Thanks, buddy!

@vrsen No, thanks to you! 🙂 You doing an amazing job!

that’s like asking Steve jobs how to code in Linux Much rather he spends his time developing then making KZfaq videos

Thanks for support!

I’ll add some instruction on GitHub and post it on pip before next video

You are extremely KIND and THOUGHFUL for doing this. THANK YOU SIR!! @@vrsen

Thanks for watching!

Yes great idea. The challenge is in allowing agents to measure the performance of other agents. Without this it’s hard to know how to modify the prompts

It could be implemented like asking the agent that is doing the task to get his own metrics and suggest improvements for himself and then the boss to implement it.@@vrsen

Great idea, but I think the problem with limited capacities and especially a limited number of repeating user-questions is that you need to repeat those measurements a lot for the same and slightly modified questions which you can't do in a working environment because a lot of offered prompt modifications will lead to worse output, which you don't want under no circumstances in your productive setup (because you already put a lot of work into the system prompts like they are ;-).

Imho modifying finally good working system prompts are not worth of trying to optimize them more and more to gain speed. But investigation the multiple steps (agent calls) a framework takes to solve more complex tasks and then trying to reduce the steps, that will increase the overall speed an save you a lot of money if you are using paid APIs. And what also may speed up complex tasks is giving the agents tools that can do more than one simple thing but maybe two or more related tasks which may be called by parameters, with arrays as output. So instead of calling multiple agents after another to each of them get one value which the next agent processes, maybe the "tool" may get all single required answers at once, give it back and only one last agent call may aggregate all that data and present the final answer. The tools could also have a parameter with which you may say that the tool will use the gained info in a pipeline and return one value or get the answers parallel an return an array. - Using async calls in the tool when it has to generate multiple parallel answers will speed up again. But it probably really needs a very well thought out and tested system prompt for letting a model decide if and when to use a tool in parallel or pipeline mode. Ouuuu... it's all so challenging :-)

Thank you

😉

It will magically happen if you just specify the folder) Docs are coming soon for sure, community probably as well

I’ll think about it🤔

The most expensive is certainly the browsing agent. You can try remove it if costs are a problem

Yeah, try it out haha

# Chat with an intelligent assistant in your terminal from openai import OpenAI # Point to the local server client = OpenAI(base_url="localhost:1234/v1", api_key="not-needed") history = [ {"role": "system", "content": "You are an intelligent assistant. You always provide well-reasoned answers that are both correct and helpful."}, {"role": "user", "content": "Hello, introduce yourself to someone opening this program for the first time. Be concise."}, ] while True: completion = client.chat.completions.create( model="local-model", # this field is currently unused messages=history, temperature=0.7, stream=True, ) new_message = {"role": "assistant", "content": ""} for chunk in completion: if chunk.choices[0].delta.content: print(chunk.choices[0].delta.content, end="", flush=True) new_message["content"] += chunk.choices[0].delta.content history.append(new_message) # Uncomment to see chat history # import json # gray_color = "\033[90m" # reset_color = "\033[0m" # print(f"{gray_color} {'-'*20} History dump {'-'*20} ") # print(json.dumps(history, indent=2)) # print(f" {'-'*55} {reset_color}") print() history.append({"role": "user", "content": input("> ")})

Same...😡

@@jjgravelleReally want to check this agent system out. Does it even run on windows? WSL2?

@@IdPreferNot1Got it to run. Make sure you CD to 'notebooks' before running 'Jupyter notebook' from the command line. Had to try it a couple times. Be aware that the displayed error is cached, so you'll see it every time you load it. But pay attention to the blue indicator on the left side, showing which step is currently active. On mine, the first step is Step 2. Run them one at a time, even if the error is still displayed. Give each step time to respond. Good luck...

Thx @jjgravelle. I'm running linux on WSL so i had to install a GUI browser into linux, then it opened.

Set model parameter accordingly in all of your agents

Working on docs. Coming soon

@@vrsen 🫡

​@vrsen how's docs going?

openai claims to not train on data generated from api calls. whether its true who knows

@@ryzikx They also claimed to not have trained their LLMs on copyrighted data and now it shows that they definitely lied to everybody about that. And: Even if they are not: They must be keeping the full data of your conversations with their models to be able to pick up a conversation in the past. - For how long are they keeping that data? Who can say what happens with the data when the company get's into whatever trouble or will be sold? If your company relies more and more on AI workflows and use a big company offering the AI services for that they will be able to create an image of your whole company's situation! Example: One day you give them a spreadsheet of your employees illness days of the last years to get an overview. Only with this data they will be able to say how many employees you have, how stressful working at your company is, at which points in time there was very much stress and maybe even guess when there's a bigger project to be rolled out. - If you take a bit more time thinking about what else could be found out with only these data then the 10sec I did now, you probably won't want to give those data to other companies any more. What when they get hacked? - Happened even to the best security companies themselves! - And could there be a better target than a company which has all/a lot of confidential data of other companies? Can't even imagine how much that data was worth. There's absolutely no chance we would send our or our client's company-data to a third party. And as a small or midsize company you even couldn't file a lawsuit against them because they just had to stretch it until you are bankrupt.

You can now install from pip

You can change this by changing model parameter when initializing an agent