braintrust 0.4.3__tar.gz → 0.5.0__tar.gz

{braintrust-0.4.3 → braintrust-0.5.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: braintrust
-Version: 0.4.3
+Version: 0.5.0
 Summary: SDK for integrating Braintrust
 Home-page: https://www.braintrust.dev
 Author: Braintrust

{braintrust-0.4.3 → braintrust-0.5.0}/src/braintrust/_generated_types.py

@@ -144,6 +144,11 @@ class AsyncScoringControlAsyncScoringControl5(TypedDict):
     triggered_xact_id: str
 
 
+class AsyncScoringControlAsyncScoringControl6(TypedDict):
+    kind: Literal['mark_attempt_failed']
+    function_ids: Sequence[Any]
+
+
 class AsyncScoringStateAsyncScoringState(TypedDict):
     status: Literal['enabled']
     token: str
@@ -484,6 +489,10 @@ class Dataset(TypedDict):
     """
     User-controlled metadata about the dataset
     """
+    url_slug: str
+    """
+    URL slug for the dataset. used to construct dataset URLs
+    """
 
 
 class DatasetEventMetadata(TypedDict):
@@ -532,6 +541,43 @@ class EnvVar(TypedDict):
     """
 
 
+class EvalStatusPageConfig(TypedDict):
+    score_columns: NotRequired[Sequence[str] | None]
+    """
+    The score columns to display on the page
+    """
+    metric_columns: NotRequired[Sequence[str] | None]
+    """
+    The metric columns to display on the page
+    """
+    grouping_field: NotRequired[str | None]
+    """
+    The metadata field to use for grouping experiments (model)
+    """
+    filter: NotRequired[str | None]
+    """
+    BTQL filter to apply to experiment data
+    """
+    sort_by: NotRequired[str | None]
+    """
+    Field to sort results by (format: 'score:<name>' or 'metric:<name>')
+    """
+    sort_order: NotRequired[Literal['asc', 'desc'] | None]
+    """
+    Sort order (ascending or descending)
+    """
+    api_key: NotRequired[str | None]
+    """
+    The API key used for fetching experiment data
+    """
+
+
+EvalStatusPageTheme: TypeAlias = Literal['light', 'dark']
+"""
+The theme for the page
+"""
+
+
 class ExperimentEventMetadata(TypedDict):
     model: NotRequired[str | None]
     """
@@ -749,20 +795,24 @@ FunctionIdRef: TypeAlias = Mapping[str, Any]
 
 
 FunctionObjectType: TypeAlias = Literal[
-    'prompt', 'tool', 'scorer', 'task', 'custom_view', 'preprocessor', 'facet'
+    'prompt', 'tool', 'scorer', 'task', 'workflow', 'custom_view', 'preprocessor', 'facet', 'classifier'
 ]
 
 
-FunctionOutputType: TypeAlias = Literal['completion', 'score', 'any']
+FunctionOutputType: TypeAlias = Literal['completion', 'score', 'facet', 'classification', 'any']
 
 
-FunctionTypeEnum: TypeAlias = Literal['llm', 'scorer', 'task', 'tool', 'custom_view', 'preprocessor', 'facet']
+FunctionTypeEnum: TypeAlias = Literal[
+    'llm', 'scorer', 'task', 'tool', 'custom_view', 'preprocessor', 'facet', 'classifier'
+]
 """
 The type of global function. Defaults to 'scorer'.
 """
 
 
-FunctionTypeEnumNullish: TypeAlias = Literal['llm', 'scorer', 'task', 'tool', 'custom_view', 'preprocessor', 'facet']
+FunctionTypeEnumNullish: TypeAlias = Literal[
+    'llm', 'scorer', 'task', 'tool', 'custom_view', 'preprocessor', 'facet', 'classifier'
+]
 
 
 class GitMetadataSettings(TypedDict):
@@ -1674,7 +1724,18 @@ class PromptDataNullishOrigin(TypedDict):
 class PromptParserNullish(TypedDict):
     type: Literal['llm_classifier']
     use_cot: bool
-    choice_scores: Mapping[str, float]
+    choice_scores: NotRequired[Mapping[str, float] | None]
+    """
+    Map of choices to scores (0-1). Used by scorers.
+    """
+    choice: NotRequired[Sequence[str] | None]
+    """
+    List of valid choices without score mapping. Used by classifiers that deposit output to tags.
+    """
+    allow_no_match: NotRequired[bool | None]
+    """
+    If true, adds a 'No match' option. When selected, no tag is deposited.
+    """
 
 
 class PromptSessionEvent(TypedDict):
@@ -2104,7 +2165,7 @@ class SpanScope(TypedDict):
 
