How does function calling with tools really work?
Вставка
- Опубліковано 9 лип 2024
- A few months ago, I made a video about how function calling works in OpenAI and then how the corresponding feature works in Ollama. There were a lot of comments that came in on that video, some of which called me an idiot or a moron for suggesting that that was how function calling works. And so in this video, I fix a mistake that I made that should make a lot of those comments go away. We'll see.
You can find the code for every video I make at github.com/technovangelist/vi.... Then find the folder name that starts with the date this video was published and a title that makes sense for what the video covers.
Be sure to sign up to my monthly newsletter at technovangelist.substack.com/...
I have a Patreon at / technovangelist
You can find the Technovangelist discord at: / discord
The Ollama discord is at / discord
(they have a pretty url because they are paying at least $100 per month for Discord. You help get more viewers to this channel and I can afford that too.) - Наука та технологія
![Lp. Последняя Реальность #107 РОДНОЙ ДОМ [Финал] • Майнкрафт](http://i.ytimg.com/vi/IK3QKzKUlHM/mqdefault.jpg)
I'm truly impressed by your explanation. As a complete beginner in this field, I found your ideas very easy to understand. You deserve a larger audience and more support. I'm grateful for experts like you who can break down complex topics and make learning accessible for newcomers like myself.
The way you explained it before is way more robust than how most frameworks/providers accomplish things with a tool use abstraction.
Love your videos! Your last vid about function calling really cleared some things up.
I used that knowledge to create a market research bot for my company! They loved it, now I've jumped from a frontend typescript tev to AI operations engineer
Nice. hope that came with a bit of a pay bump.....let me know the next thing you need and I can try to cover that too.
@technovangelist I would be stoked to see a "top 5 ollama models for different tasks" style video. I'm just using Llama 3 for everything right now.
Some tasks I would like to optimize for speed, others for depth.
When it comes to function calling with lots of highly specific parameters I find that the models that I can run in ollama are simply uncapable of following the schema i provide. Whereas openai and claude do an excellent job following a large json schema. So when it comes to a mod being "trained for function calling" I think they mean trained to follow large and strict schemas well as thats really the main difference.
You will see a difference if your function call (basically json output) needs to follow a large and strict schema. Everyones examples are too small to notice the change.
Have you tried. I just tweaked my code to use 10 parameters per function. And gave a more complicated prompt. Worked just fine. But if you are doing that you probably have bigger issues outside of the model.
I think the biggest problem for you, is that most people will read something in docs or on a blog and then claim to understand. Then attack someone even though they dont dont actually understand what they talking about. You only understand once you implement and use the functionality.
And to your 2nd point. You 100% correct theirs no reason for a model to be finetuned for function calling, i discovered function calling with gpt3.5 about 3 months after chatgpts launch.
I wish. It’s pretty clear in the docs. They just see the feature name and assume from there. Thanks for the comment.
Pure gold! Always appreciate your concise explanation and humour.
Great springboard on the subject matter. Clear, to the point.
Despite your aversion to a reasonable display mode, both of your 'tools' videos make me say 'Whoop, it's not just me. Thank you.
I have no aversion to the reasonable display mode, which of course is light mode...
The way you explain thinks..... Is soooo pedagogical . The tone and the voice nuance ..musical in the ears 😊 .
Thanks
best content creator at explaining complex things simply
Timely. I have been battling getting function calling to wotk right. Sadly, many of the examples out there don't work with different models, they seem to all assume OpenAI. I look forward to giving your approach a try!
You are correct! Function calling is actually made possible simply by the reasoning capacity, such that it is, of the model. There is nothing more than that. It is a convenient abstraction for service interactions. Instead of function calling we could just call it "if you think you need it you may ask for the following ...". BTW, this type of process reasoning is also used for agentic interactions when deciding workflows.
if your model can write code then it can call a function !
Ollama tools got merged, the day after you mentioned it :-). Thanks for the push
Nice
My favorite video of yours to date. Actually the example clarified some questions I had, so thank you. I personally hope you make more mistakes 😉
Thanks for your videos and demo code Matt... very helpful.
And sorry that some people are nasty and hateful. There os no need for that. It is sad that some people feel they have permission to vent their anger and negativity and harm others.
Hi Matt,
I watched again this video with pleasure, and it got me thinking again :-). First of all, please avoid the trap of dividing your followers into lovers and haters. You produce top-notch content, and there's no need to apologize or dramatize (appreciating you irony).
