langgraph/tutorials/introduction/

[giscus[bot]]

langgraph/tutorials/introduction/

Build language agents as graphs

https://langchain-ai.github.io/langgraph/tutorials/introduction/

[rudiranck]

Großartig!

Rudi Ranck, PhD

[babelpainterwell]

Anyone found some glitches with appending SystemMessage to 'state'? They can be correctly appended, by printing out graph.get_state() in the end, but it seems the agent can't be guided by any info in the SystemMessage.

[rudiranck]

**I've just ran a test, it works. (!) If you try to insert a systemmessage twice it breaks. SO if you wanna update a system message, you need to inform the .id of this very me

Adding a system message___________________________________________________

if count ==0:
    system_message = SystemMessage(content="Answer only in spanish language")
    graph.update_state(config, {"messages": [system_message]})
#___________________________________________________________________________

(!) If you try to insert a systemmessage twice it breaks. SO if you wanna update a system message, you need to inform the .id of this very message, so it's updated instead appended:

    new_system_message = SystemMessage(content="now you are....")
    graph.update_state(config, {"messages": [new_system_message ]}, id=....)

[babelpainterwell]

Hi thanks for the feedback. I'm not neither inserting it twice nor update the system message. I was just using it as the first message before my first human message while using Langgraph and the problem was the agent seemed not able to be guided by the system message. I updated some of my examples here. #635

[rudiranck]

check whether you're using llama 3. I had some trouble with it to understand how to use properly tools and also notice the system prompt.

[rudiranck]

also... take a look how the guy is inserting the prompt here, maybe using this works for you: https://www.kaggle.com/code/scratchpad/notebooke4904c6179/edit

[Eknathabhiram]

Hey Can we update the state in router function without returning.

[pnshann]

Hi, I am using AWS Bedrock via Boto3 to replicate this workflow. I have been successful up until Part 4: Human-in-the-loop. When I attempt to run the code with the 'interrupt_before' argument included in the graph compiler, the LLM still executes the tool instead of being interrupted. There are no error messages.

For replication purposes, I am using Claude 3 Haiku and using the ChatBedrock function to initialize the model. My environment is as follows:

• Python version == 3.9.15
• LangGraph==0.0.66
• Boto3==1.34.115

Any help or guidance you can provide would be much appreciated. Happy to provide more information. Thanks!

[rudiranck]

I had some issues like this in the past. I suggest you to update all the packages:
example if you going to run it on the jupyter:
%%capture --no-stderr
%pip install -U langchain-community
%pip install -U langchain_anthropic
%pip install -U tavily-python
...
then try to debug the code to follow the execution, and be sure the execution is really no getting out the loop 'for event in graph.stream(inputs, config):...' (that happens when it encounters the interruption)

[njt1980]

Hi - Question, can we update the state from a tool as well? I am running into some kind of issue. The state which I have defined is as follows.
class State(TypedDict):
messages: Annotated[list[AnyMessage],add_messages]
cart: list[str]

This is the tool which I have defined for updating the cart.

@tool
def add_to_cart(state: State, item: str):
"""
Add an item to the cart.
Args:
item: The item to add.
"""
state.get("cart").append(item)
cart = state.get("cart")
return {"cart":cart}

The run completes successfully but when I do a get_state on the graph, it does not include the "cart" key. Just putting part of the message in here
..U5iOoH8x'), AIMessage(content='The apple has been added to your cart.', response_metadata={'token_usage': {'completion_tokens': 10, 'prompt_tokens': 671, 'total_tokens': 681}, 'model_name': 'gpt-3.5-turbo-0125', 'system_fingerprint': None, 'finish_reason': 'stop', 'logprobs': None}, id='run-64e9aa3b-3847-4180-9132-d4347ccb12f7-0', usage_metadata={'input_tokens': 671, 'output_tokens': 10, 'total_tokens': 681})]}, next=(),..
I would have expected to see the cart value after the last AIMessage.
Will be great if I can get some direction on what might be going wrong here.
Thank you!

[rudiranck]

you should return the message from the tool as well, I guess your state is waiting for 2 params,

you should return something like, return {"messages": [response], "item": 'itemName'}

check in the documentation how to build tools ...

