Yeah, so there are definitely ways to have more control. I have videos on nested chatting, sequential chatting, and something not so new now I guess, but with Stateflow. Will help have more control
I like the way you explain things! ... Now comes the "but"... I think it'd really elevate the content if you would close out the theoretical portion of your videos with an example application of the whole concept. So, don't remove anything from the explanation segment, but round it out with a clearly defined sample project to see if all works together, and especially, what makes using these functions more valuable compared to just running the agent. I get that this is sort of already in the video, but it didn't really become clear. Also, it'd be SO awesome, if you, or somebody, would put together a video that combines the currently best agent framework and memgpt (or other memory), gets it all working locally (not local LLM, still using oai), and then shows how one could actually use that to create larger coding projects....
I appreciate what you are saying and I understand. I think that makes sense and especially for this an example would have been nice. Rounding out the video like that would be helpful 👍. Yes I’m trying to work on larger coding projects, and just make them SAAS projects. Which would basically be trying to put all the things together into a real project
Hi Tyler, Thanks for the wonderful work and the helpful content. Could you also add a use case where we can call external APIs (via REST Endpoints) using AutoGen and perform some activities. For example - I want get list of items from a website by sending a GET request OR add a new item by send a POST request with required parameters. Authorization code can also be provided thorough an env variable.
Hey thank you I appreciate it 🙌. Yeah, I have plans to soon have a remote server with an open-source llm, and create a simple application with rest apis. Then I could call the apis from somewhere else. Like a simple production app.
Hi Tyler, great work and channel! Learning a lot from you 👏 Facing one challenge: is it always the case that userproxy agent gets the function to execute? Then what happened to the whole role playing and each agent getting their own tasks to do (read function to execute). Asking because my function call is not triggering at all for a Data Engineer agent that is supposed to fetch data from sql. Thank you for your help!
Hey, thank you so much I am glad you are learning :D. So yes it seems that you have to register the functions to the userproxy in order for them to be 'known'. Then from there when the chat is initiated, the UserAgent would know which agent (say in a groupchat) would use the next function in the list that is registered. Or if you just initiated a chat with a DE agent, then when the function to fetch data (maybe it's fetch_sql_data) is next to be called, the UserAgent is supposed to chat with the DE Agent and the DE Agent executes the function. I hope that made sense, there have been changes recently with them, and I am also testing them out more.
@@TylerReedAI yeah thanks! for some reason for me, not a single time the function call is kicking off. I even copied your code as is, copied one of original small projects on git, and it is still not working. Only difference though is I'm using a local LLM with the help of LM studio. But it is hard to believe that it matters...
i am getting this error "[{'type': 'extra_forbidden', 'loc': ('body', 'functions'), 'msg': 'Extra inputs are not permitted', 'input': [{'name': 'research', 'description': 'research about a given topic, return the research material including reference links', 'parameters': {'type': 'object', 'properties': {'query': {'type': 'string', 'description': 'The topic to be researched about'}},
Hello, thanks for the lesson. But there is one thing, for some reason all the bloggers on TH-cam show how AIs talk to each other. But none of you taught a lesson on how to connect to an API like Google ADS to collect keywords, Google Drive, Google Spreadsheet or Calendar. I would be grateful for a lesson in which this is not just the chatter of artificial intelligence, but a whole work that replaces this or that activity. This will be interesting, but simply creating AI that are useless, I don’t understand the point of these agents
Hey thank you, and I totally understand where you are coming from. These are good ideas, and I have more coming. I think sometimes it's easier to get caught up in the 'how to setup' and maybe more implementations for real-world would be better. Thank you for your input, I appreciate it! I do have some GPT-Vision videos coming out where I work with simple web scraping (for intro) but will have more in depth implementations in the near future.
@@TylerReedAI Thank you, I will be very grateful, for business you need to make more complex agent connections, simple options, everyone talks about them, but I don’t understand that this is of little use in business. We need more interactions with external APIs so that the agents are truly cool and capable assistants,
Thanks for the video, really appreciate it, I have a question about Open Source LLM integration with this code, if I want to do that I just need to do the steps on your previous video and then add the model to the JSON file at the beginning I assume. But for the functions calling, first where should I put the functions , is the same as OpenAI functions list or there is a different way. Second , do you think Mistral Instruct 7b will be able to do complex tasks that have many function calls like connecting to a database and drawing graphs etc.. Thanks again wish you all the best
Thank you for spending the time to watch! Yeah, so it's fine to add all models to the json file. Even if you don't use them all, it's okay. I will be making a video also that has multiple LLM's with different base urls (config has since been renamed) and API's. So multiple roles with use different models. The Function calling can go inside the main python file, however...my video was to keep it somewhat simple as an intro, but in my opinion, I would have a separate function file that will house each function needed, then use each one in the main python file so it's less clutter. I don't have a lot of knowledge personally on Mistral, however...I've watched a couple other people discuss it and it seems like connecting to a DB, parsing the information, and making connections like you're mentioning would be possible. I saw some statistics where the Mistral 7B with fine-tuning can be much better than GPT-4. But, you know with whatever task it would be designed for. I hope some of this helped
Great explaination. I wish I had more control over the conversation structure though.
Yeah, so there are definitely ways to have more control. I have videos on nested chatting, sequential chatting, and something not so new now I guess, but with Stateflow. Will help have more control
Thanks for the answer. Im trying stateflow right now.
