langgraph/tutorials/customer-support/customer-support/ #477
giscus[bot]
bot
Replies: 40 comments 62 replies
-
|
When I was testing the code of part 4, I found that when the user inputs "Yeah I think I'd like an affordable hotel for my week-long stay (7 days). And I'll want to rent a car.", the following exception occurs: Maybe the current code does not handle this type of exception. After comparison and verification, it appears that it is related to the following two lines of code in the “class Assistant” being repeated one extra time: Could someone please help confirm this? Thank you. |
Beta Was this translation helpful? Give feedback.
-
|
I tried running the code but ran into this error: TypeError Traceback (most recent call last) TypeError: unsupported operand type(s) for |: 'type' and 'type' |
Beta Was this translation helpful? Give feedback.
-
|
Ah - this is because I'm using the typescript-style typehints here, which weren't introduced until python v 3.10 - will try to update to make the type-hinting compatibile with lower python versions |
Beta Was this translation helpful? Give feedback.
-
|
For part 2, when I add the interrupt, the AI does not reply with the details received from the Tool Message. In fact, the tool message never executes. However, when I don't include interrupt, the workflow executes as expected and the Tool Message is returned, and the AI can utilize it for a conclusion. |
Beta Was this translation helpful? Give feedback.
-
|
@Yazanjian print events from the result in the while snapshot.next:
result = None
...
_print_event(result, _printed)
snapshot = part_2_graph.get_state(config) |
Beta Was this translation helpful? Give feedback.
-
|
Very interesting tutorial. In the last part with supervisor design pattern, there are some messages that are pushing redundant information to the user, when the sub agent delegate the rest of the conversation to the host assistant. I was wondering if making the sub agents more independent in terms of handling the whole concern (e.g. finishing up everything related to hotel booking without delegating the intermediate steps to the host) would remedy that. |
Beta Was this translation helpful? Give feedback.
-
|
From where "message.id" appears? Traceback (most recent call last): When I print my message above "if message" in _print_event, it appears: PRINT EVENT MESSAGE:: ('user', {'messages': [SystemMessage(content='Voce e uma secretaria responsavel por realizar agendamentos e responder perguntas basicas sobre o caso\n Use as devidas tools para consultar datas indisponiveis ou duvidas legais frequentes, ou outras informacoes em funcao da mensagem do usuario\n Caso seja uma solicitacao de agendamento, retorne de 2 a 4 opcoes de datas disponiveis dentro do intervalo de tempo dado pela messages\n Segue sua ultima acao, a partir dela, defina se vai dar uma resposta final, ou chamar uma tool.\n pense sempre na proxima acao \n '), HumanMessage(content='Quando podemos conversar na semana que vem?')]}) |
Beta Was this translation helpful? Give feedback.
-
|
It's from the reducer function so that we can interrupt and then switch out the message based on the id. Using a Checkpointer is required here. |
Beta Was this translation helpful? Give feedback.
-
|
How to make this agent interactive, rather than feeding it a list of pre-defined user input? |
Beta Was this translation helpful? Give feedback.
-
|
You can add whisper transcription and TTS model to make it more live like or use animation API to make avator out of it . |
Beta Was this translation helpful? Give feedback.
-
|
oh, that's not what I mean... just a normal chat interaction... I type something, the chatbot responds, then I write something else, chatbot replies... |
Beta Was this translation helpful? Give feedback.
-
|
You can just pass a user question to the graph stream. So, rather than looping the question list, you can for example: question = "What are the available options on 12 June?"
events = part_4_graph.stream(
{"messages": ("user", question)}, config, stream_mode="values"
)
for event in events:
_print_event(event, _printed)You also can use graph.invoke() if you want to. For prototyping, you can use jupyter notebook or maybe adding chatbot interface, such as using Gradio, etc. |
Beta Was this translation helpful? Give feedback.
-
|
excellent, this seems useful. Thanks! |
Beta Was this translation helpful? Give feedback.
-
|
The gradio UI example from the LangraphAgent Deeplearning.ai course is one of my favorite. |
Beta Was this translation helpful? Give feedback.
