chore(weave): move builtin object schemas to weave/shared

dev_docs/BuiltinObjectClasses.md

@@ -80,7 +80,7 @@ While many Weave Objects are free-form and user-defined, there is often a need f
 
 Here's how to define and use a validated base object:
 
-1. **Define your schema** (in `weave/trace_server/interface/builtin_object_classes/your_schema.py`):
+1. **Define your schema** (in `weave/shared/builtin_object_classes/your_schema.py`):
 
 ```python
 from pydantic import BaseModel
@@ -160,10 +160,10 @@ Run `make synchronize-base-object-schemas` to ensure the frontend TypeScript typ
 
 ### Architecture Flow
 
-1. Define your schema in a python file in the `weave/trace_server/interface/builtin_object_classes/test_only_example.py` directory. See `weave/trace_server/interface/builtin_object_classes/test_only_example.py` as an example.
-2. Make sure to register your schemas in `weave/trace_server/interface/builtin_object_classes/builtin_object_registry.py` by calling `register_base_object`.
+1. Define your schema in a python file in the `weave/shared/builtin_object_classes/test_only_example.py` directory. See `weave/shared/builtin_object_classes/test_only_example.py` as an example.
+2. Make sure to register your schemas in `weave/shared/builtin_object_classes/builtin_object_registry.py` by calling `register_base_object`.
 3. Run `make synchronize-base-object-schemas` to generate the frontend types.
-   - The first step (`make generate_base_object_schemas`) will run `scripts/generate_base_object_schemas.py` to generate a JSON schema in `weave/trace_server/interface/builtin_object_classes/generated/generated_builtin_object_class_schemas.json`.
+   - The first step (`make generate_base_object_schemas`) will run `scripts/generate_base_object_schemas.py` to generate a JSON schema in `weave/shared/builtin_object_classes/generated/generated_builtin_object_class_schemas.json`.
    - The second step (yarn `generate-schemas`) will read this file and use it to generate the frontend types located in `frontends/weave/src/components/PagePanelComponents/Home/Browse3/pages/wfReactInterface/generatedBuiltinObjectClasses.zod.ts`.
 4. Now, each use case uses different parts:
    1. `Python Writing`. Users can directly import these classes and use them as normal Pydantic models, which get published with `weave.publish`. The python client correct builds the requisite payload.

scripts/generate_base_object_schemas.py

@@ -3,15 +3,14 @@
 
 from pydantic import create_model
 
-from weave.trace_server.interface.builtin_object_classes.builtin_object_registry import (
+from weave.shared.builtin_object_classes.builtin_object_registry import (
     BUILTIN_OBJECT_REGISTRY,
 )
 
 OUTPUT_DIR = (
     Path(__file__).parent.parent
     / "weave"
-    / "trace_server"
-    / "interface"
+    / "shared"
     / "builtin_object_classes"
     / "generated"
 )

tests/trace/test_llm_structured_completion_model.py

@@ -193,7 +193,7 @@ def test_llm_structured_completion_model_filtering(client: WeaveClient):
 
 
 @patch(
-    "weave.trace_server.interface.builtin_object_classes.llm_structured_model.get_weave_client"
+    "weave.shared.builtin_object_classes.llm_structured_model.get_weave_client"
 )
 def test_llm_structured_completion_model_predict_text_response(mock_get_client):
     """Test the predict function with mocked LLM API response for text format."""
@@ -245,7 +245,7 @@ def test_llm_structured_completion_model_predict_text_response(mock_get_client):
 
 
 @patch(
-    "weave.trace_server.interface.builtin_object_classes.llm_structured_model.get_weave_client"
+    "weave.shared.builtin_object_classes.llm_structured_model.get_weave_client"
 )
 def test_llm_structured_completion_model_predict_json_response(mock_get_client):
     """Test the predict function with mocked LLM API response for JSON format."""
