feat(genai): support optional seed parameter

libs/genai/langchain_google_genai/_common.py

@@ -230,6 +230,15 @@ class _BaseGoogleGenerativeAI(BaseModel):
             ```
     """  # noqa: E501
 
+    seed: int | None = Field(
+        default=None,
+    )
+    """
+    Seed used in decoding.
+
+    If not set, the request uses a randomly generated seed.
+    """
+
     @property
     def lc_secrets(self) -> dict[str, str]:
         # Either could contain the API key

libs/genai/langchain_google_genai/chat_models.py

@@ -2213,6 +2213,7 @@ def _prepare_params(
                     or thinking_level is not None
                     else None
                 ),
+                "seed": self.seed,
             }.items()
             if v is not None
         }

libs/genai/langchain_google_genai/llms.py

@@ -83,6 +83,7 @@ def validate_environment(self) -> Self:
             transport=self.transport,
             additional_headers=self.additional_headers,
             safety_settings=self.safety_settings,
+            seed=self.seed,
         )
 
         return self

libs/genai/tests/integration_tests/test_chat_models.py

@@ -1384,3 +1384,34 @@ def test_chat_google_genai_invoke_respects_max_tokens(max_tokens: int) -> None:
     assert output_tokens <= max_tokens, (
         f"Expected output_tokens <= {max_tokens}, got {output_tokens}"
     )
+
+
+@pytest.mark.flaky(retries=3, delay=1)
+def test_seed_provides_reproducibility() -> None:
+    llm = ChatGoogleGenerativeAI(model=_MODEL, thinking_budget=0)
+    n_generations = 3
+
+    # Explicit seed improves reproducibility
+    actual_results = set()
+    for _ in range(n_generations):
+        result = llm.invoke(
+            "Provide a number between 0 and 100.",
+            generation_config={
+                "top_p": 1,
+                "temperature": 1.0,
+                "max_output_tokens": 10,
+                "seed": 42,
+            },
+        )
+        actual_results.add(result.content)
+    assert len(actual_results) == 1
+
+    # Lack of seed means it's very unlikely to get the same number 3 times in a row
+    actual_results = set()
+    for _ in range(n_generations):
+        result = llm.invoke(
+            "Provide a number between 0 and 100.",
+            generation_config={"top_p": 1, "temperature": 1.0, "max_output_tokens": 10},
+        )
+        actual_results.add(result.content)
+    assert len(actual_results) > 1

libs/genai/tests/unit_tests/test_chat_models.py

@@ -3170,3 +3170,20 @@ def test_kwargs_override_thinking_level() -> None:
     config = llm._prepare_params(stop=None, thinking_level="high")
     assert config.thinking_config is not None
     assert config.thinking_config.thinking_level == "high"
+
+
+def test_seed_initialization() -> None:
+    # Test explicitly provided seed
+    llm = ChatGoogleGenerativeAI(
+        model=MODEL_NAME,
+        google_api_key=SecretStr(FAKE_API_KEY),
+        seed=42,
+    )
+    assert llm.seed == 42
+
+    # Test default seed
+    llm = ChatGoogleGenerativeAI(
+        model=MODEL_NAME,
+        google_api_key=SecretStr(FAKE_API_KEY),
+    )
+    assert llm.seed is None