What's the limit on customizing the state for convenience

[intern162]

If I want to create a state that holds the following information:

Message: a list
Content: a string read from a file
Sender: the name of the agent
Here's the process flow I envision:

The reader agent reads the content from a file, setting the sender as the reader.
The state is then passed to the researcher who fact-checks and updates the content, updating the sender to the researcher.
The state is then passed to the writer who writes the content and then clears it, sending the state back to the reader.
How can one implement such logic and functionality? Are states this flexible? How would the agent_node function, as shown in the tutorial, look in this context?

[Sarthak-Kakkar-03]

I do not think it is possible to do that. The models are incapable of accomplishing this, even with Langgraph.

[intern162]

Really? I thought this application should be pretty basic for Langgraph to accomplish.

[Glinte]

It is probably possible. Your description is hard to understand but I think the thing you're trying to make is roughly just a finite state machine. As long as you write appropriate reducers it should work.

[intern162]

Hey!
I have posted an easier to read description, what do you mean by reducers?
Is making a state like this possible, I am having trouble understanding the structure of getting the llm to update the state's custom structure appropriately.

[Glinte]

https://langchain-ai.github.io/langgraph/concepts/low_level/#reducers

[hwchase17]

Would something like this work?

from typing import TypedDict, List, Annotated
from langgraph.graph import StateGraph, END
import operator

# Updated custom state with a counter
class CustomState(TypedDict):
    messages: Annotated[List[dict], operator.add]
    content: str
    sender: str
    counter: int  # New field to track iterations

# Updated agent functions
def reader_agent(state: CustomState) -> CustomState:
    file_content = "This is some content read from a file."
    return {
        "content": file_content,
        "sender": "reader",
        "messages": [{"role": "system", "content": "Content read from file."}],
        "counter": state["counter"] + 1  # Increment counter
    }

def researcher_agent(state: CustomState) -> CustomState:
    updated_content = f"Fact-checked: {state['content']}"
    return {
        "content": updated_content,
        "sender": "researcher",
        "messages": [{"role": "system", "content": "Content fact-checked and updated."}],
        "counter": state["counter"] + 1  # Increment counter
    }

def writer_agent(state: CustomState) -> CustomState:
    return {
        "content": "",
        "sender": "writer",
        "messages": [{"role": "system", "content": f"Content written: {state['content']}"}],
        "counter": state["counter"] + 1  # Increment counter
    }

# Updated router function with ending condition
def router(state: CustomState) -> str:
    if state["counter"] >= 3:  # End after 3 iterations (one full cycle)
        return END
    elif state["sender"] == "reader":
        return "researcher"
    elif state["sender"] == "researcher":
        return "writer"
    elif state["sender"] == "writer":
        return "reader"
    else:
        return END

# Create the graph
workflow = StateGraph(CustomState)

# Add nodes
workflow.add_node("reader", reader_agent)
workflow.add_node("researcher", researcher_agent)
workflow.add_node("writer", writer_agent)

# Set the entrypoint
workflow.set_entry_point("reader")

# Add edges
workflow.add_conditional_edges("reader", router)
workflow.add_conditional_edges("researcher", router)
workflow.add_conditional_edges("writer", router)

# Compile the graph
app = workflow.compile()

# Run the graph
result = app.invoke({"messages": [], "content": "", "sender": "", "counter": 0})
print(result)

[intern162]

Hey!
Thank you so much for taking time to out to respond to this, this is one of the best examples for Langgraph state usage I have seen.
Very helpful for beginners!

[hwchase17]

this was AI generated :) I'll share my "LangGraph engineer" project soon!

[intern162]

That would be great!
I am trying to just structure the workflow to use LLMs right now. Making it use tools like the File System for these processes.
So it automatically runs, fixes information, and writes. I will probably extract information selectively from the fields in the state for LLM invocations, dependent on their role.
If you have any ideas/examples of something useful for that, I would really appreciate it!
Thinking of using Langchain's with structure output with pydantic classes.

[hwchase17]

langgraph engineer: https://github.com/hwchase17/langgraph-engineer