[dominik-roemer]

Having the same problem and question.
"you should return something like, return {"messages": [response], "item": 'itemName'}"
Didn't solve the problem.

[zallesov]

This is the longest Hello World program I've seen so far....
A but overwhelming for a first tutorial.

[hinthornw]

Thanks for voicing the feedback! Would you prefer that they be broken into seven separate tutorials? Each part is fairly self-contained. Could also put the quick start more prominently in the docs so this isn't the absolute first thing you run into.

[jhines2k7]

It's not obvious that the notebook from your intro is downloadable. Could you add the link to download the notebook and a call to action in the Conclusion? I know you've added it to the top of the article, but most people (including myself) will either not notice it, forget that it is there.

[pizzacoffeecode]

@hinthornw since you're looking for feedback I absolutely love this format. I've needed to get up to speed quickly for production and this was perfect. I get the OP's point about Hello World, but I think there is a place for this as a quick start also.

Having separate guides is great, but when you don't know what you don't know, this is a really nice format.

[winstones2]

"Table of contents" is not complete: there are 7 parts in this page, but there are only 3 parts in the "Table of contents".
The tutorial helps me a lot.

[hinthornw]

Oh good catch!

[vegetablest]

Hello, I've noticed that starting from the third part, the tools node in the visualization is displayed as action. Is this a mistake?

[hinthornw]

Fixed! I had rerun all the code without re-visualizing after changing the node name. Thanks for the eagle eye :)

[hidwata]

I'd like to ensure that the number of tokens received by the LLM is less than or equal to, for example, 10K. For this, I want the ‘MemorySaver’ to forget old histories. Alternatively, if the number of tokens exceeds the limit when the conversation history is passed from checkpointer to LLM, I want to delete the old history and then pass it on. How can I do this?

[hidwata]

I found the answer.
https://langchain-ai.github.io/langgraph/how-tos/managing-conversation-history/

[DionysusG]

the URL is outdated, unfortunately

[DionysusG]

https://langchain-ai.github.io/langgraph/how-tos/memory/manage-conversation-history/

[hidwata]

Thank you for showing the new link. It:s much more organized than before.

[duob-ai]

When using another search tool like "GoogleSerperResults" or "GoogleSearchResults" I'm always receiving the error: TypeError: Got an unexpected keyword argument '__arg1'

I can't figure out why.

[SamthinkGit]

Hi!
I've just started using LangGraph and I wanted to remark this guide is super helpful ^^.

PD: It is true as other comments say that maybe is a bit long for a quickstart, however it is didactic and easy to follow.

[arzzahid66]

hi everyone i want to create langgraph application without using tools just llm and stages like if stage 1 function intract with user like medical bot and after that stage 2nd llm ask user about problems and stage 3rd ask llm about patient age gender and 4th stage predict diagnose so give me idea or code how can i do that

[ITSAIDI]

Maybe you can do it with just kind of loop with an LLM ?

[BellaBe]

Create 4 agents/nodes, each with its own prompt, create state to keep track of each agent response, if required chain response from one agent into a prompt of another agent, exit the graph after the 4th agent response. It is a pretty straightforward solution. Tools usage is not mandatory.

[ITSAIDI]

Great Job Langraph Team !!

• I have a problem when I use Ollama models instead of Claud 3 you have been used in this Tuto, The Agent get stuck in infinite loop of Function Calling , it seems like he dosn't see the Meesage from the Tool and keep requesting it ?
  any Help please ???

[Mehrtash-Souri]

I have the same problem with llama3,

[Mehrtash-Souri]

I'm going to get Mistral with Ollama and check this with Mistral. maybe mistral does not have this problem

[ITSAIDI]

Yes, it's depending to the model I think, I try Mistral Nemo 12b m give good results.

[Mehrtash-Souri]

when I use llama3 for :
Part 2: Enhancing the Chatbot with Tools
The Agent get stuck in infinite loop of Function Calling.
With "ollama run mistral" i geo mistral in ollama. then with below function:

def getMistralLLMForToolCall():
return ChatMistralAI(model=llm_model,
endpoint="http://192.168.1.10:8080/v1",
temperature=0,
verbose=True,
api_key="mistral",
format="json",

                )

llm = getMistralLLMForToolCall()

I create instance of ChatMistralAI , and use it as llm in tutorial code it works without stucking on tool calling.
I hope this help.

regards.

