I like these short videos that cover key ideas or something valuable in a concise way. For example, that LangChain prompt template was pretty handy to know about. So many videos out there about this kind of material are 30m to an hour or more. Too many of them belabor the point and meander into low value topics. So, good job on that front. Lastly, you should use dynamics compression on your audio. Some parts were much louder than others. Hope this helps!
I think for tasks , react is a general purpose worker agent .. but function calling is only a tool . So it cannot be compared . We should understand the difference between tools and function calls ! As tools generally take a single string input amd give a string output and functions can take args and return args ! So : ReAct ! The correct way ! To do a react chain ! You need to define a self-query tool , which takes a query input . So you should create a tool which uses this tool , ie : each tool now is a mini agent , so the only thing different is the prompt : So your think tool ! First we need a planner tool, ie uses the think tool and a planner preset prompt ! This is the thoughts ! Then you need your action tool : This is a mini toolbox , framed as actions : so the model would call the action with the parameters , and the reponse would be formatted as a observation ! Also your final answer tool: This takes your final answer and formats it into something lovely 😍 So each step is a tool: Now we can male a executable toolchain ! So if we create a tool which is basically a graph ! From the thought to the action to the final answer with a loop for thought (do i meed another tool call) 🤙.. we have made the react process into a toolchain ! So we force the process with chain /graph. Or we can just give the model the tools instead ! But three tools and one tools ? I think by giving the model just the single chain we optimise its workflow : Now we can also make selfquery-agents for other agents such as refiner and coder , So we can add these as tools or create another toolchain from the self-query Tools .. based on the graph set up in the langgraph docs ! So now we have two tool chains or a collection of self-query-agents ? We can give thr model the agents to choose ots own destiny or , we can create an intent detector tool ! A conditional node ! ... Or tool: So now we create a new process which goes for the intent node first and fron the output of the intent node the correct chain can be selected , ao we can have multiple chains fixed processes . And our intent detector detect of we need to use a agentchain or graph , or construct a graph with some tools ! So : the advanced model would alao have a rag ( or local memory tool) So we create a new set of tools which are based on the rag : So our search tools and file loader tools etc can manage local memory , perform querys etc . Agan we should deploy the same set up ! ... Tools to perform tasks and chain whicj perform multi turn tasks , such as application devlopment , ETL, finetuning , information gathering .. large sumary writing .. We can note also that some of these intesive multi turn tasks such as producing a dissertation etc . Would take a research step , (which utilizes websercher chains etc) and a sumary writing stage , whichw oudl also have its owm subchains , so these combined intesive tasks would take days or hrs! .. bit the output produced would be of the highest calibur ! It has been said that the work produced by gpt is not great quality especaily for scientific . But its only about how to implement a methodolgy . As seen above we implented a reAct . But this cohld have been a chain of thoughts , or forrest of thoughts . Or a sema or Crisp datamining methodolgy ! So by segmenting tasks into mini tools or agents or processes and chains in fact the output would have been produced multiple models or agents prompted differently to spexific role and purpose , even with personality injection and self crtique .. the output ia totally I recognizable and infact even Unique !
can't you just combine them both to get the best of both worlds? i guess you could also bind the tools when invoking the react prompt, so that the model would call a necessary tool based on the final result decision?
I like these short videos that cover key ideas or something valuable in a concise way. For example, that LangChain prompt template was pretty handy to know about. So many videos out there about this kind of material are 30m to an hour or more. Too many of them belabor the point and meander into low value topics. So, good job on that front. Lastly, you should use dynamics compression on your audio. Some parts were much louder than others. Hope this helps!
will address this in the next videos :) thanks a bunch!
I think for tasks , react is a general purpose worker agent .. but function calling is only a tool . So it cannot be compared .
We should understand the difference between tools and function calls !
As tools generally take a single string input amd give a string output and functions can take args and return args !
So :
ReAct !
The correct way !
To do a react chain !
You need to define a self-query tool , which takes a query input .
So you should create a tool which uses this tool , ie : each tool now is a mini agent , so the only thing different is the prompt :
So your think tool !
First we need a planner tool, ie uses the think tool and a planner preset prompt !
This is the thoughts !
Then you need your action tool :
This is a mini toolbox , framed as actions : so the model would call the action with the parameters , and the reponse would be formatted as a observation !
Also your final answer tool:
This takes your final answer and formats it into something lovely 😍
So each step is a tool:
Now we can male a executable toolchain !
So if we create a tool which is basically a graph ! From the thought to the action to the final answer with a loop for thought (do i meed another tool call) 🤙.. we have made the react process into a toolchain ! So we force the process with chain /graph.
Or we can just give the model the tools instead ! But three tools and one tools ? I think by giving the model just the single chain we optimise its workflow :
Now we can also make selfquery-agents for other agents such as refiner and coder ,
So we can add these as tools or create another toolchain from the self-query Tools .. based on the graph set up in the langgraph docs ! So now we have two tool chains or a collection of self-query-agents ? We can give thr model the agents to choose ots own destiny or , we can create an intent detector tool ! A conditional node ! ... Or tool:
So now we create a new process which goes for the intent node first and fron the output of the intent node the correct chain can be selected , ao we can have multiple chains fixed processes . And our intent detector detect of we need to use a agentchain or graph , or construct a graph with some tools !
So : the advanced model would alao have a rag ( or local memory tool)
So we create a new set of tools which are based on the rag :
So our search tools and file loader tools etc can manage local memory , perform querys etc . Agan we should deploy the same set up ! ... Tools to perform tasks and chain whicj perform multi turn tasks , such as application devlopment , ETL, finetuning , information gathering .. large sumary writing ..
We can note also that some of these intesive multi turn tasks such as producing a dissertation etc .
Would take a research step , (which utilizes websercher chains etc) and a sumary writing stage , whichw oudl also have its owm subchains , so these combined intesive tasks would take days or hrs! .. bit the output produced would be of the highest calibur !
It has been said that the work produced by gpt is not great quality especaily for scientific .
But its only about how to implement a methodolgy .
As seen above we implented a reAct . But this cohld have been a chain of thoughts , or forrest of thoughts . Or a sema or Crisp datamining methodolgy !
So by segmenting tasks into mini tools or agents or processes and chains in fact the output would have been produced multiple models or agents prompted differently to spexific role and purpose , even with personality injection and self crtique .. the output ia totally I recognizable and infact even Unique !
So a rapid change of thoughts !! Asap !
Just a quick question for open weather map langchain agent which one will be good
Thank for your comments
I didn’t get why with ReAct we have more control. Isn’t LLM still responsible to selecting the tool?
please can you share the website sources,papers of what you explained?
very good explanation
Very informative. keep them coming!
Could have added eamples where either was a better choice.
Nice Video. I subscribed!
can't you just combine them both to get the best of both worlds?
i guess you could also bind the tools when invoking the react prompt, so that the model would call a necessary tool based on the final result decision?