 
 SpanType: TypeAlias = Literal[
-    'llm', 'score', 'function', 'eval', 'task', 'tool', 'automation', 'facet', 'preprocessor'
+    'llm', 'score', 'function', 'eval', 'task', 'tool', 'automation', 'facet', 'preprocessor', 'classifier'
 ]
 """
 Type of the span, for display purposes only
@@ -2384,6 +2445,7 @@ AsyncScoringControl: TypeAlias = (
     | AsyncScoringControlAsyncScoringControl3
     | AsyncScoringControlAsyncScoringControl4
     | AsyncScoringControlAsyncScoringControl5
+    | AsyncScoringControlAsyncScoringControl6
 )
 
 
@@ -2530,6 +2592,43 @@ class DatasetEvent(TypedDict):
     """
 
 
+class EvalStatusPage(TypedDict):
+    id: str
+    """
+    Unique identifier for the eval status page
+    """
+    project_id: str
+    """
+    Unique identifier for the project that the eval status page belongs under
+    """
+    user_id: NotRequired[str | None]
+    """
+    Identifies the user who created the eval status page
+    """
+    created: NotRequired[str | None]
+    """
+    Date of eval status page creation
+    """
+    deleted_at: NotRequired[str | None]
+    """
+    Date of eval status page deletion, or null if the eval status page is still active
+    """
+    name: str
+    """
+    Name of the eval status page
+    """
+    description: NotRequired[str | None]
+    """
+    Textual description of the eval status page
+    """
+    logo_url: NotRequired[str | None]
+    """
+    URL of the logo to display on the page
+    """
+    theme: EvalStatusPageTheme
+    config: EvalStatusPageConfig
+
+
 class Experiment(TypedDict):
     id: str
     """
@@ -3228,6 +3327,7 @@ class View(TypedDict):
         'prompts',
         'tools',
         'scorers',
+        'classifiers',
         'logs',
         'monitor',
         'for_review',

{braintrust-0.4.3 → braintrust-0.5.0}/src/braintrust/framework.py

@@ -1280,6 +1280,29 @@ async def _run_evaluator_internal(
     filters: list[Filter],
     stream: Callable[[SSEProgressEvent], None] | None = None,
     state: BraintrustState | None = None,
+):
+    # Start span cache for this eval (it's disabled by default to avoid temp files outside of evals)
+    if state is None:
+        from braintrust.logger import _internal_get_global_state
+
+        state = _internal_get_global_state()
+
+    state.span_cache.start()
+    try:
+        return await _run_evaluator_internal_impl(experiment, evaluator, position, filters, stream, state)
+    finally:
+        # Clean up disk-based span cache after eval completes and stop caching
+        state.span_cache.dispose()
+        state.span_cache.stop()
+
+
+async def _run_evaluator_internal_impl(
+    experiment,
+    evaluator: Evaluator,
+    position: int | None,
+    filters: list[Filter],
+    stream: Callable[[SSEProgressEvent], None] | None = None,
+    state: BraintrustState | None = None,
 ):
     event_loop = asyncio.get_event_loop()
 
@@ -1290,11 +1313,13 @@ async def _run_evaluator_internal(
             {**parent_propagated},
             {"span_attributes": {"purpose": "scorer"}},
         )
+        # Strip trace from logged input - it's internal plumbing that shouldn't appear in spans
+        logged_input = {k: v for k, v in kwargs.items() if k != "trace"}
         with root_span.start_span(
             name=name,
             span_attributes={"type": SpanTypeAttribute.SCORE, "purpose": "scorer"},
             propagated_event=merged_propagated,
-            input=dict(**kwargs),
+            input=logged_input,
         ) as span:
             score = scorer
             if hasattr(scorer, "eval_async"):
@@ -1415,6 +1440,77 @@ async def _run_evaluator_internal(
                 tags = hooks.tags if hooks.tags else None
                 root_span.log(output=output, metadata=metadata, tags=tags)
 