-
|
In the section where the primary assistant is defined, the 'route_primary_assistant' function is missing one output type 'enter_book_car_rental' in the literal. |
Beta Was this translation helpful? Give feedback.
-
|
You are right. Nice catch with the enter_book_car_rental, but I think the output does not matter in the literal and there is something wrong with the Assistant class by repeating two lines of code. |
Beta Was this translation helpful? Give feedback.
-
|
Yes Literal just specifies the possible outputs. It shouldn't affect the logic of the function. In the Assistant class, will a simple fix like removing the repeated codes work? Thanks. |
Beta Was this translation helpful? Give feedback.
-
|
yeah I have another link where this class is used ,and there they dont use repeating codes ,and it works fine |
Beta Was this translation helpful? Give feedback.
-
|
I thought the double take here: messages = state["messages"] + [("user", "Respond with a real output.")]
state = {**state, "messages": messages}
messages = state["messages"] + [("user", "Respond with a real output.")]
state = {**state, "messages": messages}Was intentional to account for the expected |
Beta Was this translation helpful? Give feedback.
-
|
I don't really understand what this is for. Wierdly in the video that LangChain made for this tutorial at this link: https://www.youtube.com/watch?v=b3XsvoFWp4c&list=PLfaIDFEXuae16n2TWUkKq5PgJ0w6Pkwtg&index=18 |
Beta Was this translation helpful? Give feedback.
-
|
I wonder how this chatbot run in production? If many people chat in the same time, does state will be mixed and chatbots responses are wrong for others people? |
Beta Was this translation helpful? Give feedback.
-
|
Every user has passenger_id and thread_id, so the identity is unique |
Beta Was this translation helpful? Give feedback.
-
|
I was attempting to create the graph in part 4, swapping out the llm with OllamaFunction with the relatively new bind_tools() feature. However, when invoking the llm after it is delegated to another assistant, it keep throwing an error "Failed to parse response from model". Was wondering if there are any solution for this? |
Beta Was this translation helpful? Give feedback.
-
|
Which model are you using from ollama? |
Beta Was this translation helpful? Give feedback.
-
|
I have tried with a few models, llama3 70b instruct, phi3 14b and mistral 7b. |
Beta Was this translation helpful? Give feedback.
-
|
Llama 3 70b is a fine tuned for function calling ,It should work ,I guess. |
Beta Was this translation helpful? Give feedback.
-
|
I'm interested in creating a graph based structure which allows for direct tools calls without the llm seeing it. Perhaps having a tool deciding assistant in the middle of the graph, one for direct and one for regular. Would love any advice. Example : graph TD |
Beta Was this translation helpful? Give feedback.
-
|
Looks good to me,bro |
Beta Was this translation helpful? Give feedback.
-
|
Hey Jaid, thanks for the feedback. Would love to connect as I'm having some difficulty implementing this. I understand something similar to Part 4 would be the path forward, but not sure how to set it up. |
Beta Was this translation helpful? Give feedback.
-
|
When I ask bot a question where the LLM outputs more than one tool_call objects, and them were to be routed to different nodes by some routing function, I get the following error:
It happens because the routing method can only route a message to one node, but sometimes the LLM expects to reach tools in diferent nodes, and ends up getting answer from only one tool. Example, in the following route function: route = tools_condition(state)
if route == END:
return END
tool_calls = state["messages"][-1].tool_calls
did_cancel = any(tc["name"] == CompleteOrEscalate.__name__ for tc in tool_calls)
if did_cancel:
return "leave_skill"
tool_names = [t.name for t in book_excursion_safe_tools]
# here is checking if all tool_calls match tools from safe node
if all(tc["name"] in tool_names for tc in tool_calls):
return "book_excursion_safe_tools"
return "book_excursion_sensitive_tools"In this case, if LLM calls a safe tool and a sensitive tool, the rounting function will end up routing to sensitive tools. Is there a way to resolve that? Should I prompt the LLM to always outputs only one tool_call at once? @hinthornw |
Beta Was this translation helpful? Give feedback.
