Skip to main content

Build a Chatbot

Prerequisites

This guide assumes familiarity with the following concepts:

Note: this tutorial previously built a chatbot using RunnableWithMessageHistory. You can access this version of the tutorial in the v0.2 docs.

Overview​

We'll go over an example of how to design and implement an LLM-powered chatbot. This chatbot will be able to have a conversation and remember previous interactions.

Note that this chatbot that we build will only use the language model to have a conversation. There are several other related concepts that you may be looking for:

  • Conversational RAG: Enable a chatbot experience over an external source of data
  • Agents: Build a chatbot that can take actions

This tutorial will cover the basics which will be helpful for those two more advanced topics, but feel free to skip directly to there should you choose.

Setup​

Jupyter Notebook​

This guide (and most of the other guides in the documentation) uses Jupyter notebooks and assumes the reader is as well. Jupyter notebooks are perfect for learning how to work with LLM systems because oftentimes things can go wrong (unexpected output, API down, etc) and going through guides in an interactive environment is a great way to better understand them.

This and other tutorials are perhaps most conveniently run in a Jupyter notebook. See here for instructions on how to install.

Installation​

For this tutorial we will need langchain-core and langgraph:

pip install langchain-core langgraph

For more details, see our Installation guide.

LangSmith​

Many of the applications you build with LangChain will contain multiple steps with multiple invocations of LLM calls. As these applications get more and more complex, it becomes crucial to be able to inspect what exactly is going on inside your chain or agent. The best way to do this is with LangSmith.

After you sign up at the link above, make sure to set your environment variables to start logging traces:

export LANGCHAIN_TRACING_V2="true"
export LANGCHAIN_API_KEY="..."

Or, if in a notebook, you can set them with:

import getpass
import os

os.environ["LANGCHAIN_TRACING_V2"] = "true"
os.environ["LANGCHAIN_API_KEY"] = getpass.getpass()

Quickstart​

First up, let's learn how to use a language model by itself. LangChain supports many different language models that you can use interchangeably - select the one you want to use below!

pip install -qU langchain-openai
import getpass
import os

os.environ["OPENAI_API_KEY"] = getpass.getpass()

from langchain_openai import ChatOpenAI

model = ChatOpenAI(model="gpt-4o-mini")

Let's first use the model directly. ChatModels are instances of LangChain "Runnables", which means they expose a standard interface for interacting with them. To just simply call the model, we can pass in a list of messages to the .invoke method.

from langchain_core.messages import HumanMessage

model.invoke([HumanMessage(content="Hi! I'm Bob")])
API Reference:HumanMessage
AIMessage(content='Hi Bob! How can I assist you today?', additional_kwargs={'refusal': None}, response_metadata={'token_usage': {'completion_tokens': 10, 'prompt_tokens': 11, 'total_tokens': 21, 'completion_tokens_details': {'reasoning_tokens': 0}}, 'model_name': 'gpt-4o-mini-2024-07-18', 'system_fingerprint': 'fp_1bb46167f9', 'finish_reason': 'stop', 'logprobs': None}, id='run-7e4594a6-354f-45bf-8b83-e7185b50447f-0', usage_metadata={'input_tokens': 11, 'output_tokens': 10, 'total_tokens': 21})

The model on its own does not have any concept of state. For example, if you ask a followup question:

model.invoke([HumanMessage(content="What's my name?")])
AIMessage(content="I'm sorry, but I don't have access to personal information about users unless you've shared it with me in this conversation. How can I assist you today?", additional_kwargs={'refusal': None}, response_metadata={'token_usage': {'completion_tokens': 30, 'prompt_tokens': 11, 'total_tokens': 41, 'completion_tokens_details': {'reasoning_tokens': 0}}, 'model_name': 'gpt-4o-mini-2024-07-18', 'system_fingerprint': 'fp_1bb46167f9', 'finish_reason': 'stop', 'logprobs': None}, id='run-6b433482-b9f5-4a18-8089-d92719445402-0', usage_metadata={'input_tokens': 11, 'output_tokens': 30, 'total_tokens': 41})

Let's take a look at the example LangSmith trace

We can see that it doesn't take the previous conversation turn into context, and cannot answer the question. This makes for a terrible chatbot experience!