Let me delve into a point that emerged in your demo/experiment, which is, in my opinion, more significant than the function calling "issue". You verified that the majority of open-source models available on Ollama are able to produce the expected JSON. That's somewhat surprising to me. This demonstrates, as you suggested, that the OpenAI function-calling fine-tuned models are just marketing, but wait. I remember that the old GPT-3.5 OpenAI "instruct" models like "text-davinci-003" were able to produce JSON (so function calling-JSON if you will), but subsequent chat COMPLETION models (fine-tuned for lists of system/assistant/user messages) weren't! So, my guess is that OpenAI released the function-calling fine-tuned models later to correct the chat-completion fine-tuning?! Ironic again.
But, back to the Ollama models-I'm still perplexed. Are these optimized for both (at the same run-time) CHAT completion and "function calling" (aka JSON outputs)? This could maybe be a topic for another video...?
By the way, it would be kind of you to share the code on your GitHub repo as usual, but anyway the video is absolutely explanatory.
I'll take a look at the "Tools" #5284 Ollama PR. In my opinion, standardization could help the community around Ollama, even if you demonstrated that any user-made schema does the job.
Thanks always for sharing great content. I appreciate your effort.
Chapeau
Giorgio
Awesome! I am glad there are people like you to simplify and RE-explain the basics to the "writers". ☺
I really appreciate you coming and stepping on the trolls' feet. Perfect 👌
I see no reason to get excited about incompetent comments.
Just chill and explain nicely 👏
Considering openAI as the ONLY solution is not smart. In a lot of usecases you can get away with opensource models like llama3, mixtral, deepseek etc. And try not to blame Ollama its just a library to run quantized open source model locally, and give you a API interface just like OpenAI 😆
It is incredible how some think OpenAI is the only solution that deserves to exist.
that's amazingly simple, nice. I'm guessing one scenario where someone would still want an agent-framework is if the framework was a low/no-code workflow.
I'd love to see a video on whether running models with GPTQ quantization is worthwhile. Most explanations I've seen amount to "GPTQ is for GPUs, GGML is for cpus" without saying why GPTQ is completely neglected in projects like ollama, or if there is even a meaningful advantage to either at this point.
quantized models are fine .. they work as well as the original full precision in general !!
Speed is ALWAYS dependant on the system !
The libraries are creating abstractions over a document and we can forget where the abstraction layer ends and where the substrate begins.
I'm going to try out this pattern. Lots of libraries make it easy to swap out models expecting completions vs conversations... fewer libraries have a nice clean way to swap out models that handle function calling differently.
As you grow in popularity, you may experience that your closest supporters will apply the greatest scrutiny. It doesn't mean you're disliked, no matter the perception of tone.
@@emmanuelgoldstein3682 problem was only felt like it was incomplete ! .. as everybody has been giving the same incomplete tutorial ..
Welcome to the real world 🙂
I suggest a disconnected engagement approach.
Love your videos and your style.
what do you mean by disconnected approach?
@@technovangelist 'Disconnected Engagement'. Stay fully engaged with what you are working on and your goals but disconnected from trolls, detractors and negative feedback. All the best with your channel.
Got it. Thanks. Luckily the negative is a small fraction of the rest of the comments. And I don’t spend too much time on it. I had fun with this one though. Thanks for the comment.
Matt, it seems that you go manually through YT comments... would it be possible to use AI to help you with that somehow? 🤔
You make it sound like reading my comments is something I would want to avoid. Ideas come from comments. Connection comes from comments. This would be the last thing I would ever want to outsource to an ai or other human.
So glad you made this video! Could you perhaps go into why apps like the ones that assist with coding or app creation that use function calling may fail with local models, but work seamlessly with the cloud models? I think this is an area where people are struggling based on the many issues I see in Github repos.
I think a lot of folks don’t realize that function calling is possible in ollama. There are folks who seem intent on spreading the notion that function calling is more than it is. And so they kind of brute force their way through rather than taking the simpler approach. But that’s just a guess. Can you point me to some of the issues you have seen?
Matt - haters suck. You’re doing great and it’s awesome that you’re willing to share your wisdom and knowledge. Please ignore the jerks, we’re surrounded by ass holes.
If I ignore them I don’t get to do fun things like this video.
Hello Matt thanks for the updated example, i had been stuck there too, but never thought once about insulting you beause of my lack of experience 🙂.
Please can you point to the websearch tool used?
Searxng
I accept your apology.
Whew!
100% clear. thanks
Glad it helped
can you explain or some useful resources to learn more about introspection and reflecion?
Like a boss!
haters will hate, but you rock man! ty for your videos !
Is it possible to use function calling with tools with Open-Webui?
i think that ollama implementation of function calling is by forcing `{` tokens at the starting to force the model to generate function call.
correct me if i am wrong.
