Hi Markus, Is Function calling now deprecated? I see openai has now introduced tools to address API Assistants. Should function calling still work or tools should be adopted.
Hi Saurabh, hope you are doing well! I have heard that as well, but I did not find any information about it being deprecated. The assistant API introduces "state" into the API with threads, so it works different and function calling can be used a tool. You got a nice timing, I will release a video about the Assitant API tomorrow :)
@codingcrashcourses8533 how can you use the Functions Agent with FastAPI streaming? I'm using some previous FastAPI tutorials that were previously working fine for streaming, but no longer work with the Functions Agent. Im not sure why???
Great videos! Thank you! Would you be able to help with an example where the Open AI functions agent uses a tool to access vector db, has conversational buffer memory and is customised with a prompt template (e.g. you are a useful X etc.)? I’ve tried but langchain docs are hard to parse 😂😅
haha. I also some got some uscases where I don´t have the correct chain avaiable out of the box. I just use the normal LLMChain with a cutom prompt in this case and retrieve documents "by hand", for example top_n. History is also just a List of Classes with HumanMessage and AIMessage then where I append new messages myself. Not the perfect solution, but for custom prompts this might be the way to go currently.
@@izaac__ I use the following custom class to create a final input for the OpenAIChat class, maybe it helps you :-). class PromptBuilder: def __init__(self, system_message: str, human_message_template: PromptTemplate): self.system_message = system_message self.human_message_template = human_message_template def inject_into_prompt(self, template: PromptTemplate, **kwargs) -> str: return template.format(**kwargs) def build_final_prompt( self, messages: List[str], query: str, dokumente: str ) -> List[BaseMessage]: final_prompt = [SystemMessage(content=self.system_message)] for index, message in enumerate(messages): if index % 2 == 0: final_prompt.append(HumanMessage(content=message)) else: final_prompt.append(AIMessage(content=message)) newest_message = self.inject_into_prompt( template=self.human_message_template, query=query, dokumente=dokumente, ) final_prompt.append(HumanMessage(content=newest_message)) return final_prompt
That was not the function, that was the model. The actual output is ALWAYS returned by the LLM. The model will just decide whether to retrieve additional information from functions or not.
![ARGENTINA vs. BOLIVIA [6-0] | RESUMEN | ELIMINATORIAS SUDAMERICANAS | FECHA 10](http://i.ytimg.com/vi/C3wPtKNbAOA/mqdefault.jpg)
Hi Markus,
Is Function calling now deprecated? I see openai has now introduced tools to address API Assistants. Should function calling still work or tools should be adopted.
Hi Saurabh, hope you are doing well! I have heard that as well, but I did not find any information about it being deprecated. The assistant API introduces "state" into the API with threads, so it works different and function calling can be used a tool.
You got a nice timing, I will release a video about the Assitant API tomorrow :)
@@codingcrashcourses8533 i have notifications turned on for your videos. So looking forward to it :)
@codingcrashcourses8533 how can you use the Functions Agent with FastAPI streaming? I'm using some previous FastAPI tutorials that were previously working fine for streaming, but no longer work with the Functions Agent. Im not sure why???
Never used streaming with fastapi before, sorry :/
@@codingcrashcourses8533 ah.. darn
Great videos! Thank you! Would you be able to help with an example where the Open AI functions agent uses a tool to access vector db, has conversational buffer memory and is customised with a prompt template (e.g. you are a useful X etc.)? I’ve tried but langchain docs are hard to parse 😂😅
haha. I also some got some uscases where I don´t have the correct chain avaiable out of the box. I just use the normal LLMChain with a cutom prompt in this case and retrieve documents "by hand", for example top_n. History is also just a List of Classes with HumanMessage and AIMessage then where I append new messages myself. Not the perfect solution, but for custom prompts this might be the way to go currently.
Ok great will try that! Glad I’m not the only one haha Thank you for the quick response
@@izaac__ I use the following custom class to create a final input for the OpenAIChat class, maybe it helps you :-).
class PromptBuilder:
def __init__(self, system_message: str, human_message_template: PromptTemplate):
self.system_message = system_message
self.human_message_template = human_message_template
def inject_into_prompt(self, template: PromptTemplate, **kwargs) -> str:
return template.format(**kwargs)
def build_final_prompt(
self, messages: List[str], query: str, dokumente: str
) -> List[BaseMessage]:
final_prompt = [SystemMessage(content=self.system_message)]
for index, message in enumerate(messages):
if index % 2 == 0:
final_prompt.append(HumanMessage(content=message))
else:
final_prompt.append(AIMessage(content=message))
newest_message = self.inject_into_prompt(
template=self.human_message_template,
query=query,
dokumente=dokumente,
)
final_prompt.append(HumanMessage(content=newest_message))
return final_prompt
@@codingcrashcourses8533 awesome, thanks again :)
in your first example, given that the 'functions' list was about pizza, how was it able to return information about the capital of France?
That was not the function, that was the model. The actual output is ALWAYS returned by the LLM. The model will just decide whether to retrieve additional information from functions or not.
How is OpenAI function different from lang chain StructureTool?
One uses a special React prompt, OpenAI functions provide the native functionality from openai.
Hmm langchain feels like a hack now a days :/
Why like a hack?
@@codingcrashcourses8533 as you mentioned using openai functions feels hacky
Try to read its documentation. You will be sure it is a hack :D
Please be careful with your API key and showing it, someone may take it and abuse it, I advise changing your key now that you have made it public
Try to use it :)
@@codingcrashcourses8533 lol glad it was fake