LangGraph Builder - Utility and abstraction to quickly build graphs #26973

[rvndbalaji]

In my org, I implemented a utility class to dynamically build LangGraph graphs
and abstracted away some of the common details

It is no means perfect, and is probably incomplete,
but is there a way for you guys to implement this in langchain-core or langgraph??

I have this raw implementation here - https://gist.github.com/rvndbalaji/2f50bf3d2fc635bf2f36f71c724e3347

This can be used as a reference to implement a proper utility class

body = GraphConfig(
    graph_id=3,
    name="Graph",
    description="Graph",
    session_id=None,
    agents=[
        AgentConfig(
                    name="Researcher", 
                    prompt="You are a research assistant", 
                    llm="claude-sonnet-3.5", 
                    tools=[], 
                    color="#FF871F"
                   ),
        AgentConfig(
                    name="Writer", 
                    prompt="You are a great writer",
                    llm="claude-sonnet-3.5", 
                    tools=[], 
                    color="#FF871F"
                   ),
    ],
    edges=[
        GraphEdge(source="start", target="Researcher"),
         GraphEdge(source="Researcher", target="Writer"),
        GraphEdge(source="Writer", target="end")
    ]
)


graph = GraphBuilder(graph_config=body)
graph.build()
response = GraphResponse(
    config = graph.get_compiled_config(),
    base64 = graph.get_image(xray=True, mermaid=True)
)
response

@baskaryan @hwchase17

[bhavan-kaya]

Hi @rvndbalaji,

I'm also trying to implement something similar for a different use case.

In my case, the goal is to create a drag-and-drop flow builder, similar to Langflow (worth exploring), and then dynamically generate agents based on the configurations.

The key challenge is to retrieve all the configurations in real time, dynamically create the tools, and then construct an agent from them. Your implementation seems quite similar to what I need. Would you mind explaining how you did this?

[rvndbalaji]

@bhavan-kaya
I have provided a user interface with the following

1. Add Agent button & Build Graph button
2. There is a d3 graph which is rendered. When user adds an agent on the top, a node is added to the graph using d3
3. When user selects the node on the graph, a sidebar appears that allows the user to provide name of agent, prompt, selecting tools from drop down, selecting LLM etc. (In my case, the tools are ready-made and pre-exist in backend)
4. The user can save this as config (as json in the DB)
5. The exact same json is used to construct the request and send it to the python service and build the graph as shown in my examples
6. Once the graph compiles properly, I will allow the user to call the graph.run method via the UI
7. The user should be able to see the streamed outputs of each agent on the UI

Hope this helps. Drag and drop is best, but its better to have a button to add by selecting the agent in a new drop down in your case. Atleast for an MVP

List of Agents
[{"llm": "", "name": "start", "color": "#B2AFEB", "tools": [], "prompt": ""}, {"llm": "claude-sonnet-3.5", "name": "Researcher", "color": "#67D995", "tools": [], "prompt": "You are a research assistant. Your task is to search for information and summarize it concisely. Once research is done, give the text to the translator agent"}, {"llm": "claude-sonnet-3.5", "name": "Reviewer", "color": "#f094d3", "tools": [], "prompt": "You will review the outputs of other agents"}, {"llm": "claude-sonnet-3.5", "name": "Translator", "color": "#ffac4d", "tools": [], "prompt": "You are an expert translator"}, {"llm": "", "name": "end", "color": "#99DBEA", "tools": [], "prompt": ""}]

List of edges
[{"source": "Reviewer", "target": "end"}, {"source": "Researcher", "target": "Reviewer"}, {"source": "start", "target": "Researcher"}, {"source": "Translator", "target": "Reviewer"}]

[bhavaniit24]

Thank you, @rvndbalaji This was really helpful! I appreciate it.

[levi-katarok]

This is really cool! Thanks for sharing

[hwchase17]

this is very cool! we probably wont implement as part of langgraph (we want langgraph to remain more low level) or langchain-core (even more low level), but there is 100% room for DevX like this, and even better if built on top of LangGraph (you get all the persistence, human-in-hte-loop, observability etc for free)

If you want to make this into a package/repo of your own, we would def share!