To get around this, we need to pass the entire conversation history into the model. Let's see what happens when we do that:

from langchain_core.messages import AIMessage

model.invoke(
    [
        HumanMessage(content="Hi! I'm Bob"),
        AIMessage(content="Hello Bob! How can I assist you today?"),
        HumanMessage(content="What's my name?"),
    ]
)
API Reference:AIMessage
AIMessage(content='Your name is Bob! How can I help you today, Bob?', additional_kwargs={'refusal': None}, response_metadata={'token_usage': {'completion_tokens': 14, 'prompt_tokens': 33, 'total_tokens': 47, 'completion_tokens_details': {'reasoning_tokens': 0}}, 'model_name': 'gpt-4o-mini-2024-07-18', 'system_fingerprint': 'fp_e9627b5346', 'finish_reason': 'stop', 'logprobs': None}, id='run-1d3afb4f-74a6-489d-a594-bd6d806cb144-0', usage_metadata={'input_tokens': 33, 'output_tokens': 14, 'total_tokens': 47})

And now we can see that we get a good response!

This is the basic idea underpinning a chatbot's ability to interact conversationally. So how do we best implement this?

Message persistence​

LangGraph implements a built-in persistence layer, making it ideal for chat applications that support multiple conversational turns.

Wrapping our chat model in a minimal LangGraph application allows us to automatically persist the message history, simplifying the development of multi-turn applications.

LangGraph comes with a simple in-memory checkpointer, which we use below. See its documentation for more detail, including how to use different persistence backends (e.g., SQLite or Postgres).

from langgraph.checkpoint.memory import MemorySaver
from langgraph.graph import START, MessagesState, StateGraph

# Define a new graph
workflow = StateGraph(state_schema=MessagesState)


# Define the function that calls the model
def call_model(state: MessagesState):
    response = model.invoke(state["messages"])
    return {"messages": response}


# Define the two nodes we will cycle between
workflow.add_edge(START, "model")
workflow.add_node("model", call_model)

# Add memory
memory = MemorySaver()
app = workflow.compile(checkpointer=memory)

We now need to create a config that we pass into the runnable every time. This config contains information that is not part of the input directly, but is still useful. In this case, we want to include a thread_id. This should look like:

config = {"configurable": {"thread_id": "abc123"}}

This enables us to support multiple conversation threads with a single application, a common requirement when your application has multiple users.

We can then invoke the application:

query = "Hi! I'm Bob."

input_messages = [HumanMessage(query)]
for event in app.stream({"messages": input_messages}, config, stream_mode="values"):
    event["messages"][-1].pretty_print()
================================ Human Message =================================

Hi! I'm Bob.
================================== Ai Message ==================================

Hi Bob! How can I assist you today?
query = "What's my name?"

input_messages = [HumanMessage(query)]
for event in app.stream({"messages": input_messages}, config, stream_mode="values"):
    event["messages"][-1].pretty_print()
================================ Human Message =================================

What's my name?
================================== Ai Message ==================================

Your name is Bob! How can I help you today?

Great! Our chatbot now remembers things about us. If we change the config to reference a different thread_id, we can see that it starts the conversation fresh.

config = {"configurable": {"thread_id": "abc234"}}

input_messages = [HumanMessage(query)]
for event in app.stream({"messages": input_messages}, config, stream_mode="values"):
    event["messages"][-1].pretty_print()
================================ Human Message =================================

What's my name?
================================== Ai Message ==================================

I'm sorry, but I don't know your name. How can I assist you today?

However, we can always go back to the original conversation (since we are persisting it in a database)

config = {"configurable": {"thread_id": "abc123"}}

input_messages = [HumanMessage(query)]
for event in app.stream({"messages": input_messages}, config, stream_mode="values"):
    event["messages"][-1].pretty_print()
================================ Human Message =================================

What's my name?
================================== Ai Message ==================================

Your name is Bob. What would you like to talk about?

This is how we can support a chatbot having conversations with many users!

tip

For async support, update the call_model node to be an async function and use .astream when invoking the application:

async def call_model(state: MessagesState):
    response = await model.ainvoke(state["messages"])
    return {"messages": response}


workflow = StateGraph(state_schema=MessagesState)


# Define the function that calls the model
def call_model(state: MessagesState):
    response = model.invoke(state["messages"])
    return {"messages": response}


# Define the two nodes we will cycle between
workflow.add_edge(START, "model")
workflow.add_node("model", call_model)

# Add memory
memory = MemorySaver()
app = workflow.compile(checkpointer=memory)

Right now, all we've done is add a simple persistence layer around the model. We can start to make the more complicated and personalized by adding in a prompt template.

