kenkenpa 0.2.0

Generate a StateGraph of LangGraph that can be compiled from structured data.

kenkenpa

Generate a StateGraph of LangGraph that can be compiled from structured data.

Note

This document is a translation of a text originally written in Japanese.

Purpose of this Library

This library was created to handle multiple AI agents within a single application.

Examples of use:

• When you want to build AI agents with different purposes for each room
• When you want to provide a UI for end-users to build AI agents by pre-providing nodes as components
• When you want to develop a project assistant that supports different development flows for each code repository
• When you want to apply different approval flows for each categorized document

Installation

pip install kenkenpa

Usage Example

I will explain how to use kenkenpa with the example of React-Agent.
https://langchain-ai.github.io/langgraph/

In addition to React-Agent, several implementation patterns of LangGraph are posted on kenkenpa_example.

---

Define the Tool node as usual.

from langchain_core.tools import tool

@tool
def search(query: str):
    """Call to surf the web."""

    if "sf" in query.lower() or "san francisco" in query.lower():
        return "It's 60 degrees and foggy."
    return "It's 90 degrees and sunny."

Define the factory function for the Tool node.

from langgraph.prebuilt import ToolNode

tools = {
    "search_function":search,
    }

def gen_tool_node(factory_parameter,flow_parameter):
    functions = factory_parameter['functions']

    tool_functions = []
    for function in functions:
        tool_functions.append(tools[function])

    tool_node = ToolNode(tool_functions)
    return tool_node

Define the factory function for the agent node.

from langchain_openai import ChatOpenAI

def gen_agent(factory_parameter,flow_parameter):
    functions = factory_parameter['functions']

    tool_functions = []
    for function in functions:
        tool_functions.append(tools[function])

     # Setting up the LLM
    model = ChatOpenAI(
        model="gpt-4o-mini"
    )

    model = model.bind_tools(tool_functions)

    # Define the function that calls the model
    def call_model(state):
        messages = state['messages']
        response = model.invoke(messages)
        # We return a list, because this will get added to the existing list
        return {"messages": [response]}

    return call_model

Define a function to evaluate whether the last message is a tool call instead of should_continue.

def is_tool_message(state, config, **kwargs):
    """Evaluate whether the last message is a tool call."""
    messages = state['messages']
    last_message = messages[-1]
    if last_message.tool_calls:
        return True
    return False

Create structured data representing a StateGraph.

graph_settings = {
    "graph_type":"stategraph",
    "flow_parameter":{
        "name":"React-Agent",
        "state" : [
            {
                "field_name": "messages", # state name
                "type": "list", # type(*1)
                "reducer":"add_messages" # reducer(*2)
            },
        ],
    },
    "flows":[
        { # agent node(*3)
            "graph_type":"node",
            "flow_parameter":{
                "name":"agent",
                "factory":"agent_node_factory",
            },
            "factory_parameter" : {
                "functions":[
                    "search_function",
                ],
            },
        },
        { # tools node(*3)
            "graph_type":"node",
            "flow_parameter":{
                "name":"tools",
                "factory":"tool_node_factory",
            },
            "factory_parameter":{
                "functions":[
                    "search_function",
                ],
            },
        },
        {# edge START -> agent
            "graph_type":"edge",
            "flow_parameter":{
                "start_key":"START",
                "end_key":"agent"
            },
        },
        {# coditional edge 
            "graph_type":"configurable_conditional_edge",
            "flow_parameter":{
                "start_key":"agent",
                "conditions":[
                    {
                        # Transition to the tools node if the result of is_tool_message is True.
                        "expression": {
                            "eq": [{"type": "function", "name": "is_tool_message_function"}, True], # *4
                        },
                        "result": "tools"
                    },
                    {"default": "END"} 
                ]
            },
        },
        {# edge tools -> agent
            "graph_type":"edge",
            "flow_parameter":{
                "start_key":"tools",
                "end_key":"agent"
            },
        },
    ]
}

Generate a StateGraphBuilder from graph_settings.

from langchain_core.messages import HumanMessage
from langgraph.graph import  add_messages
from langgraph.checkpoint.memory import MemorySaver

from kenkenpa.builder import StateGraphBuilder

# Instantiate the StateGraphBuilder
stategraph_builder = StateGraphBuilder(graph_settings)

# Some preparations are needed to generate a compilable StateGraph.
# *1. Register the types to be used. (list is reserved, but it is written for explanation purposes.)
#stategraph_builder.add_type("list", list)

# *2. Register the reducers to be used.
stategraph_builder.add_reducer("add_messages",add_messages)

# *3. Register the Node Factories.
stategraph_builder.add_node_factory("agent_node_factory",gen_agent)
stategraph_builder.add_node_factory("tool_node_factory",gen_tool_node)