I don't know the details but I am 95% sure that has nothing to do with it. I am pretty sure its a gbnf grammar that was set up back in October.
Can I say that, if my prompt is clear enough, I can have function calling using any module? Since it’s just helping the software to decide which functions to call, right?
Thanks for the explanation, it’s mind blowing to me
The important part is to use format:json, and specify to output as json in the prompt.
@@technovangelist interesting, our makes me wonder how does ollama guarantee the output from any LLM model will be in json format?
I dont use OpenAI function calling at all - it's just a wrapper for JSON conversion and interpretation of the output, and I'd rather keep control of that myself to make it more portable between LLMs. Why would anyone write something that is locked to a LLM interface definition when we live in such a turbulent world. I'd encourage everyone to do the same. I cant honestly see any benefit in using the "function call" feature versus rolling your own.
What is the name of local search api, you used ?
Searxng
Great!!!
🔥🔥🔥
Sooo, that was exactly what I wrote on your other video? That you can use ANY model for this, as long as it returns json in the response. As it's the "calling party" that actually runs the code/function, it has nothing to do with the model itself. But you claimed this was "added" to later models? I am confused.
Hmm not sure what other video you are referring to. But if I said something that sounded like I suggested it was added in the model I was simply not stating what I meant clearly. Function calling was added to ollama in October or November. So later than the initial release in June. That’s what I would have meant.
@@technovangelist Ok, you wrote that it was added in Llama 2, which is a model. If you meant Ollama, it makes more sense. However, what exactly prevented me from doing this with the very first version of ollama? As long as I make my own scripts that talks directly to the ollama API, why would I not be able to "ask it to return json" and simply run functions in my script based on the response? That is the part that I still do not get. Why would any type of "support for function calling" need to be added to either the model or the "wrapper" (ollama in this case) for it to work?
If you did that at the beginning the answer would have probably been something like: “sure, here is the json: {…”. It wouldn’t have been just the json. Folks were adding instructions like no prose etc to get the model to follow the instructions
@@technovangelist That's odd. It worked perfectly fine for me to say "only respond with a json object, nothing else" even on the very first models. Anyways, doesn't really matter.
Really Gemma can do this? From the examples I've seen that model is pretty dumb, so if an SLM such as this can do function calling I'm impressed
That was gemma2 I think.
actually that bit was interesting to see that every single model produced not just the correct output but the right out put....
personal;y i have found that using such techniques means after you get your final response you will need to unload and rerload the model or clear the cache so the model can prepare for the next question ?
If you are having to unload and reload there must be something very strange with your setup. Is this with ollama? Have you updated to the latest versions? There is no need to do such things.
I dont know if i will ever be forgive you for this. How could do this to us 😂❤
lol
Yeah, I am not getting consistent function names. Model keeps changing them. Parameters are good. So for me it’s not stable no matter the model I use
interesting. would love to see the code you are running. I haven't been able to get it to fail ever.
@@technovangelist I got it now. Forgot to stringify the json object.
So basically I have it working in JavaScript, including agents
😆I felt like this was a dev version of this vid: ua-cam.com/video/0Szj21arytU/v-deo.html. Really enjoyed this one. As always, thanks for putting out great and useful content!
yeah function calling is just making llm to choose what function to use and specifying the required param as structured output. I am amazed how dumb people are ... just try to code up simple example and run it.
no not dumb .... they are many components to an ai system you can just use inputs and outputs ... but there is alot more you can do with a base model !
as we amy see a tutorial or example of your Mistral model , flying your RC helecopter !
most of youtuber doesn't care about what their viewer agrees or disagree (and how they said it) but you handle it as if they are part of a....? community... ? a.... companion along ollama adventures... ?
in the first place most youtuber doesn't care and move on to next video at the end those nasty words are just a comments and when newer videos come up those disagree-er would come back to watch newer video.... it also happen with wes roth channel, david saphiro channel, even Kamph channel....
well anyway, I've been watching UA-cam unreasonable long, i don't have local TV or Netflix all i got are smart tvs, android boxes, tablets all over my places, at the office at my room at my home at my car everywhere they all mostly 24 hour playing UA-cam videos. and matt you're the only one after all these years a youtuber whom really care and serious about what you said and the recent event was surprisingly handled in deferent degrees. you're treating your channel in different way it is interesting way of youtube-ing don't stop matt, unless you got private issue.
I think that is part of my background as an evangelist or as some companies call it, a dev advocate, though that’s a misleading name. Build a community, have conversations, relay feedback back to the team. I have incorporated so much feedback into my videos every time. Thanks for the comment and thanks for noticing.