VertexAISearchTool is not accessible

[naitik-parekh]

I have created the following agent hierarchy using ADK 1.13.0.

The code snippets of agents definition are as follows:

agent_2 = Agent()
agent_3 = Agent()
agent_4 = Agent()

agent_1 = Agent(
    sub_agents = [agent_2, agent_3, agent_4]
)

agent_6 = Agent(
    tools = [VertexAiSearchTool()]
)

agent_5 = Agent(
    tools = [AgentTool(agent=agent_6, skip_summarization=True)]
)

agent_7 = Agent(
    tools=[VertexAiSearchTool()]
)

root_agent = Agent(
    sub_agents=[
        agent_1,
        agent_5,
    ],
    tools=[
        AgentTool(agent=agent_7, skip_summarization=True)
    ],
)

The VertexAiSearchTool under agent_6 is not working. The agent_6 replies that tool is not found or not accessible.

I am familiar with the limitation that "Built-in tools cannot be used within a sub-agent". In the agent hierarchy I have created, is the VertexAiSearchTool under agent_6 not working due to this limitation? I have used AgentTool for agent_6 to mitigate the limitation. But is it the case that as agent_5 is a sub-agent, any configuration under it cannot use built-in tools?

[naitik-parekh]

I have initialized VertexAiSearchTool with data_store_id. For this configuration I do not think there is a need to map the data store to a search app. Correct me if I am wrong.

Apparently, after creating a new search app in the AI Applications Console and linking it to the data store used by agent_6's VertexAiSearchTool, the agent has been intermittently retrieving responses from the data store. On several occasions, agent_6 returns an LLM response such as, “The tool I need to use ............ is not working right now".

Is there a need to have the data store connected to any search app for it to be used in VertexAiSearchTool?

When the LLM response says that tool is not working, I do not see any errors/failures in the ADK debug logs. Is this expected behaviour? What can be checked further to find the issue?

[adk-bot]

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

TLDR: Your analysis is correct. The VertexAiSearchTool is failing due to a known limitation in ADK where built-in tools cannot be used within a sub-agent's hierarchy. Your agent_6 is nested under agent_5, which is a sub-agent, thus causing the issue.

---

Hello! Thanks for your detailed question.

Your understanding is spot on. The problem you're encountering with VertexAiSearchTool in agent_6 is indeed due to the ADK limitation that built-in tools cannot be used within a sub-agent. [1]

Let's break down your agent hierarchy:

1. root_agent is your top-level agent.
2. agent_5 is defined as a sub_agent within root_agent.
3. agent_6, which contains the VertexAiSearchTool, is used as a tool by agent_5.

Because agent_5 is a sub-agent, this restriction applies to it and any agents used as tools underneath it, like agent_6. Using AgentTool to wrap agent_6 is the correct approach for making an agent act as a tool, but it does not bypass this specific sub-agent limitation.

Your implementation of agent_7 is the correct and recommended pattern for using an agent with a built-in tool: it is added to the tools list of the root_agent (which is not a sub-agent). This works as expected.

To resolve this, you would need to restructure your hierarchy so that the agent using VertexAiSearchTool is not nested under a sub-agent. For instance, you could make the agent containing the search functionality parallel to agent_7 and add it directly to the root_agent's tools list.

[1] https://google.github.io/adk-docs/tools/built-in-tools/

[naitik-parekh]

Thank you for clarification.

Your explanation does not match my observation. As mentioned in my previous comment, the vertexAiSearchTool (with agent_6) is working intermittently after connecting the corresponding data store with a search app.

In my initial attempts, the vertexAiSearchTool (with agent_6) was not working at all. After some debugging I have found out that the data store I had created with structured data was not indexed implicitly after creation. On connecting the data store to a search app, the documents are indexed. After that, the vertexAISearchTool (with agent_6) is working intermittently.

To me it appears to be some issue around using a data store with structured data.

PS.
The vertexAiSearchTool (with agent_7) used data store with unstructured data. This is working smoothly and the documents are also indexed automatically without connect the data store to search app.

[LBUPD33]

I am encountering the same issue using ADK v1.19.

I am running a sequential agent that triggers a parallel step. This parallel step calls two different LLM agents, each using a VertexAiSearchTool connected to a specific datastore.

Current implementation

graph TD
    %% Entry Point
    User([User Query]) --> Root[root_agent<br>SequentialAgent]

    %% Main Flow
    subgraph "Sequential Execution"
        Root --> Parallel[parallel_search<br>ParallelAgent]
        Parallel --> Synthesis[synthesis_agent<br>LlmAgent]
    end

    %% Parallel Execution Details
    subgraph "Parallel Execution"
        Parallel --> InternalAgent[internal_search_agent<br>LlmAgent]
        Parallel --> ElearningAgent[elearning_search_agent<br>LlmAgent]
    end

    %% Tools and Datastores
    subgraph "Tools & Data Sources"
        InternalAgent --> |Uses| Tool1[VertexAiSearchTool<br>Internal Datastore]
        ElearningAgent --> |Uses| Tool2[VertexAiSearchTool<br>E-learning Datastore]
    end

    %% Data Flow / Context Sharing
    InternalAgent -.-> |Output: internal_search_results| Synthesis
    ElearningAgent -.-> |Output: elearning_search_results| Synthesis

    %% Final Output
    Synthesis --> Final([Final Response])

Loading

When these search agents are called in parallel, it appears that only one Vertex AI Datastore is reachable/accessed. The other one fails to retrieve data.

I am unable to use the standard DiscoveryEngine tool because my datastores are located in EU endpoints.

Has anyone found a workaround or a fix for this parallel execution issue?

[LBUPD33]

@adk-bot Any idea ?

[adk-bot]

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

TLDR: You are running into the "built-in tools cannot be used within a sub-agent" limitation. However, there's a specific workaround for VertexAiSearchTool that should solve your problem. You need to add the parameter bypass_multi_tools_limit=True to the agent that uses the tool.

---

Hello! You've correctly identified the core limitation that "Built-in tools cannot be used within a sub-agent".

In your hierarchy, agent_5 is a sub-agent of your root_agent. When agent_5 uses agent_6 via an AgentTool, agent_6 and its tools are considered part of that sub-agent's execution path. This is why the VertexAiSearchTool within agent_6 is not accessible by default.

Fortunately, the ADK provides a specific exception for GoogleSearchTool and VertexAiSearchTool. You can bypass this limitation by initializing the agent that directly holds the tool with a special parameter. [1]

To fix this, you should modify the definition of agent_6 as follows:

agent_6 = Agent(
    tools = [VertexAiSearchTool()],
    bypass_multi_tools_limit=True
)

By adding bypass_multi_tools_limit=True, you enable the VertexAiSearchTool to be used even when it's nested within a sub-agent structure. This should resolve the "tool not found" error.

[1] https://google.github.io/adk-docs/tools/built-in-tools/