# *4. Register the evaluation functions as well.
stategraph_builder.add_evaluete_function("is_tool_message_function", is_tool_message,)

# You can obtain a compilable StateGraph with the gen_stategraph() method.
stategraph = stategraph_builder.gen_stategraph()

# From here, write the code following the general usage of LangGraph.
memory = MemorySaver()
app =  stategraph.compile(checkpointer=memory,debug=False)

print(f"\ngraph")
app.get_graph().print_ascii()
final_state = app.invoke(
    {"messages": [HumanMessage(content="what is the weather in sf")]},
    config={"configurable": {"thread_id": 42}}
)
print(final_state["messages"][-1].content)

Execution Result

        +-----------+         
        | __start__ |         
        +-----------+         
              *               
              *               
              *               
          +-------+           
          | agent |           
          +-------+           
         .         .          
       ..           ..        
      .               .       
+-------+         +---------+ 
| tools |         | __end__ | 
+-------+         +---------+ 

The current weather in San Francisco is 60 degrees and foggy.

Definition of graph settings

Definition of stategraph (kenkenpa.models.stategraph.KStateGraph)

It represents a StateGraph.

{
    "graph_type":"stategraph",
    "flow_parameter":{
        "name":"GraphName",
        "state" : [
            {
                "field_name": "messages",
                "type": "list",
                "reducer":"add_messages"
            },
        ],
    },
    "flows":[
        #stategraph | node | edge | configurable_conditional_edge | configurable_conditional_entory_point
    ]
}

graph_type

• type: str
• desc: Specifies the graph type. Fixed as stategraph.

flow_parameter

• type: Dict
• desc: Data referenced when generating the StateGraph.

name

• type: str
• desc: The name of the StateGraph. In the case of a subgraph, it will be the name displayed in CompiledGraph. get_graph().

state

• type: List[Dict]
• desc: Describes the definition of custom states. Details are provided later.

flows

• type: List[Dict]
• desc: A list of definitions for StateGraph (SubGraph), node, and edge that make up the StateGraph.

Definition of state (kenkenpa.models.stategraph.KState)

This is the state definition referenced by the StateGraph. State is generated as TypedDict.

"state" : [
    {
        "field_name": "messages",
        "type": "list",
        "reducer":"add_messages"
    },
]

# This definition corresponds to the following.
from typing import Annotated
from typing_extensions import TypedDict
from langgraph.graph import  add_messages

class State(TypedDict):
    messages: Annotated[list, add_messages]

# Actual access
def is_tool_message(state, config, **kwargs):
    messages = state['messages']

field_name

• type: str
• desc: This is the field name.

type

• type: str
• his is a string representing the key for the field type.

The following types are pre-registered: int,float,complex,str,list,tuple,dict,set,frozenset,bool

To use types other than the reserved ones, you need to register the type with the StateGraphBuilder.

#...
#"state" : [
#    {
#        "field_name": "somestate_field",
#        "type": "SomeType_Key",
#    }
#]
#...

class SomeType(TypedDict):
    column: str

stategraph_builder = StateGraphBuilder(graph_settings)
stategraph_builder.add_type("SomeType_Key",SomeType)

# You can also pass it to the constructor.
type_map = {
    "SomeType_Key": SomeType
}

stategraph_builder = StateGraphBuilder(
    graph_settings = graph_settings,
    types = type_map
    )

reducer

• type: str
• This is a string representing the key for the reducer.

To use a reducer, you need to register it with the StateGraphBuilder.Full code

import operator

#...
#"state" : [
#    {
#        "field_name": "messages",
#        "type": "list",
#        "reducer":"operator_add"
#    },
#    {
#        "field_name": "custom_reduser",
#        "type": "list",
#        "reducer":"reducer_a_key"
#    },
#]
#...

def custom_reducer_a(left,right):
    if left is None:
        left = []
    if not right:
        return []
    return left + right

stategraph_builder = StateGraphBuilder(graph_settings)
stategraph_builder.add_reducer("operator_add",operator.add)
stategraph_builder.add_reducer("reducer_a_key",custom_reducer_a)

# You can also pass it to the constructor.
reducer_map = {
    "operator_add": operator.add,
    "reducer_a_key": custom_reducer_a
}

stategraph_builder = StateGraphBuilder(
    graph_settings = graph_settings,
    reducers = reducer_map
    )

Definition of node (kenkenpa.models.node.KNode)

This represents a single node.

