Native tool call support for Mistral AI API and topic notebook. (#2135)

* Support for Mistral AI API and topic notebook.

* formatting

* formatting

autogen/agentchat/conversable_agent.py

@@ -696,8 +696,8 @@ class ConversableAgent(LLMAgent):
                 id_key = "name"
             else:
                 id_key = "tool_call_id"
-
-            func_print = f"***** Response from calling {message['role']} \"{message[id_key]}\" *****"
+            id = message.get(id_key, "No id found")
+            func_print = f"***** Response from calling {message['role']} ({id}) *****"
             print(colored(func_print, "green"), flush=True)
             print(message["content"], flush=True)
             print(colored("*" * len(func_print), "green"), flush=True)
@@ -714,7 +714,7 @@ class ConversableAgent(LLMAgent):
             if "function_call" in message and message["function_call"]:
                 function_call = dict(message["function_call"])
                 func_print = (
-                    f"***** Suggested function Call: {function_call.get('name', '(No function name found)')} *****"
+                    f"***** Suggested function call: {function_call.get('name', '(No function name found)')} *****"
                 )
                 print(colored(func_print, "green"), flush=True)
                 print(
@@ -726,9 +726,9 @@ class ConversableAgent(LLMAgent):
                 print(colored("*" * len(func_print), "green"), flush=True)
             if "tool_calls" in message and message["tool_calls"]:
                 for tool_call in message["tool_calls"]:
-                    id = tool_call.get("id", "(No id found)")
+                    id = tool_call.get("id", "No tool call id found")
                     function_call = dict(tool_call.get("function", {}))
-                    func_print = f"***** Suggested tool Call ({id}): {function_call.get('name', '(No function name found)')} *****"
+                    func_print = f"***** Suggested tool call ({id}): {function_call.get('name', '(No function name found)')} *****"
                     print(colored(func_print, "green"), flush=True)
                     print(
                         "Arguments: \n",
@@ -1309,6 +1309,12 @@ class ConversableAgent(LLMAgent):
                 )
             for tool_call in extracted_response.get("tool_calls") or []:
                 tool_call["function"]["name"] = self._normalize_name(tool_call["function"]["name"])
+                # Remove id and type if they are not present.
+                # This is to make the tool call object compatible with Mistral API.
+                if tool_call.get("id") is None:
+                    tool_call.pop("id")
+                if tool_call.get("type") is None:
+                    tool_call.pop("type")
         return extracted_response
 
     async def a_generate_oai_reply(
@@ -1525,7 +1531,6 @@ class ConversableAgent(LLMAgent):
         message = messages[-1]
         tool_returns = []
         for tool_call in message.get("tool_calls", []):
-            id = tool_call["id"]
             function_call = tool_call.get("function", {})
             func = self._function_map.get(function_call.get("name", None), None)
             if inspect.iscoroutinefunction(func):
@@ -1543,13 +1548,24 @@ class ConversableAgent(LLMAgent):
                     loop.close()
             else:
                 _, func_return = self.execute_function(function_call)
-            tool_returns.append(
-                {
-                    "tool_call_id": id,
+            content = func_return.get("content", "")
+            if content is None:
+                content = ""
+            tool_call_id = tool_call.get("id", None)
+            if tool_call_id is not None:
+                tool_call_response = {
+                    "tool_call_id": tool_call_id,
                     "role": "tool",
-                    "content": func_return.get("content", ""),
+                    "content": content,
                 }
-            )
+            else:
+                # Do not include tool_call_id if it is not present.
+                # This is to make the tool call object compatible with Mistral API.
+                tool_call_response = {
+                    "role": "tool",
+                    "content": content,
+                }
+            tool_returns.append(tool_call_response)
         if tool_returns:
             return True, {
                 "role": "tool",

website/.gitignore

@@ -19,6 +19,8 @@ docs/topics/code-execution/*.mdx
 docs/topics/task_decomposition.mdx
 docs/topics/prompting-and-reasoning/*.mdx
 docs/topics/non-openai-models/*.mdx
+docs/topics/non-openai-models/**/*.py
+docs/topics/non-openai-models/**/*.svg
 
 # Misc
 .DS_Store

website/docs/topics/non-openai-models/about-using-nonopenai-models.md

@@ -21,7 +21,8 @@ These proxy servers can be cloud-based or running locally within your environmen
 By using cloud-based proxy servers, you are able to use models without requiring the hardware
 and software to run them.
 
-These providers can host open source/weight models, like [Hugging Face](https://huggingface.co/),
+These providers can host open source/weight models, like [Hugging Face](https://huggingface.co/)
+and [Mistral AI](https://mistral.ai/),
 or their own closed models.
 
 When cloud-based proxy servers provide an OpenAI-compatible API, using them in AutoGen
@@ -32,7 +33,8 @@ authentication which is usually handled through an API key.
 Examples of using cloud-based proxy servers providers that have an OpenAI-compatible API
 are provided below:
 
-- [together.ai example](/docs/topics/non-openai-models/cloud-togetherai)
+- [Together AI example](/docs/topics/non-openai-models/cloud-togetherai)
+- [Mistral AI example](/docs/topics/non-openai-models/cloud-mistralai)
 
 
 ### Locally run proxy servers