+                # Create trace object for scorers
+                from braintrust.trace import LocalTrace
+
+                async def ensure_spans_flushed():
+                    # Flush native Braintrust spans
+                    if experiment:
+                        await asyncio.get_event_loop().run_in_executor(
+                            None, lambda: experiment.state.flush()
+                        )
+                    elif state:
+                        await asyncio.get_event_loop().run_in_executor(None, lambda: state.flush())
+                    else:
+                        from braintrust.logger import flush as flush_logger
+
+                        await asyncio.get_event_loop().run_in_executor(None, flush_logger)
+
+                    # Also flush OTEL spans if registered
+                    if state:
+                        await state.flush_otel()
+
+                experiment_id = None
+                if experiment:
+                    try:
+                        experiment_id = experiment.id
+                    except:
+                        experiment_id = None
+
+                trace = None
+                if state or experiment:
+                    # Get the state to use
+                    trace_state = state
+                    if not trace_state and experiment:
+                        trace_state = experiment.state
+                    if not trace_state:
+                        # Fall back to global state
+                        from braintrust.logger import _internal_get_global_state
+
+                        trace_state = _internal_get_global_state()
+
+                    # Access root_span_id from the concrete SpanImpl instance
+                    # The Span interface doesn't expose this but SpanImpl has it
+                    root_span_id_value = getattr(root_span, "root_span_id", root_span.id)
+
+                    # Check if there's a parent in the context to determine object_type and object_id
+                    from braintrust.span_identifier_v3 import SpanComponentsV3, span_object_type_v3_to_typed_string
+
+                    parent_str = trace_state.current_parent.get()
+                    parent_components = None
+                    if parent_str:
+                        try:
+                            parent_components = SpanComponentsV3.from_str(parent_str)
+                        except Exception:
+                            # If parsing fails, parent_components stays None
+                            pass
+
+                    # Determine object_type and object_id based on parent or experiment
+                    if parent_components:
+                        trace_object_type = span_object_type_v3_to_typed_string(parent_components.object_type)
+                        trace_object_id = parent_components.object_id or ""
+                    else:
+                        trace_object_type = "experiment"
+                        trace_object_id = experiment_id or ""
+
+                    trace = LocalTrace(
+                        object_type=trace_object_type,
+                        object_id=trace_object_id,
+                        root_span_id=root_span_id_value,
+                        ensure_spans_flushed=ensure_spans_flushed,
+                        state=trace_state,
+                    )
+
                 score_promises = [
                     asyncio.create_task(
                         await_or_run_scorer(
@@ -1426,6 +1522,7 @@ async def _run_evaluator_internal(
                                 "expected": datum.expected,
                                 "metadata": metadata,
                                 "output": output,
+                                "trace": trace,
                             },
                         )
                     )

{braintrust-0.4.3 → braintrust-0.5.0}/src/braintrust/functions/invoke.py

@@ -3,7 +3,7 @@ from typing import Any, Literal, TypedDict, TypeVar, overload
 from sseclient import SSEClient
 
 from .._generated_types import FunctionTypeEnum
-from ..logger import Exportable, get_span_parent_object, login, proxy_conn
+from ..logger import Exportable, _internal_get_global_state, get_span_parent_object, login, proxy_conn
 from ..util import response_raise_for_status
 from .constants import INVOKE_API_VERSION
 from .stream import BraintrustInvokeError, BraintrustStream
@@ -243,6 +243,8 @@ def init_function(project_name: str, slug: str, version: str | None = None):
     :param version: Optional version of the function to use. Defaults to latest.
     :return: A function that can be used as a task or scorer.
     """
+    # Disable span cache since remote function spans won't be in the local cache
+    _internal_get_global_state().span_cache.disable()
 
     def f(*args: Any, **kwargs: Any) -> Any:
         if len(args) > 0:

braintrust-0.5.0/src/braintrust/functions/test_invoke.py

@@ -0,0 +1,61 @@
+"""Tests for the invoke module, particularly init_function."""
+
+
+from braintrust.functions.invoke import init_function
+from braintrust.logger import _internal_get_global_state, _internal_reset_global_state
+
+
+class TestInitFunction:
+    """Tests for init_function."""
+
+    def setup_method(self):
+        """Reset state before each test."""
+        _internal_reset_global_state()
+
+    def teardown_method(self):
+        """Clean up after each test."""
+        _internal_reset_global_state()
+
+    def test_init_function_disables_span_cache(self):
+        """Test that init_function disables the span cache."""
+        state = _internal_get_global_state()
+
+        # Cache should be disabled by default (it's only enabled during evals)
+        assert state.span_cache.disabled is True
+
+        # Enable the cache (simulating what happens during eval)
+        state.span_cache.start()
+        assert state.span_cache.disabled is False
+
+        # Call init_function
+        f = init_function("test-project", "test-function")
+
+        # Cache should now be disabled (init_function explicitly disables it)
+        assert state.span_cache.disabled is True
+        assert f.__name__ == "init_function-test-project-test-function-latest"
+
+    def test_init_function_with_version(self):
+        """Test that init_function creates a function with the correct name including version."""
+        f = init_function("my-project", "my-scorer", version="v1")
+        assert f.__name__ == "init_function-my-project-my-scorer-v1"
+
+    def test_init_function_without_version_uses_latest(self):
+        """Test that init_function uses 'latest' in name when version not specified."""
+        f = init_function("my-project", "my-scorer")
+        assert f.__name__ == "init_function-my-project-my-scorer-latest"
+
+    def test_init_function_permanently_disables_cache(self):
+        """Test that init_function permanently disables the cache (can't be re-enabled)."""
+        state = _internal_get_global_state()
+
+        # Enable the cache
+        state.span_cache.start()
+        assert state.span_cache.disabled is False
+
+        # Call init_function
+        init_function("test-project", "test-function")
+        assert state.span_cache.disabled is True
+
+        # Try to start again - should still be disabled because of explicit disable
+        state.span_cache.start()
+        assert state.span_cache.disabled is True

