feat(llm): GeminiLLM declares supports_structured_output = True

CHANGELOG.md

@@ -2,10 +2,15 @@
 
 ## Next
 
+
 ### Fixed
 
 - Fixed `BaseGeminiLLM`/`GeminiLLM` silently dropping token usage on every call: `invoke`/`ainvoke` (both the string and message-list paths, sync and async) built `LLMResponse` from `response.text` alone, ignoring `response.usage_metadata` entirely. `LLMResponse.usage` is now populated from it, matching `AnthropicLLM`/`OpenAILLM`/`VertexAILLM`.
 
+### Changed
+
+- (**breaking**) `BaseGeminiLLM`/`GeminiLLM.supports_structured_output` is now `True` — it already supported structured output via `response_schema`/`response_mime_type`, but never declared the capability flag, so pipeline components checking it fell back to prompt-based JSON. `SimpleKGPipeline`, `SchemaFromTextExtractor`, and `LLMEntityRelationExtractor` now enable structured output by default with `GeminiLLM`. To keep the previous behavior, construct the extractors with `use_structured_output=False`.
+
 ## 1.19.0
 
 ### Added

src/neo4j_graphrag/llm/google_genai_llm.py

@@ -87,6 +87,8 @@ class BaseGeminiLLM(LLMBase, abc.ABC):
     constructing the ``client`` SDK instance.
     """
 
+    supports_structured_output: bool = True
+
     client: "genai.Client"
 
     def __init__(

tests/unit/llm/test_google_genai_llm.py

@@ -208,6 +208,16 @@ def test_gemini_llm_is_base_gemini_llm_subclass() -> None:
     assert issubclass(GeminiLLM, BaseGeminiLLM)
 
 
+def test_gemini_llm_supports_structured_output(
+    mock_genai: Tuple[MagicMock, MagicMock],
+) -> None:
+    """GeminiLLM already supports response_schema/response_mime_type; the
+    capability flag must say so, or SimpleKGPipeline/SchemaFromTextExtractor/
+    LLMEntityRelationExtractor silently fall back to prompt-based JSON."""
+    llm = GeminiLLM("gemini-2.0-flash")
+    assert llm.supports_structured_output is True
+
+
 def test_gemini_llm_init_only_constructs_client(
     mock_genai: Tuple[MagicMock, MagicMock],
 ) -> None: