Return tool call resulting object using langchain.prebuilt.ToolNode

[saminhozs]

Hi all,

I hope that all is well.

I'm using Langchain to build a MultiAgent chatbot and utilizing the prebuilt ToolNode function to inject state using InjectedState. However, ToolNode automatically converts the tool's result to a string for use in a ToolMessage. How can I store the tool's result in its original type (like a DataFrame) while still using ToolNode, or how can I create a custom tool node that supports state injection?

Example of a graph defined to use ToolNode:

data1_graph = StateGraph(AgentState)
data1_graph.add_node('llm', self.data1_llm)
data1_graph.add_node('action', ToolNode(self.tools_data1))
data1_graph.add_node('respond', self.data1_structure)
data1_graph.add_conditional_edges('llm', self.data1_router)
data1_graph.add_edge(START, 'llm')
data1_graph.add_edge('action', 'llm')
data1_graph.add_edge('respond', END)
self.graph1 = data1_graph.compile(checkpointer=self.memory)

Example of a tool for which I would like to store the result in pd.DataFrame type along storing it in a ToolMessage:

class RevenueSegment(BaseModel):
    isin: str = Field(description = "ISIN of the company for which we want to find the main revenue segment")
    n: int = Field(description = "Number of top business revenue segments we would like to identify")

@tool(args_schema=RevenueSegment)
def main_revenue_segment(isin: str, n: int) -> str:
    """Function that allows you to find the n main revenue segments of a company"""
    query = f"""
    SELECT DISTINCT(L6_ID), SUM(REVENUE_PCT) AS SHARE_REVENUE
    FROM
        XDW.IR_AI_SANDBOX.PB_RBICS
    WHERE
        ISIN = '{isin}'
    GROUP BY
        L6_ID
    ORDER BY
        SHARE_REVENUE DESC
    LIMIT {n}
    """
    data = query_snowflake_for_data(query, mode = 1)
    
    return data

Thank you so much in advance !!

[RoderickHearn]

Question: Are you using any llm platforms to generate code or answer questions?

I’m just starting my langchain journey and am wondering how various code generating gpts will work for the build. I’m building a multimodal ai assistant and want to use Langchain primarily for persistent memory and RAG.