{braintrust-0.4.3 → braintrust-0.5.0}/src/braintrust/generated_types.py

@@ -1,4 +1,4 @@
-"""Auto-generated file (internal git SHA 87ac73f4945a47eff2d4e42775ba4dbc58854c73) -- do not modify"""
+"""Auto-generated file (internal git SHA 21146f64bf5ad1eadd3a99d186274728e25e5399) -- do not modify"""
 
 from ._generated_types import (
     Acl,
@@ -29,6 +29,9 @@ from ._generated_types import (
     Dataset,
     DatasetEvent,
     EnvVar,
+    EvalStatusPage,
+    EvalStatusPageConfig,
+    EvalStatusPageTheme,
     Experiment,
     ExperimentEvent,
     ExtendedSavedFunctionId,
@@ -136,6 +139,9 @@ __all__ = [
     "Dataset",
     "DatasetEvent",
     "EnvVar",
+    "EvalStatusPage",
+    "EvalStatusPageConfig",
+    "EvalStatusPageTheme",
     "Experiment",
     "ExperimentEvent",
     "ExtendedSavedFunctionId",

{braintrust-0.4.3 → braintrust-0.5.0}/src/braintrust/logger.py

@@ -47,12 +47,9 @@ from urllib3.util.retry import Retry
 from . import context, id_gen
 from .bt_json import bt_dumps, bt_safe_deep_copy
 from .db_fields import (
-    ASYNC_SCORING_CONTROL_FIELD,
     AUDIT_METADATA_FIELD,
     AUDIT_SOURCE_FIELD,
     IS_MERGE_FIELD,
-    MERGE_PATHS_FIELD,
-    SKIP_ASYNC_SCORING_FIELD,
     TRANSACTION_ID_FIELD,
     VALID_SOURCES,
 )
@@ -101,6 +98,14 @@ from .xact_ids import prettify_xact
 Metadata = dict[str, Any]
 DATA_API_VERSION = 2
 
+
+class DatasetRef(TypedDict, total=False):
+    """Reference to a dataset by ID and optional version."""
+
+    id: str
+    version: str
+
+
 T = TypeVar("T")
 TMapping = TypeVar("TMapping", bound=Mapping[str, Any])
 TMutableMapping = TypeVar("TMutableMapping", bound=MutableMapping[str, Any])
@@ -396,6 +401,11 @@ class BraintrustState:
             ),
         )
 
+        from braintrust.span_cache import SpanCache
+
+        self.span_cache = SpanCache()
+        self._otel_flush_callback: Any | None = None
+
     def reset_login_info(self):
         self.app_url: str | None = None
         self.app_public_url: str | None = None
@@ -452,6 +462,21 @@ class BraintrustState:
 
         return self._context_manager
 
+    def register_otel_flush(self, callback: Any) -> None:
+        """
+        Register an OTEL flush callback. This is called by the OTEL integration
+        when it initializes a span processor/exporter.
+        """
+        self._otel_flush_callback = callback
+
+    async def flush_otel(self) -> None:
+        """
+        Flush OTEL spans if a callback is registered.
+        Called during ensure_spans_flushed to ensure OTEL spans are visible in BTQL.
+        """
+        if self._otel_flush_callback:
+            await self._otel_flush_callback()
+
     def copy_state(self, other: "BraintrustState"):
         """Copy login information from another BraintrustState instance."""
         self.__dict__.update({
@@ -1297,7 +1322,7 @@ def init(
     project: str | None = None,
     experiment: str | None = None,
     description: str | None = None,
-    dataset: Optional["Dataset"] = None,
+    dataset: Optional["Dataset"] | DatasetRef = None,
     open: bool = False,
     base_experiment: str | None = None,
     is_public: bool = False,
@@ -1410,12 +1435,19 @@ def init(
             args["base_exp_id"] = base_experiment_id
         elif base_experiment is not None:
             args["base_experiment"] = base_experiment
-        else:
+        elif merged_git_metadata_settings and merged_git_metadata_settings.collect != "none":
             args["ancestor_commits"] = list(get_past_n_ancestors())
 