[BellaBe]

the following implementation in "Part 1: Build a Basic Chatbot" section is throwing an error, probably it should not be used for stream mode as stream mode is returning partial results where assistant response is still empty. With graph.invoke method it works as expected

    user_input = input("User: ")
    if user_input.lower() in ["quit", "exit", "q"]:
        print("Goodbye!")
        break
    for event in graph.stream({"messages": ("user", user_input)}):
        for value in event.values():
            print("Assistant:", value["messages"][-1].content)

[BellaBe]

Part 3: Chat bot with memory example doesn't work with ChatGroq client, groq api complains about types, probably because of tool content being an array

BODY==== {'messages': [{'role': 'user', 'content': 'My name is Bella'}, {'role': 'assistant', 'content': '', 'tool_calls': [{'type': 'function', 'id': 'call_zxw4', 'function': {'name': 'tavily_search_results_json', 'arguments': '{"query": "Bella name meaning"}'}}]}, {'role': 'tool', 'content': [{'url': 'https://nameberry.com/b/girl-baby-name-bella', 'content': 'Bella is a girl\'s name of Italian origin meaning "beautiful". It is a diminutive of Isabella and other names with the suffix -bella, and has been popularized by the Twilight series and other celebrities.'}, {'url': 'https://www.thinkbabynames.com/meaning/0/Bella', 'content': 'Bella as a girls\' name is pronounced BELL-ah. It is of Italian and Latin origin, and the meaning of Bella is "beautiful". Also a short form of Isabel or Isabella (Hebrew) "God\'s promise". Author/politician Bella Abzug; author Belva Plain. Also form of Arabella.'}], 'tool_call_id': 'call_zxw4'}], 'model': 'mixtral-8x7b-32768', 'frequency_penalty': NOT_GIVEN, 'function_call': NOT_GIVEN, 'functions': NOT_GIVEN, 'logit_bias': NOT_GIVEN, 'logprobs': NOT_GIVEN, 'max_tokens': NOT_GIVEN, 'n': 1, 'parallel_tool_calls': NOT_GIVEN, 'presence_penalty': NOT_GIVEN, 'response_format': NOT_GIVEN, 'seed': NOT_GIVEN, 'stop': None, 'stream': False, 'temperature': 1e-08, 'tool_choice': NOT_GIVEN, 'tools': [{'type': 'function', 'function': {'name': 'tavily_search_results_json', 'description': 'A search engine optimized for comprehensive, accurate, and trusted results. Useful for when you need to answer questions about current events. Input should be a search query.', 'parameters': {'type': 'object', 'properties': {'query': {'description': 'search query to look up', 'type': 'string'}}, 'required': ['query']}}}], 'top_logprobs': NOT_GIVEN, 'top_p': NOT_GIVEN, 'user': NOT_GIVEN}

groq.BadRequestError: Error code: 400 - {'error': {'message': "'messages.2' : for 'role:tool' the following must be satisfied[('messages.2.content' : value must be a string)]", 'type': 'invalid_request_error'}}

[vbarda]

please update to 0.1.12 -- we pushed a fix for this!

[BellaBe]

Thank you @vbarda, appreciate!

[ivar-os]

Hi, great introduction so far.

In part 5 it says:
new_messages = [
# The LLM API expects some ToolMessage to match its tool call. We'll satisfy that here.
ToolMessage(content=answer, tool_call_id=existing_message.tool_calls[0]["id"]),
# And then directly "put words in the LLM's mouth" by populating its response.
AIMessage(content=answer),

However, I got the error "ToolMessage not defined" (which seems to be the case, indeed). What should be used instead?

