I watched a few tutorials but every one just went through plain code, but this man is exceptional he put in so much effort to make us really understand what is going on
I watched 4 or so tutorial videos about functions and they just really didn't explain it well. Best one so far, and I really feel I get it now. Important thing to realize when starting off is that the descriptions are of central importance, you're essentially prompting GPT in the descriptions for how to respond. Functions seem to basically just be a way to format GPTs responses in a much more rigid manner and use their AI in a more controlled way to get the kind of information you want so you can use it how you want.
"This should put you ahead of most people in understanding how these function calling methods work." I have to give it to you, you did an excellent job breaking it down. I've been working with OpenAI models for 8 months, primarily building a RAG application but also dabbling in Text to SQL. Didn't really understand how functions worked until this video. And I particularly like the fact that you went through the business logic of this thing rather than the programming specific logic. I am a PHP programmer, so seeing the logic flow visually was super-helpful. Great job. Bravo!
Wow, what an incredible video! I've watched countless tutorials and read numerous articles on OpenAI function calling, but your explanation was crystal clear and finally made everything click for me. Thank you so much for sharing your knowledge in such an understandable and engaging way. You're a lifesaver! Keep up the fantastic work!
I have a question: do I have to write code repeatedly to make LLM respond to the first, second, third, or more responses? can't LLM get the function that should be executed in a prompt at once?
Thank you! Good move and idea to be able to call functions and get results for processing with OpenAI. Opens up a lot of doors with integrating LLM to application backends.
great tutoiral and explanation. thanks! I also want to learn how you use the ipython session to view the contents of your variables, I'm just learning how to use VS Code so that is a really neat feature. Its still so abstract that its hard to wrap your brain around it. but it is what it is :D
Wow, Dave! This is absolutely awesome! Your talent and dedication are truly inspiring. Thanks a lot for sharing this incredible content with us. This is exactly what I need right now. Keep up the fantastic work, my friend! I'll definitely be eagerly waiting for more amazing videos from you. Kudos to you and your amazing talent!
This is one of the most comprehensive tutorials that breaks down function calling in a simple manner. In the conversational example, you had to manually call the function 4 times to get the response you wanted. Is there a way to do this so that we can just iterate over the responses and the AI automatically knows when to stop iterating?
Thanks for explaining the function calling principle clearly. I am wondering how I can use function calling to call a RESTful API that requires an OAuth mechanism. For example, in your demonstration, to connect to an airline and buy a ticket, the user would need to log in and authenticate to be able to purchase a ticket. It would be great if you could describe this process or at least give me some guidance. I know it is possible to use ChatGPT and define custom GPTs, using the OpenAPI specification to call a REST API over HTTP(S). I have already implemented a custom GPT that calls a REST API requiring OAuth, and when authentication is needed, it prompts the user to log in. However, I do not know how to use or implement the same thing in my project and handle the OAuth mechanism when it is needed.
Do you know how I can implement user confirmation flow? Say that I have a function with few parameters, and I want to get confirmation from user if the function can be called with the parameters
Dave, fantastic video! Thanks for sharing. I was wondering if you could make a future video about the ChatGPT - Code Interpreter. I'm really curious to know if it's reliable enough to use in my Data Science workflow. Currently, I find myself doing a lot of copying and pasting between GPT-4 and Python, so I'm excited to see if the Code Interpreter can streamline this process.
I have one question, for example, i have a json file that stores location with code name. How to train the model that first of all it need to lookup that file and get out the location code before selecting as params for get_fight_info function ?
All brilliant and excellent learning material! But at 22:08 you have written the user input very nicely yourself, with all required parameters in place. This obviously is not the real life scenario. In my experience if the user omits one of the two required parameters the system will ask for it... and forget the already supplied parameter. Anyway, I profited here by learning how the functions work, how multiple functions work, and what is that "auto" parameter. Thanks.
for the langchain example, how do you know when to stop? What if i dont know theres only 4 questions? what if there are multiple questions? how will i know when i have successfully answered all the users questions?