-
|
We should change some hard-coded tool_calls[0] to a loop like this: "messages": [
ToolMessage(
tool_call_id=tc["id"], // instead of hard-coded tools[0]["id"]
content=f"API call denied by user. Reasoning: '{user_input}'. Continue assisting, accounting for the user's input.",
)
for tc in tool_calls
]so when multiple tool call involved, every tool call id could be included into our request message. |
Beta Was this translation helpful? Give feedback.
-
|
I got the same errors. Any updates on this? |
Beta Was this translation helpful? Give feedback.
-
|
One implementation to tackle the issue is to disable parallel tool calling. |
Beta Was this translation helpful? Give feedback.
-
|
When I run part4 using openai, I got error: Any help would be appreciated. |
Beta Was this translation helpful? Give feedback.
-
|
Problem solved, it because my modification of pop_dialog_state function. |
Beta Was this translation helpful? Give feedback.
-
|
Hi all! For part 4, we replace: graph_builder.add_edge("fetch_user_info", "primary_assistant") with the code below. I'm trying to understand why this is needed. It looks like we can enter the user_info node from what is called a "delegated state" (from the comments) where the state is already initialized. Can anyone explain this to me? Thank you! graph_builder.add_conditional_edges("fetch_user_info, route_to_workflow) def route_to_workflow(state: State) -> Literal["primary_assistant", |
Beta Was this translation helpful? Give feedback.
-
|
hello , |
Beta Was this translation helpful? Give feedback.
-
|
Yeah , I tested it out it is compatable |
Beta Was this translation helpful? Give feedback.
-
|
Can you please share an example of how to chain the graph with FastAPI? |
Beta Was this translation helpful? Give feedback.
-
|
Hi all - I'm trying to solve the issue when the primary assistant calls multiple tools. Right now when that happens the code produces the following error: openai.BadRequestError: Error code: 400 - {'error': {'message': "An assistant message with 'tool_calls' must be followed by tool messages responding to each 'tool_call_id'. The following tool_call_ids did not have response messages: call_2HBvFjVImXIiszYJKVvP9xvU", 'type': 'invalid_request_error', 'param': 'messages', 'code': None}} I think one way to solve this could be to add an additional node between the primary assistant and the four assistant entry nodes, where the job of this additional node is to loop through all the tool calls that came from the primary assistant (so without invoking an LLM). This probably also requires pushing the tool_call_id's to the "stack", which can be done as part of the output of the primary assistant. The "tool looper" can then make sure that a ToolMessage is added for each tool call id that we track on the stack. The node that is used to leave a special assistant (pop_dialog_state) would likely have to route to the looper node, and then the looper node can route to the primary assistant once all tools call ids have been satisfied. I wish I could paste in a slide to visualize the above - let me know if there is a way to do this pls. What are your thoughts on this? Is there an existing/simpler design pattern for this issue? Thank you JDB |
Beta Was this translation helpful? Give feedback.
-
|
@siddicky That sounds interesting, but I still don't get why it's giving an error when calling Part 4🧐. Anyway, I'm planning to let it go for now…… |
Beta Was this translation helpful? Give feedback.
-
|
Any update on this problem? |
Beta Was this translation helpful? Give feedback.
-
|
Hey siddicky how did you implemented in UI using streamlit where let say Update my flight to sometime next week then I will search for flights departing later today and arriving next week. Then, I will update the user's ticket to the new flight. Error: IntegrityError('datatype mismatch') The search tool returned an error. I will now expand the search to include all departure airports and arrival airports. Error: IntegrityError('datatype mismatch') The search tool returned an error again. I will now expand the search to include all departure airports and arrival airports, and also expand the date range to include the next two weeks. Error: IntegrityError('datatype mismatch') The search tool returned an error for the third time. I will now expand the search to include all departure airports and arrival airports, and also expand the date range to include the next three weeks. in this in UI we have to only print the human message and the AImessage not the tools message so how can we built that if you have any thing can you help me ?? |
Beta Was this translation helpful? Give feedback.