Prompt templates​

Prompt Templates help to turn raw user information into a format that the LLM can work with. In this case, the raw user input is just a message, which we are passing to the LLM. Let's now make that a bit more complicated. First, let's add in a system message with some custom instructions (but still taking messages as input). Next, we'll add in more input besides just the messages.

First, let's add in a system message. To do this, we will create a ChatPromptTemplate. We will utilize MessagesPlaceholder to pass all the messages in.

from langchain_core.prompts import ChatPromptTemplate, MessagesPlaceholder

prompt = ChatPromptTemplate.from_messages(
    [
        (
            "system",
            "You talk like a pirate. Answer all questions to the best of your ability.",
        ),
        MessagesPlaceholder(variable_name="messages"),
    ]
)

We can now update our application to incorporate this template:

workflow = StateGraph(state_schema=MessagesState)


def call_model(state: MessagesState):
    chain = prompt | model
    response = chain.invoke(state)
    return {"messages": response}


workflow.add_edge(START, "model")
workflow.add_node("model", call_model)

memory = MemorySaver()
app = workflow.compile(checkpointer=memory)

We invoke the application in the same way:

config = {"configurable": {"thread_id": "abc345"}}
query = "Hi! I'm Jim."

input_messages = [HumanMessage(query)]
for event in app.stream({"messages": input_messages}, config, stream_mode="values"):
    event["messages"][-1].pretty_print()
================================ Human Message =================================

Hi! I'm Jim.
================================== Ai Message ==================================

Ahoy there, Jim! What brings ye to me ship today? Be there a question ye seek to answer, or perhaps a tale ye wish to hear? Arrr!
query = "What is my name?"

input_messages = [HumanMessage(query)]
for event in app.stream({"messages": input_messages}, config, stream_mode="values"):
    event["messages"][-1].pretty_print()
================================ Human Message =================================

What is my name?
================================== Ai Message ==================================

Yarr, ye be callin' yerself Jim, if I be rememberin' rightly! A fine name for a landlubber, indeed! What else can I do fer ye, matey?

Awesome! Let's now make our prompt a little bit more complicated. Let's assume that the prompt template now looks something like this:

prompt = ChatPromptTemplate.from_messages(
    [
        (
            "system",
            "You are a helpful assistant. Answer all questions to the best of your ability in {language}.",
        ),
        MessagesPlaceholder(variable_name="messages"),
    ]
)

Note that we have added a new language input to the prompt. Our application now has two parameters-- the input messages and language. We should update our application's state to reflect this:

from typing import Sequence

from langchain_core.messages import BaseMessage
from langgraph.graph.message import add_messages
from typing_extensions import Annotated, TypedDict


class State(TypedDict):
    messages: Annotated[Sequence[BaseMessage], add_messages]
    language: str


workflow = StateGraph(state_schema=State)


def call_model(state: State):
    chain = prompt | model
    response = chain.invoke(state)
    return {"messages": [response]}


workflow.add_edge(START, "model")
workflow.add_node("model", call_model)

memory = MemorySaver()
app = workflow.compile(checkpointer=memory)
API Reference:BaseMessage
config = {"configurable": {"thread_id": "abc456"}}
query = "Hi! I'm Bob."
language = "Spanish"

input_messages = [HumanMessage(query)]
for event in app.stream(
    {"messages": input_messages, "language": language},
    config,
    stream_mode="values",
):
    event["messages"][-1].pretty_print()
================================ Human Message =================================

Hi! I'm Bob.
================================== Ai Message ==================================

¡Hola, Bob! ¿Cómo puedo ayudarte hoy?

Note that the entire state is persisted, so we can omit parameters like language if no changes are desired:

query = "What is my name?"

input_messages = [HumanMessage(query)]
for event in app.stream(
    {"messages": input_messages, "language": language},
    config,
    stream_mode="values",
):
    event["messages"][-1].pretty_print()
================================ Human Message =================================

What is my name?
================================== Ai Message ==================================

Tu nombre es Bob. ¿Hay algo más en lo que te pueda ayudar?

To help you understand what's happening internally, check out this LangSmith trace.

Managing Conversation History​

One important concept to understand when building chatbots is how to manage conversation history. If left unmanaged, the list of messages will grow unbounded and potentially overflow the context window of the LLM. Therefore, it is important to add a step that limits the size of the messages you are passing in.

Importantly, you will want to do this BEFORE the prompt template but AFTER you load previous messages from Message History.

We can do this by adding a simple step in front of the prompt that modifies the messages key appropriately, and then wrap that new chain in the Message History class.