{
    "graph_type":"node",
    "flow_parameter":{
        "name":"agent",
        "factory":"agent_node_factory",
    },
    "factory_parameter" : {
        "functions":[
            "search_function",
        ],
    },
}

graph_type

• type: str
• desc: Specifies the graph type. Fixed to node.

flow_parameter

• type: Dict
• desc: Data referenced when generating the StateGraph.

name

• type: str
• desc: The name of the node. This will be the name displayed in CompiledGraph.get_graph().

factory

• type: str
• factory: Represents the factory function that generates the runnable functioning as a node.

factory_parameter

• type: Optional[Dict]
• desc: Parameters passed to the factory function.

The factory function is defined to accept factory_parameter and flow_parameter.

#{
#    "graph_type":"node",
#    "flow_parameter":{
#        "name":"chatbot",
#        "factory":"gen_chatbot_agent_key",
#    },
#    "factory_parameter" : {
#        "functions":[
#            "search_function",
#        ],
#    },
#}
# node
def chatbot(state,config):
    return {"messages":[llm.invoke(state["messages"])]}

# factory
def gen_chatbot_agent(factory_parameter,flow_parameter):
    return chatbot

stategraph_builder = StateGraphBuilder(graph_settings)
stategraph_builder.add_node_factory("gen_chatbot_agent_key",gen_chatbot_agent)

stategraph = stategraph_builder.gen_stategraph()

You can also define it as a class acceptable by LangGraph.

class Chatbot():
    def __init__(self,factory_parameter,flow_parameter):
        pass

    def __call__(self,state,config):
        return {"messages":[llm.invoke(state["messages"])]}

stategraph_builder = StateGraphBuilder(graph_settings)
stategraph_builder.add_node_factory("gen_chatbot_agent_key",Chatbot)

stategraph = stategraph_builder.gen_stategraph()

# You can also pass the factory function map to the constructor.
factory_map = {
    "gen_chatbot_agent_key": Chatbot,

}

stategraph_builder = StateGraphBuilder(
    graph_settings = graph_settings,
    node_factorys = factory_map
    )

Definition of edge (kenkenpa.models.edge.KEdge)

Represents a single edge.

{
    "graph_type":"edge",
    "flow_parameter":{
        "start_key":"START",
        "end_key":"test_node_a"
    },
}

graph_type

• type: str
• desc: Specifies the graph type. Fixed to edge.

flow_parameter

• type: Dict
• desc: Data referenced when generating the StateGraph.

start_key

• type: Union[List[str],str]
• desc: Represents the starting point of the edge. Either start_key or end_key can be a list, but not both.

end_key

• type: Union[List[str],str]
• desc: Represents the endpoint of the edge. Either start_key or end_key can be a list, but not both.

Definition of configurable_conditional_edge (kenkenpa.models.configurable_conditional_edge.KConfigurableConditionalEdge)

This is the definition of a conditional edge.

{
    "graph_type":"configurable_conditional_edge",
    "flow_parameter":{
        "start_key":"agent",
        "conditions":[
            {
                "expression": {
                    "eq": [{"type": "function", "name": "is_tool_message_function"}, True],
                },
                "result": "tools"
            },
            {"default": "END"} 
        ]
    }
}

graph_type

• type: str
• desc: Specifies the graph type. Fixed to configurable_conditional_edge.

flow_parameter

• type: Dict
• desc: Data referenced when generating the StateGraph.

start_key

• type: str
• desc: The starting point of the edge.

conditions

• type: List[Union[KConditionExpression,KConditionDefault]]
• desc: Describes the conditional expressions to determine the next node. Details are provided below.

Definition of conditions (kenkenpa.models.conditions.KConditions)

This is the definition to determine the next node based on the result of the evaluation expression. Let's explain in a bit more detail.

"conditions":[
    {
        "expression": {
            "eq": [{"type": "function", "name": "is_tool_message_function"}, True],
        },
        "result": "tools"
    },
    {"default": "END"} 
]

conditions List

You can define a list of evaluation expressions, each with an expression and a result. If all expressions evaluate to false, the node specified in default will be transitioned to.

When routing based on the value of the state, it can be described as follows:

"conditions":[
    {
        "expression": {
            "eq": [{"type": "state_value", "name": "next_node"}, "node_a"],
        },
        "result": "node_a"
    },
    {
        "expression": {
            "eq": [{"type": "state_value", "name": "next_node"}, "node_b"],
        },
        "result": "node_b"
    },
    {
        "expression": {
            "eq": [{"type": "state_value", "name": "next_node"}, "node_c"],
        },
        "result": "node_c"
    },
    {"default": "END"} 
]

The conditions evaluate all expressions and specify all results that evaluate to True as the next nodes. In the following example, both node_a and node_b are returned.

"conditions":[
    {
        "expression": {
            "eq": [True, True],
        },
        "result": "node_a"
    },
    {
        "expression": {
            "eq": [True, True],
        },
        "result": "node_b"
    },
    {"default": "END"} 
]

• expression Comparison and logical expressions can be used.

