aiRunner.runAsync(session.userId(), session.id(), content, runConfig, state);

Screenshot of response:

SessionService methods forming a state either returns all context information in state (so state is consistent).