At 14:23, you use eval(function_call.name) to reference to function. Since there is only one function, wouldn't it be better to do without eval and call the function(**params) directly?
Huge thanks for the useful video 😊. I'm wondering whether you know how to deal with missing data within the prompt, for instance for the get flight info, imagine the user doesn't provide the destination, is the LLM has the ability to detect that, and back to the user ask him to provide missing info which is in this case the destination name?
whenever I provide output.tool_calls.name with name key it give the error that expecting object and string has been given could you please guide me regarding this
Hi, I have trained a model, but when I try to integrate it into my project, I receive an error message: "The model gpt-3.5-turbo-0125:model1234 does not exist or you do not have access." How can I overcome this issue? Can you guide me on this?
ua-cam.com/video/zulGMYg0v6U/v-deo.htmlsi=OeTp5Kf-wH_1OHH4&t=863 This feature was the reason I migrated from PyCharm to VSCode, and the best thing is I've never had to touch a bloody Jupyter notebook again (well, except for in Google Colab but that's force majeure).
That is awesome. Quick question: Is a second_completion necessary for the iteration on the user response? Can we not just use the function-call to also provide this iterated version of the response and read that out at the same time? So the question is, I guess, is the function applied AFTER the response is generated or does the function heavily restrict the response BEFORE it is generated. Given the statistical nature of the response mechanism I don't really see how it can do this beforehand and still guarantee a consistent outcome. But if it is after the response, then it is basically a way to iterate on the generated response by either restructuring or even semantically refelct on it (?)
Hey Dave! Have you gotten gpt to run a function twice? For instance if the prompt is "Create two events.. here are their details..." But the function call only creates 1 event at a time, and gpt only returns details for one at a time.
Try passing "two" as an int parameter in your function (and function_description) as an iteration amount and have the return value a list instead. I have no idea if this would work, I'm still getting a grasp on function calling.
This will definitely change things as the LangChain tools are essentially doing this under the hood. They will likely update the library to pick tools based on function calling.
I have a major issue, I automated the function calling but it does not always call the correct function, I need to leave it on auto so it decides when it is time to call a function. How can I make sure it always calls correct function? Function description is not working. It just calls the functions randomly
Thanks Dave, very educational. I am in the process of converting the code to the latest openai API (1.6) as the completion API has changed. This is a good exercise as it forces me to rewrite the syntax of the functions and of the API calls. I've got it running but the last part with Langchain isn't working, it crashes at llm.predict_messages, no idea why. Cheers.
Hi Dave, I am developing ai agent and using function calling. After getting data from api to by function. LLM need to do series of steps on function call response data but llm is only doing first manipulation step Can you help on this issue.
can u compare theme between programming helper 1990s vs CHAT GPT the programming helper there's no errors code depend on ur personal if u not believe me u can try programming helper and plus u r helping people
In real life, some times the user won t provide all the info you need for calling your function. How would you do to make GPT ask again for the missing information before calling the function.
You make one of the defined functions "get_complete_info(function_def, required_fields)" which itself calls an AI response to request this from the user. Then process the response and return the required_fields to the original function.
OMG this is soooo cool! How amazing is this?? Love it. So I just add a random function the way it should be and the calling system is expected to return the right response..... YAYYY!°!!!111🍆💦
Still super unstable( and can't help u reduce gpt requests and even consume more tokens. I hope this helps to transform original answer into answers with params: like { "message": { "role": "assistant", "content": "reply", "": "param0", "": "param01" } }
👉🏻Learn more about data science and AI: www.datalumina.io/newsletter
I watched a few tutorials but every one just went through plain code, but this man is exceptional he put in so much effort to make us really understand what is going on
I watched 4 or so tutorial videos about functions and they just really didn't explain it well. Best one so far, and I really feel I get it now. Important thing to realize when starting off is that the descriptions are of central importance, you're essentially prompting GPT in the descriptions for how to respond. Functions seem to basically just be a way to format GPTs responses in a much more rigid manner and use their AI in a more controlled way to get the kind of information you want so you can use it how you want.
