Description
Currently when calling Span.log the metadata property must be a dictionary. The input property does not have this restriction and instead goes through _to_bt_safe (here) which dumps the model.
It would be nice if the same flow was applied to metadata.
Reproduction
Run the following script:
from pydantic import BaseModel
import braintrust
class M(BaseModel):
foo: str = "bar"
logger = braintrust.init_logger(project="repro", set_current=False)
with logger.start_span(name="x") as span:
span.log(input=M(), metadata=M()) # input ok, metadata raises
Expected
It should accept the Pydantic model in the metadata property.
Observed
It throws ValueError: metadata must be a dictionary.
Description
Currently when calling
Span.logthemetadataproperty must be a dictionary. Theinputproperty does not have this restriction and instead goes through_to_bt_safe(here) which dumps the model.It would be nice if the same flow was applied to
metadata.Reproduction
Run the following script:
Expected
It should accept the Pydantic model in the metadata property.
Observed
It throws
ValueError: metadata must be a dictionary.