Add AgentExecutor to the core - Take a look to the LangGraph4j

[bsorrentino]

Hi

I've drafted a first preview release of a LangGraph implementation in Java ( take a look here )

To validate such implementation I've build on it an AgentExecutor, that is the feature behind the Agents strategy in original Langchain project, using langchain4j.

I would to know if there is interest to add AgentExecutor implementation in core module (that would be great) so We can discuss the better way to do this

[langchain4j]

HI, @bsorrentino

Yes, we are looking into adding more features for building agents, so this is very timely!
We are also looking into slightly different concept: implementing a state machine (which is different in nature than a graph in langgraph), you can check it here and here.

Regarding LangGraph(4j), could you please provide a good example of a use case where modelling the logic as a graph is beneficial? All the examples that I've seen so far do not really require a graph (and cycles), they can be implemented as chains.

I will dive into it more deeply probably next week, but in general: yes, would be cool to have it in the core!

[bsorrentino]

Hi @langchain4j, thanks for feedback.

I first assumed that since langchain4j was inspired by langchain, following the langgraph strategy was in accordance with the goals of the project.

Regarding LangGraph(4j), could you please provide a good example of a use case where modelling the logic as a graph is beneficial? All the examples that I've seen so far do not really require a graph (and cycles), they can be implemented as chains.

Yes I'll do, however the AgentExecutor itself use a cycle, as diagram below:

---
title: AgentExecutor (with LangGraph)
---
stateDiagram-v2 
    [*] --> agent : prompt
    state if_state <<choice>>

    agent --> if_state : reasoning
    if_state --> [*]: finish
    if_state --> action : function call
    action --> agent : result
    

[langchain4j]

I first assumed that since langchain4j was inspired by langchain, following the langgraph strategy was in accordance with the goals of the project.

It is inspired by many projects + we also have our own perspective, so it does not have to be 1:1.

Regarding the function call cycle, a similar cycle is already implemented inside the AI Services. So registered tools/functions are called automatically untill LLM decides to stop and provide an output. I am not sure this use case requires an explicit graph. So it would be great to see another example.

Just to be clear: I am not agains the idea of LangGraph, and I have heard some positive feedback about it, so I am open for this idea.
But I am yet to see a good use case where graph really helps (as opposed to a regular if/else/for/while/switch flows written in plain Java without graph, which can be tested and debugged much easier). Most of the examples online do not require cycles, even LangChain recommends using LCEL in such cases.

[bsorrentino]

I've released another use case using LangGraph4j take a look 👀: Image To PlantUML Diagram

[langchain4j]

Great stuff! But it is a DAG as well 😆 so graph is not really a must

[bsorrentino]

Great stuff! But it is a DAG as well 😆 so graph is not really a must

@langchain4j thank for feedback. In the correction subgraph there is a cycle, anyway I understood that LangGraph is not in the LangChain4j's near future strategy.

Thanks again and keep going with your amazing project.

[langchain4j]

Sorry, I did not notice the cycle in the correction, my bad.

LangGraph is still in the future strategy, but I just don't have a capacity right now to address it.
LangGraph is much more advanced thing and we are still trying to cover pretty basic features.

Your project looks very interesting and I am sure there are users of LC4J that will benefit from it.
I was planning to create a langchain4j-community repository for community extensions (kind of an incubator).
Would you be interested in adding your modules there?
Thanks

[langchain4j]

I mean, to put langgraph4j into langchain4j-community (or keep it in your repos and jsut add a link there) for now and make LC4J users aware of it (I can include it in one of the next release notes and tweet about it).
Once it matures and we get more feedback from users, we could see if this can be added to the core.