"This should put you ahead of most people in understanding how these function calling methods work." I have to give it to you, you did an excellent job breaking it down. I've been working with OpenAI models for 8 months, primarily building a RAG application but also dabbling in Text to SQL. Didn't really understand how functions worked until this video. And I particularly like the fact that you went through the business logic of this thing rather than the programming specific logic. I am a PHP programmer, so seeing the logic flow visually was super-helpful. Great job. Bravo!
Great that you don't just use the default examples, many channels are guilty of that! I have just subscribed!
Wow, what an incredible video! I've watched countless tutorials and read numerous articles on OpenAI function calling, but your explanation was crystal clear and finally made everything click for me. Thank you so much for sharing your knowledge in such an understandable and engaging way. You're a lifesaver! Keep up the fantastic work!
This is also very useful for having structured data to send to a front end in RAG applications.
Amazing tutorial! The gradual scaling in complexity really helped me understand the concepts in clear terms. Thanks 😊
I have a question: do I have to write code repeatedly to make LLM respond to the first, second, third, or more responses? can't LLM get the function that should be executed in a prompt at once?
Thank voor de video man! Heb al een poosje geen openai meer gebruikt maar deze functies zijn wel echt heel interessant
honestly this is actually 10/10 thanks dude
This is a brilliant walkthrough of the concept. Couldn't have asked for a better one. Excellent video Dave!
Thank you! Good move and idea to be able to call functions and get results for processing with OpenAI. Opens up a lot of doors with integrating LLM to application backends.
Thanks man, i dont understand the documentation from the page, your video helps a lot
The first 10 minutes really helped me a lot!
great tutoiral and explanation. thanks! I also want to learn how you use the ipython session to view the contents of your variables, I'm just learning how to use VS Code so that is a really neat feature. Its still so abstract that its hard to wrap your brain around it. but it is what it is :D
Wow, Dave! This is absolutely awesome! Your talent and dedication are truly inspiring. Thanks a lot for sharing this incredible content with us. This is exactly what I need right now. Keep up the fantastic work, my friend! I'll definitely be eagerly waiting for more amazing videos from you. Kudos to you and your amazing talent!
I totally agree. I have exactly the same opinion, you wrote it very well!
What a great way to start my day, thanks!! 🙏🏻
Dave, love your videos best notification of the day, always! Finally got access to GPT-4 api! Let’s go test out some function calling!
Nice video thank you man. I'll need to dive in some more, its an interesting topic
Great video. Just came in so clutch for me. Thanks a ton Dave. I appreciate you putting this together.
🙏🏻
Thank you for a nice overview of function calling!
This is one of the most comprehensive tutorials that breaks down function calling in a simple manner. In the conversational example, you had to manually call the function 4 times to get the response you wanted. Is there a way to do this so that we can just iterate over the responses and the AI automatically knows when to stop iterating?
clear explanation of function calling. well done!
Wow, this is so cool man! I run a tech startup where I'm looking to add chatbot functionality. I would love to chat more with you on this
Hi Dave, great Function Calling tutorial video, thanks a lot!
Great tutorial, at 3:35 how were you able to view the Completion variable using a keyboard shortcut? is it some kind of extension?
Great demonstration, have like good sir'
Excellent explanation. Thank you :)
Hey Dave, can i use the function calling in assistants too? So function parameter would be set in the client.beta.threads.messages.create i think?
Hey Dave, could you please tell me, what's the shortcut key you are using to run the selected line of code directly in the interactive window?
PLEASE YESS!! hahah i was exactly looking for that comment man haha
Thanks for explaining the function calling principle clearly.
