Merge remote-tracking branch 'origin/main' into gatekeeper/fix-4260-1

python/python/lancedb/functions.py

@@ -256,8 +256,16 @@ class SecretBinding(_RemoteValue):
 
 
 class FunctionSignature(_RemoteValue):
+    """Ordered inputs, output, and initialization fields of a Function.
+
+    ``initialization`` lists the fields of the single row each Function
+    instance is created with. Values come from each binding, not from the
+    Function version; the tuple is empty when the Function takes none.
+    """
+
     inputs: tuple[FunctionParameter, ...]
     output: FunctionOutput
+    initialization: tuple[FunctionParameter, ...] = ()
 
 
 class PythonEnvironmentSpec(_RemoteValue):
@@ -478,6 +486,10 @@ class FunctionApplication(_OpenRemoteValue):
     inputs: tuple[ApplicationInput, ...]
     output: FunctionOutput
     columns: Mapping[str, str] = Field(default_factory=dict)
+    # Transport JSON: the service validates it against the Function's
+    # initialization schema and persists the Arrow row in the binding, so it is
+    # never hashed and floating-point values are allowed.
+    initialization: Mapping[str, Any] = Field(default_factory=dict)
 
     def _known_dict(self) -> dict[str, Any]:
         value = super()._known_dict()
@@ -560,6 +572,15 @@ class FunctionBinding(_RemoteValue):
     assignment: Optional[AssignmentMapping] = None
     input_schema: Optional[Mapping[str, Any]] = None
     output_schema: Optional[Mapping[str, Any]] = None
+    initialization: Optional[str] = None
+    """Standard base64 of the validated one-row Arrow IPC initialization stream."""
+
+    def initialization_row(self) -> Optional[dict[str, Any]]:
+        """The initialization values this binding creates instances with."""
+        if self.initialization is None:
+            return None
+        reader = pa.ipc.open_stream(base64.b64decode(self.initialization))
+        return reader.read_all().to_pylist()[0]
 
 
 class RefreshColumnResult(_RemoteValue):

python/python/tests/test_first_class_function_slice1.py

@@ -83,6 +83,18 @@ def test_public_function_values_are_in_api_reference():
             FunctionBinding,
             False,
         ),
+        (
+            "remote_initialized_function_application.json",
+            "remote_initialized_function_application.canonical.json",
+            FunctionApplication,
+            False,
+        ),
+        (
+            "remote_initialized_function_binding.json",
+            "remote_initialized_function_binding.canonical.json",
+            FunctionBinding,
+            False,
+        ),
         (
             "remote_refresh_job.json",
             "remote_refresh_result.canonical.json",
@@ -269,6 +281,35 @@ def test_binding_and_refresh_result_keep_stable_remote_fields():
     assert RefreshColumnResult.from_json(result.to_canonical_json()) == result
 
 
+def test_initialization_travels_on_the_application_and_binding():
+    application = FunctionApplication.from_json(
+        fixture("remote_initialized_function_application.json")
+    )
+    assert application.initialization["temperature"] == 0.25
+    assert application.initialization["retry"]["backoff"] == (1, 2, 4)
+    application._ensure_declarable()
+
+    binding = FunctionBinding.from_json(
+        fixture("remote_initialized_function_binding.json")
+    )
+    assert binding.initialization_row() == {
+        "model": "text-embedding-3-small",
+        "dimensions": 512,
+    }
+    assert (
+        FunctionBinding.from_json(
+            fixture("remote_function_binding.json")
+        ).initialization_row()
+        is None
+    )
+
+    signature = FunctionVersion.from_json(
+        json.dumps(job_result("remote_function_job.json"))
+    ).signature
+    assert signature.initialization == ()
+    assert "initialization" not in json.loads(signature.to_canonical_json())
+
+
 def test_function_literal_numeric_domain_matches_rust():
     with pytest.raises(ValueError, match="floating-point Function literals"):
         FunctionApplication.from_json(fixture("remote_function_application_float.json"))