• Comparison Expressions

  "operator": [operand,operand]
  

• Operators

  • equals,eq,==
  • not_equals,neq,!=
  • greater_than,gt,>
  • greater_than_or_equals,gte,>=
  • less_than,lt,<
  • less_than_or_equals,lte,<=

• Operands

  • state_value Refers to the value of the state.

    {"type":"state_value", "name":"state_key"}
    

  • config_value Refers to the value of the config.

    {"type":"config_value", "name":"test_config_key"}
    

  • function Calls a predefined evaluation function.

    {"type":"function","name":"test_function","args":{"args_key":"args_value"}}
    

  • Scalar values int,float,complex,bool,str,bytes,None

• Example 1 Verifies if the result of calling the evaluation function mapped to "is_tool_message_function" is True.

  "expression": {
      "eq": [
          {"type": "function", "name": "is_tool_message_function"}, 
          True
      ],
  }
  

  To use an evaluation function, it must be mapped in the StateGraphBuilder.

  def is_tool_message(state, config, **kwargs):
      """Evaluates if the last message is a tool_call."""
      messages = state['messages']
      last_message = messages[-1]
      if last_message.tool_calls:
          return True
      return False
  
  # Generate StateGraphBuilder from graph_settings.
  stategraph_builder = StateGraphBuilder(graph_settings)
  # Register the evaluation function
  stategraph_builder.add_evaluete_function("is_tool_message_function", is_tool_message,)
  

• Example 2 Verifies if the value of the state is "evaluate_value".

  "expression": {
      "eq": [
          {"type": "state_value", "name": "state_key"}, 
          "evaluate_value"
      ],
  }
  

• Example 3 Refers to the value of the config.

  config = {
      "configurable": {
          "thread_id":42,
          "user_id": "123",
          "tool_permission": True
      }
  }
  
  "expression": {
      "eq": [
          {"type": "config_value", "name": "tool_permission"}, 
          True
      ],
  }
  

• Logical Expressions

  "and": [evaluation expression or logical expression]
  "or" : [evaluation expression or logical expression]
  "not": evaluation expression or logical expression
  

  "expression": {
      "and": [
          {"eq": [{"type": "function", "name": "is_tool_message_function"}, True]},
          {"eq": [{"type": "state_value", "name": "tool_call"}, True]},
      ]
  }
  

  "expression": {
      "or": [
          {"eq": [{"type": "function", "name": "is_tool_message_function"}, True]},
          {"eq": [{"type": "state_value", "name": "tool_call"}, True]},
      ]
  }
  

  "expression": {
      "not":{
          "eq": [{"type": "function", "name": "is_tool_message_function"}, True]
      }
  }
  

  Logical expressions can be nested.

  "expression": {
      "and":[
          {
              "or":[
                  {
                      "and":[
                          {"eq": ["10", "10"]},
                          {"eq": [True, True]},
                          ]
                  },
                  {"eq": ["10", "10"]}
                  ]
          },
          {"eq": ["10", "10"]},
      ],
  }
  

• result and default The following values can be set.

  • state_value Refers to the value of the state. The value of the state must be a str (node name).

    {"type": "state_value", "name": "state_key"}
    

  • config_value Refers to the value of the config. The value of the config must be a str (node name).

    {"type": "config_value", "name": "config_key"}
    

  • function Calls a predefined evaluation function. The return value of the function must be a str (node name) or an instance of Send.

    {"type": "function", "name": "test_function", "args": {"args_key": "args_value"}}
    

  • Scalar values str

• Example
  The evaluation function can be used for purposes utilizing the Send API as follows. Full code

# Define continue_to_jokes so that it can be called as an evaluation function.
def continue_to_jokes(state: OverallState, config, **kwargs):
    return [Send("generate_joke", {"subject": s}) for s in state["subjects"]]

# Generate the StateGraphBuilder from graph_settings.
stategraph_builder = StateGraphBuilder(graph_settings)

# Similarly, the evaluation function is also registered.
stategraph_builder.add_evaluate_function("continue_to_jokes", continue_to_jokes)

graph_settings = {
    # (omitted)

    "flows": [
        {  # conditional edge generate_topics -> continue_to_jokes
            "graph_type": "configurable_conditional_edge",
            "flow_parameter": {
                "start_key": "generate_topics",
                "path_map": ["generate_joke"],  # Specify the path_map
                "conditions": [
                    # Define only the default in the conditions of configurable_conditional_edge,
                    # and call continue_to_jokes here.
                    {"default": {"type": "function", "name": "continue_to_jokes"}}
                ]
            },
        },
    ]
}