Very interested in the "useful overhead" of langraph/langchain, but not for everything. Can you guys consider building/ or demonstrating some more conversions "in and out" of langchain/langraph and into or from straight python. Not knowing whjat is happening under the abstraction can be confusing. Can you build or is there a function to "bridge" that through maybe a pydantic model or something that you could apply at any point of your lanchain coding?
Someone from Langchain can please tell me if we can access this fonctional api without paying for access to langgraph studio? I want to use langchain fonctional api for my agent and pay as I use it with an api Key and not spending 40$ per months for access to studio
I'm assuming that the tool decorator basically adds a parameter to construct a structured input to the prompt to the LLM. is there reference for this that I can lookup? For example if I want to create types of tools or change parts of what the actual structured input looks like depending on the LLM. Using for different languages is an example or tweaking the actual structured input to optimize the LLM's peformance if it's fine tuned for a different kind of input structure.
You provide the store, but usually postgres or mongo or some other db in prod. If it were just text file or in-memory that would be mostly useless in prod
@@willfu-hinthorn example shown on your how to guide for Multi agents is not able to relate with real use case . Also the supervisor gent shown in tutorial. Please check lots of people comments there (This one is not able to use in production )
Not really sure if I see any benefit. I was already critical about Command and Interruppt. In my opinion you start to bloat the framework with edge cases and make it messy, the same what happened with LangChain. I would not want to work with code like that: attempts = 0 @task(retry=retry_policy) def get_info(): global attempts attempts += 1 Putting attempts inside the state object is much cleaner in my opinion and I dont really see why we need a "quick and dirty" way to create an Agent.
The attempts here is only used to simulate a network failure to illustrate that RetryPolicy works out of the box. We'll clarify it in the documentation.
Just the time when i actually got used to managing these states, you brought something good 👍
Glad to hear you like it! But also if you like using StateGraph, don't worry - we don't have any plans on removing it!
Very cool video. Please do a video on Agentic RAG😊
Very interested in the "useful overhead" of langraph/langchain, but not for everything. Can you guys consider building/ or demonstrating some more conversions "in and out" of langchain/langraph and into or from straight python. Not knowing whjat is happening under the abstraction can be confusing. Can you build or is there a function to "bridge" that through maybe a pydantic model or something that you could apply at any point of your lanchain coding?
What is the core benefits from switching from Graph API to Functional API? Wouldn't I just stick to Graph API if I prefer graph structure?
Someone from Langchain can please tell me if we can access this fonctional api without paying for access to langgraph studio? I want to use langchain fonctional api for my agent and pay as I use it with an api Key and not spending 40$ per months for access to studio
I'm assuming that the tool decorator basically adds a parameter to construct a structured input to the prompt to the LLM. is there reference for this that I can lookup? For example if I want to create types of tools or change parts of what the actual structured input looks like depending on the LLM. Using for different languages is an example or tweaking the actual structured input to optimize the LLM's peformance if it's fine tuned for a different kind of input structure.
Nice tutorial. Is it possible to store a pandas dataframe in long term memory? Or, it supports only texts?
We fall back to pickle in some cases for the checkpointing but I haven't specifically checked how well it works for large pandas dataframes
Where can I find Store implementations for other databases?
Please convert already existing agent tools code into this so we will get idea
where did you store long term memory? text file? database? or just in memory?
You provide the store, but usually postgres or mongo or some other db in prod.
If it were just text file or in-memory that would be mostly useless in prod
How would it work with copilot kit?
Is there a roadmap on studio for Windows?
It's been supported for a long time. `langgraph dev` - it's a python package.
Like crewai we can’t directly or simply create agents.
Need more tutorials on real use case to showcase
from langgraph.prebuilt import create_react_agent
What specifically would you like to use that would make you faster/better at building applications?
@@willfu-hinthorn example shown on your how to guide for
Multi agents is not able to relate with real use case . Also the supervisor gent shown in tutorial. Please check lots of people comments there
(This one is not able to use in production )
Not really sure if I see any benefit. I was already critical about Command and Interruppt. In my opinion you start to bloat the framework with edge cases and make it messy, the same what happened with LangChain.
I would not want to work with code like that:
attempts = 0
@task(retry=retry_policy)
def get_info():
global attempts
attempts += 1
Putting attempts inside the state object is much cleaner in my opinion and I dont really see why we need a "quick and dirty" way to create an Agent.
The attempts here is only used to simulate a network failure to illustrate that RetryPolicy works out of the box. We'll clarify it in the documentation.
I certainly won't be using this. I'll stick with StateGraph.
love interrupt()
just became messier and messier
Pretty good abstraction model I'd say
Then use the graph dsl.
Older code is no longer needed?? Right