@@ -284,7 +284,7 @@ def test_llm_structured_completion_model_predict_json_response(mock_get_client):
 
 
 @patch(
-    "weave.trace_server.interface.builtin_object_classes.llm_structured_model.get_weave_client"
+    "weave.shared.builtin_object_classes.llm_structured_model.get_weave_client"
 )
 def test_llm_structured_completion_model_predict_with_template(mock_get_client):
     """Test the predict function with message templates and template variables."""
@@ -343,7 +343,7 @@ def test_llm_structured_completion_model_predict_with_template(mock_get_client):
 
 
 @patch(
-    "weave.trace_server.interface.builtin_object_classes.llm_structured_model.get_weave_client"
+    "weave.shared.builtin_object_classes.llm_structured_model.get_weave_client"
 )
 def test_llm_structured_completion_model_predict_with_config_override(mock_get_client):
     """Test the predict function with config parameter overriding defaults."""
@@ -395,7 +395,7 @@ def test_llm_structured_completion_model_predict_with_config_override(mock_get_c
 
 
 @patch(
-    "weave.trace_server.interface.builtin_object_classes.llm_structured_model.get_weave_client"
+    "weave.shared.builtin_object_classes.llm_structured_model.get_weave_client"
 )
 def test_llm_structured_completion_model_predict_error_handling(mock_get_client):
     """Test the predict function error handling."""
@@ -656,7 +656,7 @@ def test_cast_to_message():
 
 
 @patch(
-    "weave.trace_server.interface.builtin_object_classes.llm_structured_model.get_weave_client"
+    "weave.shared.builtin_object_classes.llm_structured_model.get_weave_client"
 )
 def test_llm_structured_completion_model_predict_with_prompt(
     mock_get_client, client: WeaveClient
@@ -733,7 +733,7 @@ def test_llm_structured_completion_model_predict_with_prompt(
 
 
 @patch(
-    "weave.trace_server.interface.builtin_object_classes.llm_structured_model.get_weave_client"
+    "weave.shared.builtin_object_classes.llm_structured_model.get_weave_client"
 )
 def test_llm_structured_completion_model_prompt_takes_precedence(
     mock_get_client, client: WeaveClient

weave/flow/annotation_spec.py

@@ -1,4 +1,4 @@
-from weave.trace_server.interface.builtin_object_classes import annotation_spec
+from weave.shared.builtin_object_classes import annotation_spec
 
 # Re-export:
 AnnotationSpec = annotation_spec.AnnotationSpec

weave/flow/leaderboard.py

@@ -3,9 +3,9 @@
 from dataclasses import dataclass
 from typing import Any
 
+from weave.shared.builtin_object_classes import leaderboard
 from weave.trace.refs import OpRef
 from weave.trace.weave_client import WeaveClient, get_ref
-from weave.trace_server.interface.builtin_object_classes import leaderboard
 from weave.trace_server.trace_server_interface import CallsFilter
 from weave.utils.project_id import from_project_id
 

weave/flow/saved_view.py

@@ -6,6 +6,10 @@
 from pydantic import BaseModel
 from typing_extensions import Self
 
+from weave.shared.builtin_object_classes.saved_view import Column, Pin
+from weave.shared.builtin_object_classes.saved_view import (
+    SavedView as SavedViewBase,
+)
 from weave.trace import urls
 from weave.trace.api import publish as weave_publish
 from weave.trace.api import ref as weave_ref
@@ -20,10 +24,6 @@
 from weave.trace_server import trace_server_interface as tsi
 from weave.trace_server.common_interface import SortBy
 from weave.trace_server.interface import query as tsi_query
-from weave.trace_server.interface.builtin_object_classes.saved_view import Column, Pin
-from weave.trace_server.interface.builtin_object_classes.saved_view import (
-    SavedView as SavedViewBase,
-)
 
 KNOWN_COLUMNS = [
     "id",

weave/scorers/llm_as_a_judge_scorer.py

@@ -4,13 +4,13 @@
 
 from weave.flow.scorer import Scorer
 from weave.prompt.prompt import MessagesPrompt
+from weave.shared.builtin_object_classes.llm_structured_model import (
+    LLMStructuredCompletionModel,
+)
 from weave.trace.context.weave_client_context import get_weave_client
 from weave.trace.objectify import maybe_objectify, register_object
 from weave.trace.op import op
 from weave.trace.vals import make_trace_obj
-from weave.trace_server.interface.builtin_object_classes.llm_structured_model import (
-    LLMStructuredCompletionModel,
-)
 
 
 @register_object

weave/shared/builtin_object_classes/alert_spec.py

@@ -0,0 +1,75 @@
+from typing import Literal
+
+from pydantic import BaseModel, ConfigDict, Field
+
+from weave.shared.builtin_object_classes import base_object_def
+
+
+class WeaveMetricThresholdSpec(BaseModel):
+    """Alert specification for weave metric threshold alerts.