-
|
Here is an implementation where tool calls are only displayed in the status container. If you want to not render any tool calls then you can remove the # Chat Messages and Stream Processing with Streamlit
# Container to display chat messages
messages = st.container(border=False, height=500)
# Loop through the list of messages and display each message
for msg in msgs.messages:
# Filter messages by type to be displayed in the chat container
if msg.content and msg.type in ("ai", "assistant", "function", "human", "user"):
# Display the message with an appropriate avatar (🤖 for AI, 👤 for human)
messages.chat_message(
msg.type,
avatar="🤖" if msg.type in ("ai", "assistant") else "👤",
).write(msg.content)
# Function to stream responses from the model
def stream_response(text: str):
# Yields each message from the text stream
for msg in text:
yield msg
# Function to process and stream the response from a model or tool
def process_stream(input, config):
# Initialize an empty string to store the generated response
response_generated = ""
# Display a status message during processing
with st.status("Loading...", expanded=True) as status:
# Simulate streaming response using graph.stream with input and configuration
for output in graph.stream(input, config):
# For each output, display key and related information
for key, value in output.items():
st.markdown(f"Node: **{key.upper()}**") # Display node name
# If the node contains messages, process and display the last one
if "messages" in value:
messages = value["messages"]
if isinstance(messages[-1], AIMessage):
# Add the content of the AI message to the response
if messages[-1].content:
response_generated += messages[0].content
else:
# If the AI message contains tool call data, display the arguments in JSON format
tool_call = messages[-1].tool_calls[0]
st.json(tool_call["args"])
# Update the status label to 'Completed' when processing is done
status.update(label="Completed", state="complete", expanded=False)
return response_generated
# Capture user input from the chat input box and generate a response
if prompt := st.chat_input(placeholder="Ask me a question!"):
# Display the user message in the chat container
messages.chat_message("human", avatar="👤").write(prompt)
# Retrieve the learning module value from the session state
learning_module = st.session_state.learning_module_value
# Display the AI's response in the chat container
with messages.chat_message("ai", avatar="🤖"):
# Process the user's input to generate a response
response_generated = process_stream(
{
"messages": [HumanMessage(content=prompt)],
"learning_module": learning_module,
},
config,
)
# Stream the response using the stream_response function
st.write_stream(stream_response(response_generated))
# Add the user and AI messages to the message history
msgs.add_user_message(prompt)
msgs.add_ai_message(response_generated) |
Beta Was this translation helpful? Give feedback.
-
|
Hey siddicky i am unable to do that when i am doing the key,value for output or where in my code i have taken the event i am not gettiing any message in value here is the full code for that from typing import Annotated from langgraph.checkpoint.memory import MemorySaver class State(TypedDict): class Assistant: llm = ChatOllama(model="llama3.2:3b", temperature=1) primary_assistant_prompt = ChatPromptTemplate.from_messages( part_1_tools = [ part_1_assistant_runnable = primary_assistant_prompt | llm.bind_tools(part_1_tools) builder = StateGraph(State) thread_id = str(uuid.uuid4()) _printed = set() |
Beta Was this translation helpful? Give feedback.
-
|
Hi, when I call the graph using events = graph.stream(
{"messages": ("user", question)}, config, stream_mode="values"
)or result = graph.invoke(
{"messages": ("user", question)}, config
)It looks like I am not passing any conversation history to the LLM. |
Beta Was this translation helpful? Give feedback.
-
|
I see some history in snapshot = graph.get_state(config)
messages = snapshot.values['messages'] |
Beta Was this translation helpful? Give feedback.
-
|
Hi, |
Beta Was this translation helpful? Give feedback.
-
|
In my opinion, it use BaseModel such as ToFlightBookingAssistant as a tool, llm decide call the tool according to user input, so i have the AiMessage with tool_calls, so we must generate ToolMessage for it, otherwise some problem occor like "An assistant message with 'tool_calls' must be followed by tool messages responding to each 'tool_call_id'".You can see, entry_node funtion return ToolMessage with tool_call_id. |
Beta Was this translation helpful? Give feedback.
-
|
I can't seem to access the config using ensure_config() within the tools. Anyone else facing the same issue? |
Beta Was this translation helpful? Give feedback.