[isahers1]

Thank you for pointing this out, it should be fixed here: #1620. ToolMessage is actually part of LangChain and can be imported by writing from langchain_core.messages import ToolMessage

[Shenoy1984]

The below code dosent seem to give the tool...the call dosent seem to go, Is everyone facing this issue? it dosent seem to get tool.name. I am using OpenAI.

class BasicToolNode:
"""A node that runs the tools requested in the last AIMessage."""

def __init__(self, tools: list) -> None:
    self.tools_by_name = {tool.name: tool for tool in tools}

def __call__(self, inputs: dict):
    if messages := inputs.get("messages", []):
        message = messages[-1]
    else:
        raise ValueError("No message found in input")
    outputs = []
    for tool_call in message.tool_calls:
        tool_result = self.tools_by_name[tool_call["name"]].invoke(
            tool_call["args"]
        )
        outputs.append(
            ToolMessage(
                content=json.dumps(tool_result),
                name=tool_call["name"],
                tool_call_id=tool_call["id"],
            )
        )
    return {"messages": outputs}

tool_node = BasicToolNode(tools=[tool])
graph_builder.add_node("tools", tool_node)

[Pdocw]

I have encountered the same problem, have you solved it?

[AurianePeretti]

Hello, I wanted to tell you that I am very disapointed with your system because I had tried for maybe two hours to run just the beginning of the tutorial but each time my api key was not correct. Eventually I found out that I wasn't trying with the right key, I was trying with an openai key while I had to try with an antropic api key.... When I ran again my code I was told that 'Your credit balance is too low to access the Anthropic API. Please go to Plans & Billing to upgrade or purchase credits.' Seriously ? First of all you could had very point out that we must create an anthropic API. Secondly I'm highly unhappy because I already had paid for OpenAI now I have to pay for another API to understand the course? At least OpenAI allows you to try for free its api for a few times. I don't understand because I already used Langchain with a LLM gpt4 and it worked. I checked many times where these information were written at the start of the tutorial but I didn't see it, if it was there please forgive me. But I'm very confused because I lost a lot of time since I'll have soon an job interview tomorrow.

[HuzaifaTahir355]

Hi @AurianePeretti,
You can use open source Models as well.
How ever, There is no any issue with the tutorial. I was used "mixtral-8x7b-32768" for practice purpose (free of cost).

Below code snippet can help you out.
!pip install -qU groq langchain-groq
from langchain_groq import ChatGroq

add your API key of GROQ

os.environ["GROQ_API_KEY"] = ""
llm = ChatGroq(model="mixtral-8x7b-32768")

[usamaamjadvaival]

LangGraph 🤩 is a groundbreaking tool that beautifully bridges the gap between language and data 😎, making complex analysis remarkably intuitive 👍 😻.

[Changsoo-Park]

Getting this issue while run Part 4 Human-in-the-loop. Anybody knows the solution?

PS C:\AI_study\LangChain\CatchupAI> & c:/AI_study/LangChain/CatchupAI/.venv/Scripts/python.exe c:/AI_study/LangChain/CatchupAI/LangGraph_Tutorial/LG_QuickStart_04.py
Traceback (most recent call last):
File "c:\AI_study\LangChain\CatchupAI\LangGraph_Tutorial\LG_QuickStart_04.py", line 59, in
for event in events:
File "C:\AI_study\LangChain\CatchupAI.venv\Lib\site-packages\langgraph\pregel_init_.py", line 1223, in stream
with SyncPregelLoop(
^^^^^^^^^^^^^^^
File "C:\AI_study\LangChain\CatchupAI.venv\Lib\site-packages\langgraph\pregel\loop.py", line 679, in init
self.checkpointer_get_next_version = checkpointer.get_next_version
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
AttributeError: '_GeneratorContextManager' object has no attribute 'get_next_version'

[simon28082]

i have same issue

[usamaamjadvaival]

Getting this issue while run Part 4 Human-in-the-loop. Anybody knows the solution?

