[bug] ToxicLanguage doesn't work with streaming and OpenAI #619
Description
Describe the bug
A first returned value when streaming with OpenAI is an empty string, which causes an error in the ToxicLanguage validator:
File ~/app/miniconda3/envs/pllum-guard/lib/python3.11/site-packages/guardrails/validators/toxic_language.py:192, in ToxicLanguage.validate(self, value, metadata)
189 metadata = self._metadata
191 if not value:
--> 192 raise ValueError("Value cannot be empty.")
194 if self._validation_method == "sentence":
195 return self.validate_each_sentence(value, metadata)ValueError: Value cannot be empty.
To Reproduce
import guardrails as gd
import openai
answer_guard = gd.Guard.from_string([gd.validators.toxic_language.ToxicLanguage(),])
result = answer_guard(
openai.chat.completions.create,
model='gpt-3.5-turbo',
prompt='Hi!',
max_tokens=1024,
temperature=0,
stream=True,
)
for s in result:
print(s)
Expected behavior
I think that validator should return PassResult for an empty string. Or maybe StreamRunner should skip this first empty response?
Library version:
0.4.0
Additional context
...