I am wondering how I can use function calling to call a RESTful API that requires an OAuth mechanism. For example, in your demonstration, to connect to an airline and buy a ticket, the user would need to log in and authenticate to be able to purchase a ticket. It would be great if you could describe this process or at least give me some guidance.
I know it is possible to use ChatGPT and define custom GPTs, using the OpenAPI specification to call a REST API over HTTP(S). I have already implemented a custom GPT that calls a REST API requiring OAuth, and when authentication is needed, it prompts the user to log in. However, I do not know how to use or implement the same thing in my project and handle the OAuth mechanism when it is needed.
Excellent tutorial, right down to the point, thank you!
Perfect video, thank you very much
this is incredible. and such a helpful video! thank you, it was just what I needed.
Thank you Dave Ebbelaar, it was very helpful. What is the difference between the tools and functions?
Thank you! :) That was very clear, thank you very much! :)
great video Dave, thank you it was very useful!
Awesome tutorial. Is there a way to make OpenAI to ask for a missing information? For instance, the user only informs the loc_origin.
Great work!
Top notch info right here.
🙏🏻
Thanks a lot! How do I call a 'function' from my assistant on the N8N?
Do you know how I can implement user confirmation flow? Say that I have a function with few parameters, and I want to get confirmation from user if the function can be called with the parameters
Can you try something like this with Langfuse ?
Dave, fantastic video! Thanks for sharing. I was wondering if you could make a future video about the ChatGPT - Code Interpreter. I'm really curious to know if it's reliable enough to use in my Data Science workflow. Currently, I find myself doing a lot of copying and pasting between GPT-4 and Python, so I'm excited to see if the Code Interpreter can streamline this process.
I have one question, for example, i have a json file that stores location with code name.
How to train the model that first of all it need to lookup that file and get out the location code before selecting as params for get_fight_info function ?
Awesome video man, thanks a ton!
Glad you liked it!
At 15:30 ish, how did it suddenly get internet access? I don't see where we enabled that.
he faked an API call, and returned some hard coded data from his function
All brilliant and excellent learning material! But at 22:08 you have written the user input very nicely yourself, with all required parameters in place. This obviously is not the real life scenario. In my experience if the user omits one of the two required parameters the system will ask for it... and forget the already supplied parameter. Anyway, I profited here by learning how the functions work, how multiple functions work, and what is that "auto" parameter. Thanks.
for the langchain example, how do you know when to stop?
What if i dont know theres only 4 questions? what if there are multiple questions? how will i know when i have successfully answered all the users questions?
At 14:23, you use eval(function_call.name) to reference to function. Since there is only one function, wouldn't it be better to do without eval and call the function(**params) directly?
Huge thanks for the useful video 😊.
I'm wondering whether you know how to deal with missing data within the prompt, for instance for the get flight info, imagine the user doesn't provide the destination, is the LLM has the ability to detect that, and back to the user ask him to provide missing info which is in this case the destination name?
Is it possible to deploy a particular assistant which I have created in the assistant builder using the assistant ID ??
Thank you .
Another banger of a video!
🙏🏻⚡️
whenever I provide output.tool_calls.name with name key it give the error that expecting object and string has been given could you please guide me regarding this
Hi,
I have trained a model, but when I try to integrate it into my project, I receive an error message: "The model gpt-3.5-turbo-0125:model1234 does not exist or you do not have access."
How can I overcome this issue? Can you guide me on this?
Thank you!
THank you for ur effort
It's my pleasure!
Great!! Thank you. This helped a lot.👍
Thanks!
its assuming the parameters but is should ask the required arguments
This will work different now for Assistants?
Hi Dave, thanks for this really good video! What is the plugin you are using in this video to show the response? (the one ends with .interactive)
ua-cam.com/video/zulGMYg0v6U/v-deo.htmlsi=OeTp5Kf-wH_1OHH4&t=863
This feature was the reason I migrated from PyCharm to VSCode, and the best thing is I've never had to touch a bloody Jupyter notebook again (well, except for in Google Colab but that's force majeure).
