use gpt function

agent_jarvis/jarvis/gpt/ai_function.py

@@ -37,4 +37,7 @@ async def acall_ai_function(function: str, args: list, description: str, model:
 
     logger.debug(str(messages))
 
-    return await gpt.acreate_chat_completion(model=model, messages=messages, temperature=0)
+    msg_type, msg_content = await gpt.acreate_chat_completion(model=model, messages=messages, temperature=0)
+    if msg_type == "content":
+        return msg_content
+    return 'failed'

agent_jarvis/jarvis/gpt/gpt.py

@@ -21,6 +21,7 @@ async def acreate_chat_completion_once(
         max_tokens: int | None = None,
         deployment_id=None,
         request_timeout=40,
+        **kwargs
 ) -> str:
     """
     Create a chat completion and update the cost.
@@ -39,7 +40,8 @@ async def acreate_chat_completion_once(
             messages=messages,
             temperature=temperature,
             max_tokens=max_tokens,
-            request_timeout=request_timeout
+            request_timeout=request_timeout,
+            **kwargs
         )
     else:
         response = await openai.ChatCompletion.acreate(
@@ -47,7 +49,8 @@ async def acreate_chat_completion_once(
             messages=messages,
             temperature=temperature,
             max_tokens=max_tokens,
-            request_timeout=request_timeout
+            request_timeout=request_timeout,
+            **kwargs
         )
     if CFG.debug_mode:
         logger.debug(f"Response: {response}")
@@ -65,12 +68,13 @@ async def acreate_chat_completion(
         max_tokens: int = None,
         request_timeout: int = 40,
         num_retries=3,
-        on_single_request_timeout: Callable = None
+        on_single_request_timeout: Callable = None,
+        **kwargs
 ):
     """Create a chat completion using the OpenAI API
 
     Args:
-        messages (List[Message]): The messages to send to the chat completion
+        messages (List[dict]): The messages to send to the chat completion
         model (str, optional): The model to use. Defaults to None.
         temperature (float, optional): The temperature to use. Defaults to 0.9.
         max_tokens (int, optional): The max tokens to use. Defaults to None.
@@ -100,6 +104,7 @@ async def acreate_chat_completion(
                     temperature=temperature,
                     max_tokens=max_tokens,
                     request_timeout=request_timeout,
+                    **kwargs
                 )
             else:
                 response = await acreate_chat_completion_once(
@@ -108,6 +113,7 @@ async def acreate_chat_completion(
                     temperature=temperature,
                     max_tokens=max_tokens,
                     request_timeout=request_timeout,
+                    **kwargs
                 )
             break
         except RateLimitError:
@@ -129,5 +135,10 @@ async def acreate_chat_completion(
     if response is None:
         logger.error(f"Failed to get response from GPT after {num_retries} retries")
         raise RuntimeError(f"Failed to get response after {num_retries} retries")
-    resp = response.choices[0].message["content"]
-    return resp
+
+    choice_message = response.choices[0].message
+    content = choice_message.get("content")
+    if content is None:
+        return "function_call", {k: v for k, v in choice_message["function_call"].items()}
+    else:
+        return "content", content

agent_jarvis/jarvis/gpt/token_counter.py

@@ -2,6 +2,7 @@
 from __future__ import annotations
 
 from typing import List
+import json
 
 import tiktoken_async
 
@@ -33,7 +34,8 @@ async def count_message_tokens(
     elif model == "gpt-4":
         # !Note: gpt-4 may change over time. Returning num tokens assuming gpt-4-0314.")
         return await count_message_tokens(messages, model="gpt-4-0314")
-    elif model == "gpt-3.5-turbo-0301":
+    # TODO: OpenAI has not mention how to count tokens for 0613, thus, we use the former method
+    elif model == "gpt-3.5-turbo-0301" or model == "gpt-3.5-turbo-0613" or model == "gpt-3.5-turbo-16k-0613":
         tokens_per_message = (
             4  # every message follows <|start|>{role/name}\n{content}<|end|>\n
         )
@@ -51,6 +53,11 @@ async def count_message_tokens(
     for message in messages:
         num_tokens += tokens_per_message
         for key, value in message.items():
+            if not isinstance(value, str):
+                # TODO: Since openai does not mentioned how to count tokens of 'funciton_call',
+                #   and only string is countable, thus, if the value is not a `str` (`function_call`
+                #   field of a message), we convert it into json
+                value = json.dumps(value)
             num_tokens += len(encoding.encode(value))
             if key == "name":
                 num_tokens += tokens_per_name