Does Neuron Have A StateGraph for RAG

[Rndwiga]

I have an existing RAG implementation using LangChain and LangGraph Js. In my implementation, I make use of State Graphs to refine responses that a user gets. I have been going through the docs, but I have not found any mention of it. How can I achieve the same using Neuron AI?

[ilvalerione]

Neuron doesn't have a state graph. What do you mean with "refine responses that a user gets"?

[sixty-nine]

Would you please have a look at this proof-of-concept?

I have read somewhere (but cannot find the reference anymore) that it is possible to chain agents using the tools.
Introducing workflows would make it easier.
Additionally workflows allow non-linear chaining of agents: agent A needs agent B and agent C to be run before running.

I'd be pleased to create a PR if you think this would make sense to be integrated in Neuron AI.

[sixty-nine]

I don't want to develop much further without being sure that's what you were thinking of:

https://codeberg.org/sixty-nine/neuron-graph

Thank you in advance for your feedback.

EDIT: Don't worry about the namespaces, they can be adapted easily.

[sixty-nine]

Well it looks like I couldn't resist the call of the code. The project as it is seems usable.
Your feedback would still be very appreciated.

My goal was and remains to contribute to Neuron AI. But what I implemented with your ideas can be used completely stand-alone (although some examples explicitly use Neuron AI).

Would you like it to be integrated in Neuron AI?
Should it be a stand-alone extension?

WDYT?

[ilvalerione]

Oh it's very well made. Thank you for the commitment on the topic, that is perhaps very interesting but also challenging.

What I think will be important for the path of AI frameworks is a tool for multiagent collaborations toward a task.

This kind of graph implementation appeared to me as good trade-off between a starting point, and the uncertainty for patterns and use cases that could emerge along the way.

I also looked at projects like camel-ai, but it seems a bit out of my current time schedule.

And it is precisely on the subject of use cases that I would like to hear your opinion.

After this implementation, for what kind of automations you think it is a better component compared to what it's possible to do now?

[sixty-nine]

Thank you for the link to camel-ai, it's very interesting and inspiring.

From what I understood the original poster was asking for a way to easily chain agents. What he requested can already be done by calling each agent in turn and passing the data from the previous agent to the next.

What I implemented being a Finite State Machine everything it does can already be done with code by simply by going through the nodes calls and serializing (as in arrange in series) the code of each node.

For example the CoT example could simply be:

i = 3
while i > 0:
  run generate answer agent
  store answer
  i--

run choose answer agent with the stored answers

Trying to serialize the vending machine example would result in more intricate code:

balance = 0
product = null
done = false

while not done:
  ask for coins
  balance += coins

  while product = null:
    ask for product

  if (there is enough money):
    buy the product
    balance = balance - price
    done = true

The role of the graph is to split the business logic in small blocks that may be easier to understand.

And that's basically the use case that this would solve.

From what I understood from my quick look at the camel-ai code, they never use a graph to perform their complex tasks like Workforce or RolePlaying, it's only serialized code.

EDIT:

The idea below won't work with the current implementation because the predefined nodes cannot properly type-hint the run method.

Just an idea, that would need more thinking, but..

We could have predefined Nodes and Graphs.

We could have a predefined node that would run an agent with a message and return a message. Another predefined node that would run a node a given number of times or run a node until a condition is met. A node to store results in a database...

Then we could generate complex workflows simply by creating a graph with the right predefined nodes and not much more code.

We could have a predefined graph that takes an agent node, run it "x" times, and return the reply. That would just be a matter of creating and running the graph with the right predefined nodes.

Or a predefined graph for CoT, etc...

Of course everything above could already be implemented without graphs having predefined "higher-order" classes in NeuronAI.

[ilvalerione]

That's why the team here is skeptical, the multi agent collaboration is not strictly a chain. Camel kept our attention because you can create an agent network, and pass a "Task" to be accomplished.

What do you mean with "camel it's only serialized code"? The most interesting part is how camel start the execution of the network without a clear entry point, and then orchestrate the process.

[sixty-nine]

Camel kept our attention because you can create an agent network, and pass a "Task" to be accomplished.
What do you mean with "camel it's only serialized code"?

I'm not sure which feature of Camel you are referring to, I didn't get too much into the code of Camel but to me it looks like they are not using a graph to implement features like chain-of-thought or role-playing.

But I think there is a misunderstanding about the scope of my code.

The OP requested a way to chain agents in a graph-style way. That's what I implemented with my first approach. You can create static graphs to chain agents.

Then you suggested that each node could dynamically choose its successor, and that's what I implemented in my second repository.

As I tried to explain, everything that can be done with a graph can be done with normal code, because it's a FSM. The advantage of using a graph is that you can design simple nodes doing a single task and chain them to obtain a more complex task. That's a user level improvement in how they can architecture their code.

What you are asking now is: would the graph help us (the framework team) in implementing more complex tasks.

My point of view is that it could improve the overall readability/architecture of the code of complex tasks, but I wouldn't personally use a graph to implement them in the framework. As an example I explained that Camel does not (as far as I understand) use graph to implement those tasks.

The goal of my code is not to help maintainers, but to help end users with their code architecture.

If you don't see an interest from your part in such an addition to Neuron then I will keep my FSM code as a stand-alone thing.

It's too bad you pointed me on pydantic-graph when you actually didn't see a use-case for it.

[sixty-nine]

If someone is looking for a way to chain agents in a graph-style way, there is an implementation with some examples here.

[ilvalerione]

I'm sorry if we are not keeping up with your developments. We all are trying to learn. If you have more knowledge on this topics it's great to have some insights a step before. I believe it's healthy to have a discussion for the benefit of the community, documentation eventually, training contents, onboard other contributors, etc.

The point is that StateGraph isn't the only way to work on multi-agent applications. There are even more approaches we mentioned in this thread like handoff as tools, AutoGen, Semantic Kernel, CrewAI and so on. The discussion is to decide together what could be the best way to help developers solve multi-agent scenarios. Since we can learn from another, more advanced ecosystem, it's an opportunity to implement a better solution instead of doing the same errors others made in the past.

It seems there aren't solution that emerged as the best. They are just different. Thanks to your suggestions it's more clear the role of the StateGraph. Furthermore a StateGraph can be used also to create a workflow that depends by a supervisor agent that can be responsible to execute the underlying nodes in an "autonomous" way.

If you are open to send a PR it could be great to work on it.

Thank you for the contribution.

[sixty-nine]

Thank you for your reply.

Even though my Neuron-Graph implementation was initially meant to be used with NeuronAI, it turned out to be totally independent from it.

I prefer to keep it separated for few reasons:

• The repository can be used for other purposes than chaining agents
• I prefer to get rid of Github for my code
• I can apply my own coding standards, specifically PHPStan level 9.

I'd be glad to consider feature requests that could ease the integration with NeuronAI if needed.

[ilvalerione]

By the way we are adding PHPStan level 10. Consider the possible contributions and adoption it could have if it is part of the framework.