I need some help with my feedback node!

[RedouanAssakali]

I am working on building a graph using LangGraph Studio Engineer.

I have modified the code and am trying to understand how to send information in a state, but I still need some support.

The goal is to create a visually appealing price quote based on simple notes sent by the user. When the content is generated, it should ask for feedback and make modifications until the user is satisfied. Once approved, we will generate a .docx file containing the price quote. My challenge lies in handling the feedback node.

I hope someone can assist me with my graph.

The logic behind this:

`from langgraph.graph import StateGraph, MessagesState, END
from langchain_core.tools import tool
from langchain_core.messages import HumanMessage, AIMessage, SystemMessage
from langchain_openai import AzureChatOpenAI
from pydantic import BaseModel, Field
from typing import List
from IPython.display import Image, display
from dotenv import load_dotenv
import os
load_dotenv()

os.environ["AZURE_OPENAI_ENDPOINT"] = os.environ.get("OPENAI_ENDPOINT")
os.environ["AZURE_OPENAI_API_KEY"] = os.environ.get("OPENAI_API_KEY")

model = AzureChatOpenAI(
api_version="2024-05-01-preview",
model="gpt-35-turbo",
)

Define the Operation and Quote classes

class Operation(BaseModel):
activity: str = Field(description="The main activity or task to be performed")
details: str = Field(description="Detailed description of the activity")

class Quote(BaseModel):
receiver_information: str = Field(dzescription="Receiver Information, company name, address, etc.")
subject: str = Field(description="Subject of the quote")
message_introduction: str = Field(description="Introduction message")
operations_table: List[Operation] = Field(description="Operations table with activities and details")
explanation_operations: str = Field(description="Explanation of operations")
deliverables: List[str] = Field(description="Deliverables")
total_price: str = Field(description="Total price")
conditions: str = Field(description="Conditions")
conclusion: str = Field(description="Conclusion")
closing: str = Field(description="Closing")
@tool()
def fill_in_content_tool(tag_name: str, content: str) -> None:
""" Fills in the content of a tag in the document """
print(f"Filling content for {tag_name} with {content}")
@tool()
def create_table_tool(tag_name: str, data: list) -> None:
""" Creates a table in the document """
print(f"Creating table for {tag_name} with data {data}")

Define the state with a messages key

class DocumentState(MessagesState):
quote: str
docx_content: dict
is_done: bool

Define the functions for each node

def start_node(state: DocumentState) -> DocumentState:
return {"messages": state["messages"]}
quote_llm_instructions = """
Create a quote for a client. Include the following sections:

• receiver_information:

  • Extract the company name, contact person, and full address including postal code.

• subject:

  • Extract the main topic or purpose of the quote.

• message_introduction:

  • Write a polite greeting and briefly introduce the purpose of the quote. Mention any prior discussions or agreements.

• operations_table:

  • List the main activities or tasks to be performed.
  • Provide detailed descriptions of each activity.

• explanation_operations:

  • Write a detailed explanation of the operations, including the goals and expected outcomes. Mention the importance of each step and how it contributes to the overall project.

• deliverables:

  • List the specific items or results that will be delivered to the client.

• total_price:

  • State the total cost of the project, both in numerical and written form. Mention if the cost is exclusive of taxes (e.g., VAT).

• conditions:

  • List any conditions or assumptions that the quote is based on. This could include working hours, environment, and any other relevant factors.

• conclusion:

  • Provide any additional information that highlights the benefits or importance of the project. Explain why the client should be interested in this offer.

• closing:

  • Write a polite closing message, expressing confidence in the offer and looking forward to the client's response. Include the sender's name, title, and company.

if you can not provide the information fill in no information
"""

def fill_quote_node(state: DocumentState) -> DocumentState:
quote_llm = model.with_structured_output(Quote)

quote = quote_llm.invoke([SystemMessage(content=quote_llm_instructions)]+state["messages"])    
return {"docx_content": quote}

def preview_node(state: DocumentState) -> DocumentState:
"""Preview of the document"""

return {"messages": [SystemMessage(content="Preview of the document")]} 

def ask_feedback_node(state: DocumentState) -> str:
feedback = input("Do you want to modify the content? (yes/no): ")
if "it's ok" in feedback.lower():
return "generate_document_node"
else:
# InvalidUpdateError: Must write to at least one of ['messages', 'quote', 'docx_content', 'is_done']
feedback: str = input("Please provide feedback: ")
return { "modify_content_node": feedback}

def modify_content_node(state: DocumentState) -> DocumentState:
feedback = state["messages"][-1].content
fill_in_content_tool("tag_name", f"Modified content based on feedback: {feedback}")
return {"docx_content": f"Modified content based on feedback: {feedback}"}

def generate_document_node(state: DocumentState) -> DocumentState:
return {"docx_content": "Final document", "download_url": "http://example.com/download"}

Create the graph

graph = StateGraph(DocumentState)
graph.add_node("start_node", start_node)
graph.add_node("fill_quote_node", fill_quote_node)
graph.add_node("preview_node", preview_node)
graph.add_node("ask_feedback_node", ask_feedback_node)
graph.add_node("modify_content_node", modify_content_node)
graph.add_node("generate_document_node", generate_document_node)

Define the edges

graph.set_entry_point("start_node")
graph.add_edge("start_node", "fill_quote_node")
graph.add_edge("fill_quote_node", "preview_node")
graph.add_edge("preview_node", "ask_feedback_node")
graph.add_conditional_edges("ask_feedback_node", ask_feedback_node, {"generate_document_node": "generate_document_node", "modify_content_node": "modify_content_node"})
graph.add_edge("modify_content_node", "preview_node")
graph.add_edge("generate_document_node", END)

Compile the graph

app = graph.compile()

display(Image(app.get_graph().draw_mermaid_png()))

input_message = HumanMessage(
content="""Client: John Doe
XYZ Corporation
123 Main Street
12345 City

Assignment: Conducting a pilot project to develop a model for liquidity forecasting for XYZ Treasury. For this pilot, we will use ABC Company as the test case. The modeling will be based on liquidity forecast data provided in PDF format, with a monthly frequency.

Deliverable: An oral report of our findings regarding the feasibility and quality of liquidity forecasting using Machine Learning models.

Tasks:

Process data from PDFs
Develop a forecasting model
Fine-tune and optimize the model
Write and present the report
Price: 5500

""")
for event in app.stream({"messages": [input_message]}, stream_mode="values"):
print(event)

`

[gbaian10]

It's strange for ask_feedback_node to serve as both a node's action function and a path function for add_conditional_edges, as these two functions usually return different things.
There should be two separate functions to handle this.