+
+    Fields align with gorilla's WeaveMetricThresholdFilter and the
+    alert_worker's WeaveMetricFilter.  Extra keys are permitted so that
+    the schema can evolve without breaking existing stored objects.
+    """
+
+    model_config = ConfigDict(extra="allow")
+
+    alert_condition: str = Field(
+        default="THRESHOLD",
+        description="Trigger condition type (e.g. 'THRESHOLD')",
+    )
+    comparison_operator: Literal["GREATER_THAN", "LESS_THAN", "EQUAL"] = Field(
+        default="GREATER_THAN",
+        description="Direction of the threshold comparison",
+    )
+    metric_path: str = Field(
+        default="",
+        description="Dot-notation path to the metric value in the call (e.g. 'output.score')",
+    )
+    op_str: str = Field(
+        default="",
+        description="Op or scorer ref to scope this alert to; empty means all ops",
+    )
+    threshold: float = Field(
+        default=0.0,
+        description="Threshold value that triggers the alert",
+    )
+    window_size: int | None = Field(
+        default=None,
+        description="Max number of recent calls to include in the window",
+    )
+    window_duration: int | None = Field(
+        default=None,
+        description="Time window in seconds for historical calls",
+    )
+    monitor_ref: str | None = Field(
+        default=None,
+        description="Monitor ref; presence indicates alert was created from a monitor",
+    )
+    scorer_ref: str | None = Field(
+        default=None,
+        description="Scorer object ref identifying which scorer instance within a monitor to alert on. "
+        "Disambiguates scorers that share the same op class (e.g. multiple LLMAsAJudgeScorer instances). "
+        "Filterable via inputs.self on scorer calls or input_refs on CallsFilter.",
+    )
+    aggregation_function: Literal["mean", "median", "min", "max", "mode"] = Field(
+        default="mean",
+        description="Aggregation function applied to metric values within the window before threshold comparison",
+    )
+
+
+class AlertSpec(base_object_def.BaseObject):
+    op_scope: list[str] | None = Field(
+        default=None,
+        description="If provided, this alert only applies to calls from the given op refs",
+        examples=[
+            ["weave:///entity/project/op/name:digest"],
+            ["weave:///entity/project/op/name:*"],
+        ],
+    )
+
+    spec: WeaveMetricThresholdSpec = Field(
+        default_factory=WeaveMetricThresholdSpec,
+        description="Alert specification (threshold config, window config, etc.)",
+    )

weave/shared/builtin_object_classes/annotation_spec.py

@@ -0,0 +1,126 @@
+from typing import Any
+
+from pydantic import BaseModel, Field, create_model, field_validator, model_validator
+from pydantic.fields import FieldInfo
+
+from weave.shared.builtin_object_classes import base_object_def
+
+SUPPORTED_PRIMITIVES = (int, float, bool, str)
+
+
+class AnnotationSpec(base_object_def.BaseObject):
+    field_schema: dict[str, Any] = Field(
+        default={},
+        description="Expected to be valid JSON Schema. Can be provided as a dict, a Pydantic model class, a tuple of a primitive type and a Pydantic Field, or primitive type",
+        examples=[
+            # String feedback
+            {"type": "string", "maxLength": 100},
+            # Number feedback
+            {"type": "number", "minimum": 0, "maximum": 100},
+            # Integer feedback
+            {"type": "integer", "minimum": 0, "maximum": 100},
+            # Boolean feedback
+            {"type": "boolean"},
+            # Categorical feedback
+            {"type": "string", "enum": ["option1", "option2"]},
+        ],
+    )
+
+    # TODO
+    # If true, all unique creators will have their
+    # own value for this feedback type. Otherwise,
+    # by default, the value is shared and can be edited.
