Role management in the Chat Completions API

API
, ,
1

It seems like the current API docs suggest that only 3 possible roles are allowed:

user, system, assistant

Have people tried using custom roles for other entities such as “event”, “form” etc.?

Does the model understand them?

When building a Chatbot, if I have other entities that need to be in the chat history, is it then better to just dump the Chat History into a consistent string representation into a single user message and have the instructions in the system message of how to interpret them?

Approach 1:

message = [
   {"role": "system", "content": "Some instruction"},
   {"role": "user", "content": "Hi"},
   {"role": "assistant", "content": "Hello! How may I help you today"},
   {"role": "user", "content": "I want to do X"},
   {"role": "assistant", "content": "Please wait while I run an internal tool to do X"},
   {"role": "event", "content": "{'tool': 'x', 'status': 'completed'}"},
   ....
]

Approach 2:

messages = [
   {"role": "system", "content": "Some instruction on how to parse the custom chat history object"},
   {"role": "user", "content":  "
Chat History:
user: Hi
assistant: Hello, how can I help you today?
user: I want to do X.
assistant: Please what while I run an internal tool to do X
event: "{'tool': 'x', 'status': 'completed'}
"
},
]
2

There are only three types of roles available: system, user, and assistant.
However, you can include different names within the same role.

For example, you can include conversation histories of userA and userB.

completion = client.chat.completions.create(
  model="gpt-4o",
  messages=[
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "user", "name": "userA","content": "Hello!"},
    {"role": "user", "name": "userB","content": "Hi there!"}
  ]
)

print(completion.choices[0].message)

In the same way, is it possible to specify event as the value for the name key.

3 Likes
3

If you want to visualize an implementation of how, what @dignity_for_all is talking about, works in practice, here’s a short clip.

3 Likes
5

Would you recommend specifying name as “event” for the system message or the user message or the assistant message?

Does the model understand these distinctions?

6

This is amazing. How are you getting the additional metadata about the message such as message_type to the model? Do you just dump it as a JSON in the content field?

7

The simplest explanation is that I sub class from Message and add in the necessary attributes; remembering that Message class supports 16 metadata variables.

class AutoExecSubMessage(BaseMessage):
    originator:Optional[str] = Field(default="")

    # specialized for stories
    story_state:Optional[str] = Field(default="Draft")
    reference_message_id:Optional[str] = Field(default="")
    

    def update_story_state(self, story_state:str):
        self.story_state = story_state
        return super().update(story_state=self.story_state)
8

But the content field in the User message can only be a string. Do you dump it to JSON then?

9

No. It’s a string. I am not sure what thrust of the question is.

But I can’t convert an arbitary string to json. If you are asking on the way in to chat-completion, I just ignore messages that are set to be ignore.

hth

10

Sorry if I wasn’t more clear. I don’t know where your AutoExecSubMessage
fits into the content.

Do you just use the story state field or something?

Like

message = AutoExecSubMessage()

messages = [
    {'role': 'user', 'name': message.originator, 'content': message.story_state}
]
11

Oh … I see the question now.

            for message in as_thread.list_messages():
                # specialized for story
                if message.story_state == 'Draft' or message.story_state == 'Final':
                # end specialized for story

                    previous_messages.append({'role': message.role, 'content': message.content[0]['text'].value})