Very useful, thanks!
You are welcome! 🙌
I like the way you explain things! ... Now comes the "but"... I think it'd really elevate the content if you would close out the theoretical portion of your videos with an example application of the whole concept. So, don't remove anything from the explanation segment, but round it out with a clearly defined sample project to see if all works together, and especially, what makes using these functions more valuable compared to just running the agent. I get that this is sort of already in the video, but it didn't really become clear. Also, it'd be SO awesome, if you, or somebody, would put together a video that combines the currently best agent framework and memgpt (or other memory), gets it all working locally (not local LLM, still using oai), and then shows how one could actually use that to create larger coding projects....
I appreciate what you are saying and I understand. I think that makes sense and especially for this an example would have been nice. Rounding out the video like that would be helpful 👍. Yes I’m trying to work on larger coding projects, and just make them SAAS projects. Which would basically be trying to put all the things together into a real project
Hi Tyler, Thanks for the wonderful work and the helpful content. Could you also add a use case where we can call external APIs (via REST Endpoints) using AutoGen and perform some activities. For example - I want get list of items from a website by sending a GET request OR add a new item by send a POST request with required parameters. Authorization code can also be provided thorough an env variable.
Hey thank you I appreciate it 🙌. Yeah, I have plans to soon have a remote server with an open-source llm, and create a simple application with rest apis. Then I could call the apis from somewhere else. Like a simple production app.
Hi Tyler, great work and channel! Learning a lot from you 👏
Facing one challenge: is it always the case that userproxy agent gets the function to execute? Then what happened to the whole role playing and each agent getting their own tasks to do (read function to execute). Asking because my function call is not triggering at all for a Data Engineer agent that is supposed to fetch data from sql. Thank you for your help!
Hey, thank you so much I am glad you are learning :D.
So yes it seems that you have to register the functions to the userproxy in order for them to be 'known'. Then from there when the chat is initiated, the UserAgent would know which agent (say in a groupchat) would use the next function in the list that is registered. Or if you just initiated a chat with a DE agent, then when the function to fetch data (maybe it's fetch_sql_data) is next to be called, the UserAgent is supposed to chat with the DE Agent and the DE Agent executes the function.
I hope that made sense, there have been changes recently with them, and I am also testing them out more.
@@TylerReedAI yeah thanks! for some reason for me, not a single time the function call is kicking off. I even copied your code as is, copied one of original small projects on git, and it is still not working. Only difference though is I'm using a local LLM with the help of LM studio. But it is hard to believe that it matters...
What's the difference between copying a model and creating from a model?
i am getting this error "[{'type': 'extra_forbidden', 'loc': ('body', 'functions'), 'msg': 'Extra inputs are not permitted', 'input': [{'name': 'research', 'description': 'research about a given topic, return the research material including reference links', 'parameters': {'type': 'object', 'properties': {'query': {'type': 'string', 'description': 'The topic to be researched about'}},
Hello, thanks for the lesson. But there is one thing, for some reason all the bloggers on TH-cam show how AIs talk to each other. But none of you taught a lesson on how to connect to an API like Google ADS to collect keywords, Google Drive, Google Spreadsheet or Calendar. I would be grateful for a lesson in which this is not just the chatter of artificial intelligence, but a whole work that replaces this or that activity. This will be interesting, but simply creating AI that are useless, I don’t understand the point of these agents
Hey thank you, and I totally understand where you are coming from. These are good ideas, and I have more coming. I think sometimes it's easier to get caught up in the 'how to setup' and maybe more implementations for real-world would be better. Thank you for your input, I appreciate it! I do have some GPT-Vision videos coming out where I work with simple web scraping (for intro) but will have more in depth implementations in the near future.
@@TylerReedAI Thank you, I will be very grateful, for business you need to make more complex agent connections, simple options, everyone talks about them, but I don’t understand that this is of little use in business. We need more interactions with external APIs so that the agents are truly cool and capable assistants,
Thanks for the video, really appreciate it, I have a question about Open Source LLM integration with this code, if I want to do that I just need to do the steps on your previous video and then add the model to the JSON file at the beginning I assume.
But for the functions calling, first where should I put the functions , is the same as OpenAI functions list or there is a different way. Second , do you think Mistral Instruct 7b will be able to do complex tasks that have many function calls like connecting to a database and drawing graphs etc..
Thanks again wish you all the best
Thank you for spending the time to watch! Yeah, so it's fine to add all models to the json file. Even if you don't use them all, it's okay. I will be making a video also that has multiple LLM's with different base urls (config has since been renamed) and API's. So multiple roles with use different models. The Function calling can go inside the main python file, however...my video was to keep it somewhat simple as an intro, but in my opinion, I would have a separate function file that will house each function needed, then use each one in the main python file so it's less clutter.
I don't have a lot of knowledge personally on Mistral, however...I've watched a couple other people discuss it and it seems like connecting to a DB, parsing the information, and making connections like you're mentioning would be possible. I saw some statistics where the Mistral 7B with fine-tuning can be much better than GPT-4. But, you know with whatever task it would be designed for.
I hope some of this helped
@@TylerReedAI thanks a lot, really appreciate it
wut, so is it good
Yes I think functions help and allow more freedom to do things with the llm responses
Idk if this software even does@@TylerReedAI