+    unique_among_creators: bool = False
+
+    # TODO
+    # If provided, this feedback type will only be shown
+    # when a call is generated from the given op ref
+    op_scope: list[str] | None = Field(
+        default=None,
+        examples=[
+            ["weave:///entity/project/op/name:digest"],
+            ["weave:///entity/project/op/name:*"],
+        ],
+    )
+
+    @model_validator(mode="before")
+    @classmethod
+    def preprocess_field_schema(cls, data: dict[str, Any]) -> dict[str, Any]:
+        if "field_schema" not in data:
+            return data
+
+        field_schema = data["field_schema"]
+
+        temp_field_tuple = None
+        # Handle Pydantic Field
+        if isinstance(field_schema, tuple):
+            if len(field_schema) != 2:
+                raise ValueError("Expected a tuple of length 2")
+            annotation, field = field_schema
+            if (
+                not isinstance(annotation, type)
+            ) or annotation not in SUPPORTED_PRIMITIVES:
+                raise TypeError("Expected annotation to be a primitive type")
+            if not isinstance(field, FieldInfo):
+                raise TypeError("Expected field to be a Pydantic Field")
+            temp_field_tuple = (annotation, field)
+        elif field_schema in SUPPORTED_PRIMITIVES:
+            temp_field_tuple = (field_schema, Field())
+
+        if temp_field_tuple is not None:
+            # Create a temporary model to leverage Pydantic's schema generation
+            TempModel = create_model("TempModel", field=temp_field_tuple)  # noqa: N806
+
+            schema = TempModel.model_json_schema()["properties"]["field"]
+
+            if (
+                "title" in schema and schema["title"] == "Field"
+            ):  # default title for Field
+                schema.pop("title")
+
+            data["field_schema"] = schema
+            return data
+
+        # Handle Pydantic model
+        if isinstance(field_schema, type) and issubclass(field_schema, BaseModel):
+            # Read back through `data` (typed `Any`): pydantic is invisible to
+            # the type-check env, so `issubclass` cannot narrow off bare `type`.
+            data["field_schema"] = data["field_schema"].model_json_schema()
+            return data
+
+        return data
+
+    @field_validator("field_schema")
+    def validate_field_schema(cls, schema: dict[str, Any]) -> dict[str, Any]:  # noqa: N805
+        # Imported lazily: `import jsonschema` eagerly loads every installed
+        # format-checker lib (rfc3987_syntax builds a Lark grammar, ~0.45s),
+        # so keeping it out of module scope avoids paying that on `import weave`.
+        import jsonschema
+
+        # Validate the schema
+        try:
+            jsonschema.validate(None, schema)
+        except jsonschema.exceptions.SchemaError:
+            raise
+        except jsonschema.exceptions.ValidationError:
+            pass  # we don't care that `None` does not conform
+        return schema
+
+    def value_is_valid(self, payload: Any) -> bool:
+        """Validates a payload against this annotation spec's schema.
+
+        Args:
+            payload: The data to validate against the schema
+
+        Returns:
+            bool: True if validation succeeds, False otherwise
+        """
+        # Lazy import: see note in validate_field_schema (avoids ~0.45s of
+        # jsonschema format-checker imports at `import weave` time).
+        import jsonschema
+
+        try:
+            jsonschema.validate(payload, self.field_schema)
+        except jsonschema.exceptions.ValidationError:
+            return False
+        return True

weave/shared/builtin_object_classes/base_object_def.py

@@ -0,0 +1,10 @@
+import pydantic
+
+RefStr = str
+
+
+# This is just an alternative to weave.Object for the server side.
+# I _think_ this will go away once we have the full weave system on the server
+class BaseObject(pydantic.BaseModel):
+    name: str | None = None
+    description: str | None = None