That is awesome. Quick question:
Is a second_completion necessary for the iteration on the user response? Can we not just use the function-call to also provide this iterated version of the response and read that out at the same time?
So the question is, I guess, is the function applied AFTER the response is generated or does the function heavily restrict the response BEFORE it is generated. Given the statistical nature of the response mechanism I don't really see how it can do this beforehand and still guarantee a consistent outcome. But if it is after the response, then it is basically a way to iterate on the generated response by either restructuring or even semantically refelct on it (?)
any way to make it more likely to call the function?
do I have to pay for function calls as well, suppose I have 100 functions, so would each time these function cost added?
Thank you Dave, do you think we will see new kind of jobs related to OpenAi?
what is the plug-in name? It’s useful
same can you please do one video on the function call on Gemini Ai ?
Can the properties be anything else other than object?
Hey Dave! Have you gotten gpt to run a function twice? For instance if the prompt is
"Create two events.. here are their details..."
But the function call only creates 1 event at a time, and gpt only returns details for one at a time.
Try passing "two" as an int parameter in your function (and function_description) as an iteration amount and have the return value a list instead. I have no idea if this would work, I'm still getting a grasp on function calling.
Great!!
You reckon then this replace the whole need of Tools in Langchain? or the need for Langchain altogether?
This will definitely change things as the LangChain tools are essentially doing this under the hood. They will likely update the library to pick tools based on function calling.
are the python files you just ran saved somewhere ?
I have a major issue, I automated the function calling but it does not always call the correct function, I need to leave it on auto so it decides when it is time to call a function. How can I make sure it always calls correct function? Function description is not working. It just calls the functions randomly
Thanks Dave, very educational. I am in the process of converting the code to the latest openai API (1.6) as the completion API has changed. This is a good exercise as it forces me to rewrite the syntax of the functions and of the API calls. I've got it running but the last part with Langchain isn't working, it crashes at llm.predict_messages, no idea why. Cheers.
hi did you figure it out how to handle this?
what plugin you are using for interactive CLI.
More info on that here: ua-cam.com/video/3sIzCFuLgIQ/v-deo.html
Hi Dave, I am developing ai agent and using function calling.
After getting data from api to by function.
LLM need to do series of steps on function call response data but llm is only doing first manipulation step
Can you help on this issue.
can u explain what's an AI Agent
Yea but how to put this into any production
can u compare theme between programming helper 1990s vs CHAT GPT the programming helper there's no errors code depend on ur personal if u not believe me u can try programming helper and plus u r helping people
is this just for python? is it posible to do this on php?
What screen recording software do you use?
Ecamm
Gold
Hollander!
In real life, some times the user won t provide all the info you need for calling your function. How would you do to make GPT ask again for the missing information before calling the function.
You make one of the defined functions "get_complete_info(function_def, required_fields)" which itself calls an AI response to request this from the user. Then process the response and return the required_fields to the original function.
Please don't use eval, result can be something unexpected and you can end up leaking unwanted parts of code/information
Thanks for the tip!
Has anyone tried this with a webhook?
These videos would be stellar if they didn't have the background music-noise which is extremely annoying.
OMG this is soooo cool! How amazing is this?? Love it. So I just add a random function the way it should be and the calling system is expected to return the right response..... YAYYY!°!!!111🍆💦
all good, but did the bg music puts me to sleep why you put it
Sleep is good for you
please cut the very high frequency when you speak, maybe you don't hear that frequency but it's horrible, cut above 15000/15500 hz, please 🙏🙏
I will look into this, thanks! I don't hear it myself.
Wtf
Please do not recommend that folks casually use the `eval` function - that's pretty dangerous!
well explained
Still super unstable( and can't help u reduce gpt requests and even consume more tokens. I hope this helps to transform original answer into answers with params: like {
"message": {
"role": "assistant",
"content": "reply",
"": "param0",
"": "param01"
}
}