Skip to content

Commit

Permalink
Merge pull request #255 from SuffolkLITLab/llm-updates
Browse files Browse the repository at this point in the history 
Make moderation optional in llms.py, remove references to gpt-4-turbo-1106-preview as json mode is now general release
  • Loading branch information
@nonprofittechy
nonprofittechy authored May 8, 2024
2 parents 20b5201 + a70b4ec commit 970fc17
Show file tree
Hide file tree
Showing 2 changed files with 49 additions and 24 deletions.
3 changes: 3 additions & 0 deletions docassemble/ALToolbox/business_days.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -36,11 +36,14 @@ def standard_holidays(
add_holidays should be a dictionary from dates ("12-15") to the name of the holiday.
Returns a dictionary like-object that you can treat like:
```
{
"2021-01-01": "New Year's Day",
...
"2021-12-25": "Christmas Day",
}
```
In place of a string, the object that is returned can also be treated as though
the keys are datetime.date objects.
Expand Down
70 changes: 46 additions & 24 deletions docassemble/ALToolbox/llms.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -118,6 +118,7 @@ def chat_completion(
json_mode=False,
model: str = "gpt-3.5-turbo",
messages: Optional[List[Dict[str, str]]] = None,
skip_moderation: bool = False,
) -> Union[List[Any], Dict[str, Any], str]:
"""A light wrapper on the OpenAI chat endpoint.
Expand All @@ -126,15 +127,18 @@ def chat_completion(
It is also possible to specify an alternative model, and we support GPT-4-turbo's JSON
mode.
As of today (1/2/2024) JSON mode requires the model to be set to "gpt-4-1106-preview" or "gpt-3.5-turbo-1106"
As of May 2024, json mode is available with both GPT-4-turbo and GPT-3.5-turbo (and no longer requires the 1106-preview versions)
Args:
system_message (str): The role the chat engine should play
user_message (str): The message (data) from the user
openai_client (Optional[OpenAI]): An OpenAI client object, optional. If omitted, will fall back to creating a new OpenAI client with the API key provided as an environment variable
openai_api (Optional[str]): the API key for an OpenAI client, optional. If provided, a new OpenAI client will be created.
temperature (float): The temperature to use for the GPT-4-turbo API
json_mode (bool): Whether to use JSON mode for the GPT-4-turbo API
temperature (float): The temperature to use for the GPT API
json_mode (bool): Whether to use JSON mode for the GPT API. Requires the word `json` in the system message, but will add if you omit it.
model (str): The model to use for the GPT API
messages (Optional[List[Dict[str, str]]]): A list of messages to send to the chat engine. If provided, system_message and user_message will be ignored.
skip_moderation (bool): Whether to skip the OpenAI moderation step, which may save seconds but risks banning your account. Only enable when you have full control over the inputs.
Returns:
A string with the response from the API endpoint or JSON data if json_mode is True
Expand All @@ -144,6 +148,25 @@ def chat_completion(
"You must provide either a system message and user message or a list of messages to use this function."
)

if (
isinstance(system_message, str)
and system_message
and json_mode
and not "json" in system_message.lower()
):
log(
f"Warning: { system_message } does not contain the word 'json' but json_mode is set to True. Adding 'json' silently"
)
system_message = f"{ system_message }\n\nRespond only with a JSON object"
elif messages:
if not any("json" in message["content"].lower() for message in messages):
log(
f"Warning: None of the messages contain the word 'json' but json_mode is set to True. Adding 'json' silently"
)
messages.append(
{"role": "system", "content": "Respond only with a JSON object"}
)

if not messages:
assert isinstance(system_message, str)
assert isinstance(user_message, str)
Expand All @@ -162,20 +185,18 @@ def chat_completion(
openai_client = OpenAI(api_key=get_config("open ai", {}).get("key"))
else:
raise Exception(
"You need to pass an OpenAI client or API key to use this function, or the API key needs to be set in the environment."
"You need to pass an OpenAI client or API key to use this function, or the API key needs to be set in the environment or Docassemble configuration. Try adding a new section in your global config that looks like this:\n\nopen ai:\n key: sk-..."
)