PS C:\AI_study\LangChain\CatchupAI> & c:/AI_study/LangChain/CatchupAI/.venv/Scripts/python.exe c:/AI_study/LangChain/CatchupAI/LangGraph_Tutorial/LG_QuickStart_04.py Traceback (most recent call last): File "c:\AI_study\LangChain\CatchupAI\LangGraph_Tutorial\LG_QuickStart_04.py", line 59, in for event in events: File "C:\AI_study\LangChain\CatchupAI.venv\Lib\site-packages\langgraph\pregel__init__.py", line 1223, in stream with SyncPregelLoop( ^^^^^^^^^^^^^^^ File "C:\AI_study\LangChain\CatchupAI.venv\Lib\site-packages\langgraph\pregel\loop.py", line 679, in init self.checkpointer_get_next_version = checkpointer.get_next_version ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ AttributeError: '_GeneratorContextManager' object has no attribute 'get_next_version'

use this code:
from typing import Annotated

from langchain_anthropic import ChatAnthropic
from langchain_community.tools.tavily_search import TavilySearchResults
from typing_extensions import TypedDict

from langgraph.checkpoint.memory import MemorySaver
from langgraph.graph import StateGraph, START
from langgraph.graph.message import add_messages
from langgraph.prebuilt import ToolNode, tools_condition

memory = MemorySaver()

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

graph_builder = StateGraph(State)

tool = TavilySearchResults(max_results=2)
tools = [tool]
llm_with_tools = llm.bind_tools(tools)

def chatbot(state: State):
return {"messages": [llm_with_tools.invoke(state["messages"])]}

graph_builder.add_node("chatbot", chatbot)

tool_node = ToolNode(tools=[tool])
graph_builder.add_node("tools", tool_node)

graph_builder.add_conditional_edges(
"chatbot",
tools_condition,
)
graph_builder.add_edge("tools", "chatbot")
graph_builder.add_edge(START, "chatbot")

[Hui-Xiaoyang]

I copy and past the codes to try part 6, but seems it doesn't work as expected. It never interrupted beofre "human". What could be the reason?

[oniafk]

I've been trying to start with this quick start but no matter what I do it doesn't let me get the latest version, every time I run the pip install -U langgraph It gives me the version 0.0.8, I already update my python version, everything, but I just get messages like this on my terminal:

WARNING: The candidate selected for download or install is a yanked version: 'langgraph' candidate (version 0.0.8 at https://files.pythonhosted.org/packages/c5/96/bbb9a8ba05fac954037e2761ebdea4be89eb30da10455ff1c10222b1abe9/langgraph-0.0.8-py3-none-any.whl#sha256=916cb24b211e2b4fc31d7cafcad2f97f49dfab41e833c869e9ea62481eaa310e (from https://pypi.org/simple/langgraph/) (requires-python:>=3.8.1,<4.0))
Reason for being yanked:

ERROR: Could not find a version that satisfies the requirement langgraph==0.2.39 (from versions: 0.0.8)
ERROR: No matching distribution found for langgraph==0.2.39

what could I've been doing wrong?

[hypotenuse345]

Great tutorial! I have learned a lot about the major features of LangGraph from this notebook. It is also exciting to build an autonomous agent using LangGraph. I am looking forward to trying it out and seeing how it performs.

However, when I try to run the code, I encounter an annoying issue about getting the current state under graph.stream(...).

When I was trying to get the state snapshot via graph.get_state(...) or the state history via graph.get_state_history(...), it returns state with no 'values' or 'next' elements. Such issue seems to be related to the streaming mode. After graph.invoke(...), the current state or state history can actually be obtained.

StateSnapshot(values={}, next=(), config={'configurable': {'thread_id': '0'}}, metadata=None, created_at=None, parent_config=None, tasks=())

I have made some attempts to see if I can obtain the state snapshot under streaming mode. I found it can be obtained again when I iterate the events of the streaming results first. However, I am not sure if this is the correct way to do it. Does anyone else encounter this issue? Is there any better way to get the state snapshot under streaming mode?

Here is my version info:

langchain_community==0.3.3
langchain_core==0.3.12
langchain_openai==0.2.3
langgraph==0.2.39

[vishnu1729]

Hi,

I have been trying your tutorial with my own use case.

1. chatbot converts natural language user input into Cypher queries.
2. tools takes the generated Cypher from the LLM and runs it against Graph DB and renders result.

When i get to Part 4. I run one question as in the tutorial to inspect the next state. However here i get snapshot.next as (). How do i fix this?

Unfortunately i can't share a lot of code as it is work related.

