Incomplete (and broken) implementation of DiscoveryEngineSearchTool

[guillaumeblaquiere]

Describe the bug
A previous great bug fix allow to use VertexAISearch with another tools:
4485379

However, testing and deep diving into it lead me to understand a lazy and partial implementation.

Let me explain.
First, the implementation assume that all Datastore use CHUNKS.

adk-python/src/google/adk/tools/discovery_engine_search_tool.py

Line 96 in 6da7274

search_result_mode=discoveryengine.SearchRequest.ContentSearchSpec.SearchResultMode.CHUNKS,

It's a bad assumption, users can/could prefer document mode.

Same for the chunk space just after.

In addition, when the result is processed, 2 problems

1. Chunks are the only possible answer
   

   adk-python/src/google/adk/tools/discovery_engine_search_tool.py

   Line 115 in 6da7274

   chunk = item.chunk

2. The structData are not taken into account
   https://cloud.google.com/generative-ai-app-builder/docs/prepare-data#storage-unstructured

The code only take into account the GCS uri (from document metadata)

adk-python/src/google/adk/tools/discovery_engine_search_tool.py

Line 121 in 6da7274

"url": chunk.document_metadata.uri,

Here the problem I have with my own search (only a chunk element as example

      "chunk": {
        "name": "projects/630841691860/locations/global/collections/default_collection/dataStores/data-store-phenix-confluence/branches/0/documents/66983245/chunks/c11",
        "id": "c11",
        "content": "<REDACTED>",
        "documentMetadata": {
          "uri": "gs://vg1np-pf-phenix-rec-7f-care-bucket/phenix-extract/2025-08-22/content/66983245.html",
          "title": "66983245",
          "structData": {
            "uri": "https://confluence.agilefabric.fr.carrefour.com/pages/viewpage.action?pageId=66983245",
            "title": "Listes de valeurs (List of Values)",
            "space": "PHENIX Team"
          }
        },
        "relevanceScore": 0.58030599355697632
      }
    }

The TRUE uri is in the structData part, but now the agent return the gs:// one. Totally useless...
(Same for the title, by the way ....)

So now, I have the choice between

• Only VertexAISearchTool agent which works correctly
• Multi tools agent whith broken search.

To Reproduce
No real streamlined way to reproduce it:

• Create a VertexAI Search environment in DOCUMENT mode to validate the CHUNKS mode fail
• Create a VertexAI Search environment in CHUNKS mode with structData as mentionned here https://cloud.google.com/generative-ai-app-builder/docs/prepare-data#storage-unstructured

Then

1. Create an agent with a VertexAISearchTool
2. Run the agent and validate the correct search with the 2 created search engines
3. Create a multi-tool agent, for instance a root agent with your previous agent as sub-agent
4. Re-run the previous tests on the 2 created search engines and validate my description

Expected behavior
A full support of discovery engine tool, even if the VertexAISeach tool disappear and only the new one live, with full configuration, and correct support of all options

Desktop (please complete the following information):

• OS: Not relevant
• Python version: 3.13.8
• ADK version: 1.16.0

Model Information:
Gemini 2.5 pro

[xuanyang15]

@guillaumeblaquiere DiscoveryEngineSearchTool was introduced as a workaround to support using VertexAISearchTool with other tools in the same agent (there is a limitation on model that built-in tools can not be used with other tools

). It will only take effect when you use VertexAISearchTool together with other tools in the same agent. It currently only supports chunk mode, I can help add more support.

In the meantime, if your agent only uses one tool, the VertexAISearchTool, it will not be replaced by the workaround DiscoveryEngineSearchTool.

[xuanyang15]

@guillaumeblaquiere Could you please let me know how you create VertexAI Search environment in DOCUMENT mode? The built-in VertexAiSearchTool doesn't provide a way for user to specify the search mode. I want to investigate how we can support the DOCUMENT mode usage without letting user specify the search mode.

[comlhj1114]

In google-adk==1.4.1, it was possible to use both VertexAiSearchTool (with a structured data store) and GoogleSearchTool simultaneously, since both are classified as search_tools.

However, in the latest version, this setup no longer works. When multiple tools are used together, the ADK automatically calls the Vertex AI Search tool through the DiscoveryEngineSearchTool.
The problem is that the DiscoveryEngineSearchTool does not support structured data stores (only works with the chunk mode), resulting in an error.

[guillaumeblaquiere]

@xuanyang15 For my first use of VertexAI search, I used the DOCUMENT mode. In fact, you can ONLY use the CHUNK mode when you use unstructured data. If you gather documents, you are not allowed to query them in CHUNK mode and the API call will fail.

Here a piece of doc: https://cloud.google.com/generative-ai-app-builder/docs/create-data-store-es#storage-periodic-sync
(bullet point 13)

Today, I always use unstructured data with chunk and custom structData, but, for a framework, it's not suitable to have this kind of opinionated implementation.

[guillaumeblaquiere]

I tried stuff and I understand it's a Gemini limitation. To have a robust solution, it might be wise to wait for Gemini 3.0 ;)

[xuanyang15]

Thank you @guillaumeblaquiere! I will keep an eye on that.

[xuanyang15]

@comlhj1114 Thanks for reporting this issue! We have added an option to enable/disable the workaround in the latest code:

• 6da7274
• 9a6b850

I will also work with the team to add more supports.