VertexAISearchSummaryTool drops SummaryWithMetadata citations/references despite summary_include_citations=True

[glaziermag]

VertexAISearchSummaryTool drops SummaryWithMetadata citations/references despite summary_include_citations=True

Package versions

• langchain-google-community: 4.0.0
• google-cloud-discoveryengine: 0.20.0
• langchain-core: 1.4.4
• Python: 3.13.12
• langchain-google main SHA: 330f2df

Minimal corpus

Three direct-upload test documents under 100 KB:

• red: FALCON-17 / R-ALLOW-101 / https://example.test/red/v1
• blue: ORCHID-29 / B-ALLOW-202 / https://example.test/blue/v1
• green: MAPLE-33 / G-ALLOW-303 / https://example.test/green/v1

Direct Discovery Engine request

• query: What are the current release codes and policy codes?
• filter: tenant: ANY("red")
• SummarySpec.include_citations: true
• SummarySpec.summary_result_count: 3
• data store type: unstructured
• serving config: Enterprise/LLM Discovery Engine app serving config

Note: I first tried a structured datastore serving config. Direct Search returned
summary_with_metadata, but skipped generation with LLM_ADDON_NOT_ENABLED, so
that was not used as the LangChain bug confirmation. The confirmed run used a
tiny direct-upload text datastore and an Enterprise Search app with
SEARCH_ADD_ON_LLM.

Direct Discovery Engine response evidence

• summary.summary_text present: True
• summary.summary_with_metadata present: True
• citation_metadata.citations present: True
• references present: True
• reference documents: ['projects/917493783080/locations/global/collections/default_collection/dataStores/summary-citation-parity-ds-r3/branches/0/documents/doc_red_v1']
• reference URIs: []

Because this was a direct raw-bytes text upload, Discovery Engine returned
references[*].document but did not populate references[*].uri.

Excerpt:

{
  "summary": {
    "summary_text": "The current red tenant release code is FALCON-17 [1]. The red policy code is R-ALLOW-101 [1].",
    "summary_with_metadata": {
      "summary": "The current red tenant release code is FALCON-17. The red policy code is R-ALLOW-101.",
      "citation_metadata": {
        "citations": [
          {"end_index": "49", "sources": [{}]},
          {"start_index": "50", "end_index": "85", "sources": [{}]}
        ]
      },
      "references": [
        {
          "document": "projects/917493783080/locations/global/collections/default_collection/dataStores/summary-citation-parity-ds-r3/branches/0/documents/doc_red_v1"
        }
      ]
    }
  }
}

Full request/response JSON is in:

/Users/gabe/Desktop/agent-search-freshness/summary_citation_parity_runs/summary_citation_parity_20260610T233854282350Z/raw_calls.jsonl

LangChain code

from langchain_google_community.vertex_ai_search import VertexAISearchSummaryTool

tool = VertexAISearchSummaryTool(
    project_id=PROJECT_ID,
    location_id="global",
    data_store_id=DATA_STORE_ID,
    engine_data_type=0,
    summary_result_count=3,
    summary_include_citations=True,
)
result = tool.run('What are the current release codes and policy codes?')

The harness set the tool's private _serving_config to the engine/app serving
config above so the direct API and LangChain calls exercised the same
Enterprise/LLM Search request. The metadata loss itself is independent of that
override: current main still has _run() return only
response.summary.summary_text.

Current main, libs/community/langchain_google_community/vertex_ai_search.py:

def _run(self, user_query: str) -> str:
    request = self._create_search_request(user_query)
    response = self._client.search(request)
    return response.summary.summary_text

LangChain actual output

• installed return type: str
• installed returned value: 'The current red tenant release code is FALCON-17 [1]. The red policy code is R-ALLOW-101 [1].'
• main return type: str
• main returned value: 'The current red tenant release code is FALCON-17 [1]. The red policy code is R-ALLOW-101 [1].'

The wrapped LangChain _client.search call received a raw SearchResponse with
summary_with_metadata, citation_metadata.citations, and references, but
VertexAISearchSummaryTool._run() returned only response.summary.summary_text.

Expected behavior

The tool should expose summary_with_metadata, citation metadata, and references,
or provide a structured artifact/return path when summary_include_citations=True.

Actual behavior

The tool discards the metadata and returns only summary_text.

Negative controls

• Direct google-cloud-discoveryengine Search API was correct for the same request.
• No ADK, ChatVertexAI, Gemini, Vertex model endpoint, Cloud Run, GCS, BigQuery,
  Document AI, website crawl, PDF import, or LangSmith telemetry was used.
• Duplicate searches for VertexAISearchSummaryTool + citations/references and
  SearchResponse SummaryWithMetadata did not find an existing matching issue.