-
|
I am also facing the same issue. I am stuck in this. Can anyone please help us here? from langchain_core.runnables import ensure_config ensure_config() is returning |
Beta Was this translation helpful? Give feedback.
-
|
def create_tool_node_with_fallback(tools: list) -> dict: changing this function to this, works. |
Beta Was this translation helpful? Give feedback.
-
|
please pass the config to the tools directly as an argument -- see this how-to https://langchain-ai.github.io/langgraph/how-tos/pass-config-to-tools/#define-tools-and-model |
Beta Was this translation helpful? Give feedback.
-
|
Hi! |
Beta Was this translation helpful? Give feedback.
-
|
Hello! Is there any good documentation available for Part 4 based on reAct approach? Thanks! |
Beta Was this translation helpful? Give feedback.
-
|
Let me know if you find any. |
Beta Was this translation helpful? Give feedback.
-
|
Hey, My Graph is similar to : ) Any help would be appreciated to resolve this error |
Beta Was this translation helpful? Give feedback.
-
|
I have an issue on @tool
def search_flights(
departure_airport: Optional[str] = None,
arrival_airport: Optional[str] = None,
start_time: Optional[date | datetime] = None,
end_time: Optional[date | datetime] = None,
limit: int = 20
):I am getting the following error: I replace the |
Beta Was this translation helpful? Give feedback.
-
|
I have the same error. I'm using command-r-plus from cohere as LLM. |
Beta Was this translation helpful? Give feedback.
-
|
there are duplicate sentence below： messages = state["messages"] + [("user", "Respond with a real output.")] |
Beta Was this translation helpful? Give feedback.
-
|
ValueError: Unexpected message type: content. Use one of 'human', 'user', 'ai', 'assistant', 'function', 'tool', or 'system'. |
Beta Was this translation helpful? Give feedback.
-
|
The customer support program does not work from Part 1. ValueError: Unexpected message type: content. Use one of 'human', 'user', 'ai', 'assistant', 'function', 'tool', or 'system'. |
Beta Was this translation helpful? Give feedback.
-
|
Sorry, I posted to the wrong place. |
Beta Was this translation helpful? Give feedback.
-
|
In Part 3, the while loop forgets to add: _print_event(result, _printed) Add this before snapshot = part_2_graph.get_state(config), otherwise the AI response after the tool node won't be displayed. |
Beta Was this translation helpful? Give feedback.
-
Congratulations for this valuable tutorial! Regarding asking confirmatin to the Human, how to implement this feature in a web environment, where the client would be in a webchat and our agent and tools would be in a cloud microservice? Thank you! |
Beta Was this translation helpful? Give feedback.
-
|
My previous idea in Gradio was to display the line of inquiry as an AIMessage in the chat box when the confirmation appears, allowing the user to input it, and then send the user's answer back as a HumanMessage. However, due to the issue of interruptions caused by while, inquiries may continue to occur or even crash directly. I didn't know how to solve this problem at the time, so I later gave up. Perhaps you can refer to and improve this idea? |
Beta Was this translation helpful? Give feedback.
-
|
I am trying to adapt PART 4 to my use case, the main differences are different |
Beta Was this translation helpful? Give feedback.
-
|
the full state of this single run would be nice to help you |
Beta Was this translation helpful? Give feedback.
-
|
Im extending the part 4 to one of my use cases and im facing some weird behavior when the user queries something that triggers more than 1 tool from the primary assistant (the tools used to work on the conditional_branching). Those behaviors are:
im clueless on this one |
Beta Was this translation helpful? Give feedback.
-
|
Can someone guide how can I use these specialized workflows in https://github.com/langchain-ai/opengpts |
Beta Was this translation helpful? Give feedback.
-
|
Could you please provide an example of: |
Beta Was this translation helpful? Give feedback.
and others
-
langgraph/tutorials/customer-support/customer-support/
Build language agents as graphs
https://langchain-ai.github.io/langgraph/tutorials/customer-support/customer-support/
Beta Was this translation helpful? Give feedback.
All reactions