         if dataset is not None:
-            args["dataset_id"] = dataset.id
-            args["dataset_version"] = dataset.version
+            if isinstance(dataset, dict):
+                # Simple {"id": ..., "version": ...} dict
+                args["dataset_id"] = dataset["id"]
+                if "version" in dataset:
+                    args["dataset_version"] = dataset["version"]
+            else:
+                # Full Dataset object
+                args["dataset_id"] = dataset.id
+                args["dataset_version"] = dataset.version
 
         if is_public is not None:
             args["public"] = is_public
@@ -1446,7 +1478,11 @@ def init(
     # For experiments, disable queue size limit enforcement (unlimited queue)
     state.enforce_queue_size_limit(False)
 
-    ret = Experiment(lazy_metadata=LazyValue(compute_metadata, use_mutex=True), dataset=dataset, state=state)
+    ret = Experiment(
+        lazy_metadata=LazyValue(compute_metadata, use_mutex=True),
+        dataset=dataset if isinstance(dataset, Dataset) else None,
+        state=state,
+    )
     if set_current:
         state.current_experiment = ret
     return ret
@@ -1761,6 +1797,25 @@ def login(
         _state.login(app_url=app_url, api_key=api_key, org_name=org_name, force_login=force_login)
 
 
+def register_otel_flush(callback: Any) -> None:
+    """
+    Register a callback to flush OTEL spans. This is called by the OTEL integration
+    when it initializes a span processor/exporter.
+
+    When ensure_spans_flushed is called (e.g., before a BTQL query in scorers),
+    this callback will be invoked to ensure OTEL spans are flushed to the server.
+
+    Also disables the span cache, since OTEL spans aren't in the local cache
+    and we need BTQL to see the complete span tree (both native + OTEL spans).
+
+    :param callback: The async callback function to flush OTEL spans.
+    """
+    global _state
+    _state.register_otel_flush(callback)
+    # Disable span cache since OTEL spans aren't in the local cache
+    _state.span_cache.disable()
+
+
 def login_to_state(
     app_url: str | None = None,
     api_key: str | None = None,
@@ -2323,30 +2378,6 @@ def _enrich_attachments(event: TMutableMapping) -> TMutableMapping:
 
 
 def _validate_and_sanitize_experiment_log_partial_args(event: Mapping[str, Any]) -> dict[str, Any]:
-    # Make sure only certain keys are specified.
-    forbidden_keys = set(event.keys()) - {
-        "input",
-        "output",
-        "expected",
-        "tags",
-        "scores",
-        "metadata",
-        "metrics",
-        "error",
-        "dataset_record_id",
-        "origin",
-        "inputs",
-        "span_attributes",
-        ASYNC_SCORING_CONTROL_FIELD,
-        MERGE_PATHS_FIELD,
-        SKIP_ASYNC_SCORING_FIELD,
-        "span_id",
-        "root_span_id",
-        "_bt_internal_override_pagination_key",
-    }
-    if forbidden_keys:
-        raise ValueError(f"The following keys are not permitted: {forbidden_keys}")
-
     scores = event.get("scores")
     if scores:
         for name, score in scores.items():
@@ -3855,6 +3886,21 @@ class SpanImpl(Span):
         if serializable_partial_record.get("metrics", {}).get("end") is not None:
             self._logged_end_time = serializable_partial_record["metrics"]["end"]
 
+        # Write to local span cache for scorer access
+        # Only cache experiment spans - regular logs don't need caching
+        if self.parent_object_type == SpanObjectTypeV3.EXPERIMENT:
+            from braintrust.span_cache import CachedSpan
+
+            cached_span = CachedSpan(
+                span_id=self.span_id,
+                input=serializable_partial_record.get("input"),
+                output=serializable_partial_record.get("output"),
+                metadata=serializable_partial_record.get("metadata"),
+                span_parents=self.span_parents,
+                span_attributes=serializable_partial_record.get("span_attributes"),
+            )
+            self.state.span_cache.queue_write(self.root_span_id, self.span_id, cached_span)
+
         def compute_record() -> dict[str, Any]:
             exporter = _get_exporter()
             return dict(