LangChain comes with a few built-in helpers for managing a list of messages. In this case we'll use the trim_messages helper to reduce how many messages we're sending to the model. The trimmer allows us to specify how many tokens we want to keep, along with other parameters like if we want to always keep the system message and whether to allow partial messages:

from langchain_core.messages import SystemMessage, trim_messages

trimmer = trim_messages(
    max_tokens=65,
    strategy="last",
    token_counter=model,
    include_system=True,
    allow_partial=False,
    start_on="human",
)

messages = [
    SystemMessage(content="you're a good assistant"),
    HumanMessage(content="hi! I'm bob"),
    AIMessage(content="hi!"),
    HumanMessage(content="I like vanilla ice cream"),
    AIMessage(content="nice"),
    HumanMessage(content="whats 2 + 2"),
    AIMessage(content="4"),
    HumanMessage(content="thanks"),
    AIMessage(content="no problem!"),
    HumanMessage(content="having fun?"),
    AIMessage(content="yes!"),
]

trimmer.invoke(messages)
[SystemMessage(content="you're a good assistant", additional_kwargs={}, response_metadata={}),
 HumanMessage(content='whats 2 + 2', additional_kwargs={}, response_metadata={}),
 AIMessage(content='4', additional_kwargs={}, response_metadata={}),
 HumanMessage(content='thanks', additional_kwargs={}, response_metadata={}),
 AIMessage(content='no problem!', additional_kwargs={}, response_metadata={}),
 HumanMessage(content='having fun?', additional_kwargs={}, response_metadata={}),
 AIMessage(content='yes!', additional_kwargs={}, response_metadata={})]

To use it in our chain, we just need to run the trimmer before we pass the messages input to our prompt. 

workflow = StateGraph(state_schema=State)


def call_model(state: State):
    chain = prompt | model
    trimmed_messages = trimmer.invoke(state["messages"])
    response = chain.invoke(
        {"messages": trimmed_messages, "language": state["language"]}
    )
    return {"messages": [response]}


workflow.add_edge(START, "model")
workflow.add_node("model", call_model)

memory = MemorySaver()
app = workflow.compile(checkpointer=memory)

Now if we try asking the model our name, it won't know it since we trimmed that part of the chat history:

config = {"configurable": {"thread_id": "abc567"}}
query = "What is my name?"
language = "English"

input_messages = messages + [HumanMessage(query)]
for event in app.stream(
    {"messages": input_messages, "language": language},
    config,
    stream_mode="values",
):
    event["messages"][-1].pretty_print()
================================ Human Message =================================

What is my name?
================================== Ai Message ==================================

I'm sorry, but I don't know your name. You haven't told me yet!

But if we ask about information that is within the last few messages, it remembers:

config = {"configurable": {"thread_id": "abc678"}}
query = "What math problem did I ask?"
language = "English"

input_messages = messages + [HumanMessage(query)]
for event in app.stream(
    {"messages": input_messages, "language": language},
    config,
    stream_mode="values",
):
    event["messages"][-1].pretty_print()
================================ Human Message =================================

What math problem did I ask?
================================== Ai Message ==================================

You asked the math problem "what's 2 + 2."

If you take a look at LangSmith, you can see exactly what is happening under the hood in the LangSmith trace.

Streaming​

Now we've got a functioning chatbot. However, one really important UX consideration for chatbot applications is streaming. LLMs can sometimes take a while to respond, and so in order to improve the user experience one thing that most applications do is stream back each token as it is generated. This allows the user to see progress.

It's actually super easy to do this!

By default, .stream in our LangGraph application streams application steps-- in this case, the single step of the model response. Setting stream_mode="messages" allows us to stream output tokens instead:

config = {"configurable": {"thread_id": "abc789"}}
query = "Hi I'm Todd, please tell me a joke."
language = "English"

input_messages = [HumanMessage(query)]
for chunk, metadata in app.stream(
    {"messages": input_messages, "language": language},
    config,
    stream_mode="messages",
):
    if isinstance(chunk, AIMessage):
        print(chunk.content, end="|")
|Hi| Todd|!| Here|’s| a| joke| for| you|:

|Why| did| the| scare|crow| win| an| award|?

|Because| he| was| outstanding| in| his| field|!||

Next Steps​

Now that you understand the basics of how to create a chatbot in LangChain, some more advanced tutorials you may be interested in are:

  • Conversational RAG: Enable a chatbot experience over an external source of data
  • Agents: Build a chatbot that can take actions

If you want to dive deeper on specifics, some things worth checking out are:

Was this page helpful?


You can also leave detailed feedback on GitHub.