Anthropic tool support #370
Conversation
|
I think this is Anthropic specific #211 |
Ah, excellent! |
|
Also, while OpenAI doesn't reject it, I believe I remember them saying we should not do that. I suspect all (or the vast majority) of instruction tuning is focused on the system prompt at the top of the document, so we shouldn't expect it to work if we inject an additional system prompt mid-conversation. I'm +1 on rewriting everything to have system prompt first, always. |
SmittieC
force-pushed
the
cs/anthropic_tool_support
branch
from
a0ff29f
to
fcb9cbd
Compare
| @@ -185,7 +184,8 @@ def _build_chain(self) -> Runnable[dict[str, Any], Any]: | |||
|
|
|||
| @property | |||
| def prompt(self): | |||
| system_prompt = SystemMessagePromptTemplate.from_template(self.experiment.prompt_text) | |||
| prompt = self.experiment.prompt_text + "\nCurrent datetime: {current_time}" | |||
There was a problem hiding this comment.
While this is in my opinion not the most elegant solution, I can't think of another one except using a human message. I don't think we should use a human message for this though.
There was a problem hiding this comment.
No necessary to change right now but I think we should expand this a bit more.
Today's datetime on UTC time 2023-05-02T10:00:00+00:00, it's Tuesday.
| @@ -215,6 +215,7 @@ def _build_chain(self) -> Runnable[dict[str, Any], str]: | |||
| {"input": RunnablePassthrough()} | |||
| | RunnablePassthrough.assign(source_material=RunnableLambda(lambda x: self.source_material)) | |||
| | RunnablePassthrough.assign(participant_data=RunnableLambda(lambda x: self.participant_data)) | |||
| | RunnablePassthrough.assign(current_time=RunnableLambda(lambda x: str(timezone.now()))) | |||
There was a problem hiding this comment.
FYI: Only the agents will be using the knowlegde of the current time, but I think it would be nice if we can make all our gpt's time aware
|
In a future PR we can update the tool to use our ScheduledMessages |
| @@ -185,7 +184,8 @@ def _build_chain(self) -> Runnable[dict[str, Any], Any]: | |||
|
|
|||
| @property | |||
| def prompt(self): | |||
| system_prompt = SystemMessagePromptTemplate.from_template(self.experiment.prompt_text) | |||
| prompt = self.experiment.prompt_text + "\nCurrent datetime: {current_time}" | |||
There was a problem hiding this comment.
No necessary to change right now but I think we should expand this a bit more.
Today's datetime on UTC time 2023-05-02T10:00:00+00:00, it's Tuesday.
requirements/requirements.in
Outdated
| @@ -45,3 +45,4 @@ django-field-audit>=1.2.7 | |||
| turn-python>=0.2.0 | |||
| jinja2 | |||
| django-taggit | |||
| transformers | |||
There was a problem hiding this comment.
I got this:
Could not import transformers python package. This is needed in order to calculate get_token_ids. Please install it with
pip install transformers.
that originated from here.
There was a problem hiding this comment.
👍🏻 could you add a comment
Langchain is complaining about the pace where we add a system message with the date time. Appartently the system message must be the first message. For OpenAI agents this is not the case, so might be anthropic specific. Anycase, I'm thinking of just appending "the current date is {current_date}" to the prompt