[vishnu1729]

I couldn't attach images but sharing the query and response with some info changed.

user_input = "give me the email contact for users in our graph. Show me only the user information and email."
config = {"configurable": {"thread_id": "1"}}
# The config is the **second positional argument** to stream() or invoke()!
events = graph.stream(
    {"messages": [("user", user_input)]}, config, stream_mode="values"
)
for event in events:
    if "messages" in event:
        event["messages"][-1].pretty_print()

Output

================================ Human Message =================================

give me the email contact for users in our graph. Show me only the user information and email.
================================== Ai Message ==================================

cypher
MATCH (u:User)-[:EMAIL]->(email:EmailAddress)
RETURN u, email

Then if i do

snapshot = graph.get_state(config)
snapshot.next

Output

()

[vishnu1729]

this is one of those " my code doesn't work, i don't know why; my code works i don't know why" situation but it is working now. All the changes i could remember doing was defining the tools list after initializing the graph_builder and trying once without the interrupt_before

[Albert-Gao]

Need help for understanding the Part4: Human in the loop

The flow here is the user asks a question and the LLM invokes a tool call to provide an answer, from the example, I see no value provided from the new interrupt_before=["tools"] parameter.

For the following explanation, it makes things even more confusing, isnt the previous example(Part 2: Enhancing the Chatbot with Tools) tools also? according to this langsmith log (https://smith.langchain.com/public/4fbd7636-25af-4638-9587-5a02fdbb0172/r/2b55e643-f5c8-4ca8-a401-8b5a9f282a43)

Notice that unlike last time, the "next" node is set to 'tools'. We've interrupted here! Let's check the tool invocation.

Where is the interruption, please?

[LP727]

Taking the exact code in the example, in Part 3: Adding Memory to the Chatbot, my pretty print output looks like this:

================================ Human Message =================================

Hi there! My name is Will.
================================== Ai Message ==================================
Tool Calls:
tavily_search_results_json (call_17aarszw)
Call ID: call_17aarszw
Args:
query: Who is Will?

Instead of the expected :

================================�[1m Human Message �[0m=================================

Hi there! My name is Will.
==================================�[1m Ai Message �[0m==================================

Hello Will! It's nice to meet you. How can I assist you today? Is there anything specific you'd like to know or discuss?

It seems I cannot convince my llm to NOT call the tool on user_inputs. Even trying:
llm.bind_tools(tools, tool_choice="auto")
or
llm.bind_tools(tools, tool_choice="none")
Always results in the tool getting called.

So I am wondering, was there any pre-prompting used in this example? Was Part 3: Adding Memory to the Chatbot meant to be done using :
return {"messages": [llm.invoke(state["messages"])]}
instead of:
return {"messages": [llm_with_tools.invoke(state["messages"])]}
I am pretty sure I can prompt my way around this or add an extra node to my graph but what is the expected behaviour here?

[SharoMonk]

I ran into the same problem and what solves it for me is changing the LLM i was using. I think the problem from my experience is lack of standardisation in how the providers treats and return messages.

[LP727]

Was changing the model sufficient or did you have to swap the API as well? Meaning, did you have to change from a ChatAnthropic to a ChatOpenAI or was changing the model="..." enough to get it to work?

My problems where occuring using ChatOpenAI so it might be that the support for tool is not well implemented, or that the LLM fails to predict what the expected message is supposed to look like.

[ryanma9629]

for event in graph.stream({"messages": [("user", user_input)]}):

why using ("user", user_input) as input? If I change it to ("human", user_input) or just user_input, seems nothing different.

[CaptainDaVinci]

In the part 1: I get an error:

ValueError: Message dict must contain 'role' and 'content' keys, got {}

[JMARK0]

As a beginner I start Langgraph , while importing START and END from "langgraph.graph" I got received error

from langgraph.graph import StateGraph,START,END
ImportError: cannot import name 'START' from 'langgraph.graph' (C:\MyProjects\MyApps\Agents\venv_agent\Lib\site-packages\langgraph\graph_init_.py)

On checking I observed the START and END constants are in langgraph.graph.graph and langgraph.graph.state can we get corrected in tutorials examples

[Rineraj]

Did anyone try to use Azure api key because I am able to run the code using openai api key but not with Azure key