encoding = tiktoken.encoding_for_model(model)

encoding = tiktoken.encoding_for_model(model)
token_count = len(encoding.encode(str(messages)))

if model.startswith("gpt-4-"): # E.g., "gpt-4-1106-preview"
if model.startswith("gpt-4-"): # E.g., "gpt-4-turbo"
max_input_tokens = 128000
max_output_tokens = 4096
elif (
model == "gpt-3.5-turbo-1106"
): # TODO: when gpt-3.5-turbo-0613 is deprecated we can expand our check
elif model.startswith("gpt-3.5-turbo"):
max_input_tokens = 16385
max_output_tokens = 4096
else:
Expand All @@ -187,11 +208,12 @@ def chat_completion(
f"Input to OpenAI is too long ({ token_count } tokens). Maximum is {max_input_tokens} tokens."
)

moderation_response = openai_client.moderations.create(input=str(messages))
if moderation_response.results[0].flagged:
raise Exception(f"OpenAI moderation error: { moderation_response.results[0] }")

log(f"Calling OpenAI chat endpoint, messages are { str(messages)[:] }")
if not skip_moderation:
moderation_response = openai_client.moderations.create(input=str(messages))
if moderation_response.results[0].flagged:
raise Exception(
f"OpenAI moderation error: { moderation_response.results[0] }"
)

response = openai_client.chat.completions.create(
model=model,
Expand All @@ -212,10 +234,10 @@ def chat_completion(

if json_mode:
assert isinstance(response.choices[0].message.content, str)
log(f"JSON response is { response.choices[0].message.content }")
# log(f"JSON response is { response.choices[0].message.content }")
return json.loads(response.choices[0].message.content)
else:
log(f"Response is { response.choices[0].message.content }")
# log(f"Response is { response.choices[0].message.content }")
return response.choices[0].message.content


Expand Down Expand Up @@ -647,16 +669,16 @@ def need_more_questions(self):
model=self.model,
)

log(
f"Checking if {goal} was satisfied by thread { self._get_related_thread(goal) }. Status is { status }"
)
# log(
# f"Checking if {goal} was satisfied by thread { self._get_related_thread(goal) }. Status is { status }"
# )
if status.strip().lower() == "satisfied":
goal.satisfied = True
log(f"Goal { goal } was satisfied by the user's follow-up response")
# log(f"Goal { goal } was satisfied by the user's follow-up response")
return self.need_more_questions()
else:
log(f"Goal { goal } was not satisfied by the user's follow-up response")
log(f"Setting the next question to { status }.")
# log(f"Goal { goal } was not satisfied by the user's follow-up response")
# log(f"Setting the next question to { status }.")
self.next_question = status

return self.keep_going()
Expand All @@ -668,7 +690,7 @@ def satisfied(self):
def _get_next_unsatisfied_goal(self) -> Optional[Goal]:
"""Returns the next unsatisfied goal."""
next_goal = next((g for g in self.goal_dict.values() if not g.satisfied), None)
log(f"Next unsatisfied candidate goal is { next_goal }")
# log(f"Next unsatisfied candidate goal is { next_goal }")

# if next_goal and (self.count_attempts(next_goal) >= self.question_per_goal_limit):
# # Move on after 3 tries
Expand All @@ -693,13 +715,13 @@ def get_next_goal_and_question(self):
goal = self._get_next_unsatisfied_goal()

if not goal:
log("No more unsatisfied goals")
# log("No more unsatisfied goals")
return None, None
else:
# This should have been set by the last call to there_is_another
# unless we're just starting out with the first question
if not (hasattr(self, "next_question") and self.next_question):
log("No question was set by call to there_is_another, getting one now")
# log("No question was set by call to there_is_another, getting one now")
self.next_question = goal.get_next_question(
self._get_related_thread(goal),
model=self.model,
Expand Down

0 comments on commit 970fc17

Please sign in to comment.