Skip to main content

Build a Chatbot

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.

Concepts​

Here are a few of the high-level components we'll be working with:

  • Chat Models. The chatbot interface is based around messages rather than raw text, and therefore is best suited to Chat Models rather than text LLMs.
  • Prompt Templates, which simplify the process of assembling prompts that combine default messages, user input, chat history, and (optionally) additional retrieved context.
  • Chat History, which allows a chatbot to "remember" past interactions and take them into account when responding to followup questions.
  • Debugging and tracing your application using LangSmith

We'll cover how to fit the above components together to create a powerful conversational chatbot.

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​

To install LangChain run:

pip install langchain

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 interchangably - 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-3.5-turbo")

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='Hello Bob! How can I assist you today?', response_metadata={'token_usage': {'completion_tokens': 10, 'prompt_tokens': 12, 'total_tokens': 22}, 'model_name': 'gpt-3.5-turbo-0125', 'system_fingerprint': None, 'finish_reason': 'stop', 'logprobs': None}, id='run-8ecc8a9f-8b32-49ad-8e41-5caa26282f76-0', usage_metadata={'input_tokens': 12, 'output_tokens': 10, 'total_tokens': 22})

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, I don't have access to that information.", response_metadata={'token_usage': {'completion_tokens': 13, 'prompt_tokens': 12, 'total_tokens': 25}, 'model_name': 'gpt-3.5-turbo-0125', 'system_fingerprint': None, 'finish_reason': 'stop', 'logprobs': None}, id='run-4e0066e8-0dcc-4aea-b4f9-b9029c81724f-0', usage_metadata={'input_tokens': 12, 'output_tokens': 13, 'total_tokens': 25})

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 assist you today, Bob?', response_metadata={'token_usage': {'completion_tokens': 14, 'prompt_tokens': 35, 'total_tokens': 49}, 'model_name': 'gpt-3.5-turbo-0125', 'system_fingerprint': None, 'finish_reason': 'stop', 'logprobs': None}, id='run-c377d868-1bfe-491a-82fb-1f9122939796-0', usage_metadata={'input_tokens': 35, 'output_tokens': 14, 'total_tokens': 49})

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 History​

We can use a Message History class to wrap our model and make it stateful. This will keep track of inputs and outputs of the model, and store them in some datastore. Future interactions will then load those messages and pass them into the chain as part of the input. Let's see how to use this!

First, let's make sure to install langchain-community, as we will be using an integration in there to store message history.

# ! pip install langchain_community

After that, we can import the relevant classes and set up our chain which wraps the model and adds in this message history. A key part here is the function we pass into as the get_session_history. This function is expected to take in a session_id and return a Message History object. This session_id is used to distinguish between separate conversations, and should be passed in as part of the config when calling the new chain (we'll show how to do that.

from langchain_community.chat_message_histories import ChatMessageHistory
from langchain_core.chat_history import BaseChatMessageHistory
from langchain_core.runnables.history import RunnableWithMessageHistory

store = {}


def get_session_history(session_id: str) -> BaseChatMessageHistory:
if session_id not in store:
store[session_id] = ChatMessageHistory()
return store[session_id]


with_message_history = RunnableWithMessageHistory(model, get_session_history)

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 session_id. This should look like:

config = {"configurable": {"session_id": "abc2"}}
response = with_message_history.invoke(
[HumanMessage(content="Hi! I'm Bob")],
config=config,
)

response.content
Parent run 9bdaa45d-604e-4891-9b0a-28754985f10b not found for run 271bd46a-f980-407a-af8a-9399420bce8d. Treating as a root run.
'Hello Bob! How can I assist you today?'
response = with_message_history.invoke(
[HumanMessage(content="What's my name?")],
config=config,
)

response.content
Parent run 16482292-535c-449d-8a9d-d0fccf5112eb not found for run 7f2e501a-d5b4-4d8c-924b-aae9eb9d7267. Treating as a root run.
'Your name is Bob. How can I assist you today, Bob?'

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

config = {"configurable": {"session_id": "abc3"}}

response = with_message_history.invoke(
[HumanMessage(content="What's my name?")],
config=config,
)

response.content
Parent run c14d7130-04c5-445f-9e22-442f7c7e8f07 not found for run 946beadc-5cf1-468f-bac4-ca5ddc10ea73. Treating as a root run.
"I'm sorry, I don't know your name as you have not provided it."

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

config = {"configurable": {"session_id": "abc2"}}

response = with_message_history.invoke(
[HumanMessage(content="What's my name?")],
config=config,
)

response.content
Parent run 4f61611c-3875-4b2d-9f89-af452866d55a not found for run 066a30b1-bbb0-4fee-a035-7fdb41c28d91. Treating as a root run.
'Your name is Bob. How can I assist you today, Bob?'

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

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 are a helpful assistant. Answer all questions to the best of your ability.",
),
MessagesPlaceholder(variable_name="messages"),
]
)

chain = prompt | model

Note that this slightly changes the input type - rather than pass in a list of messages, we are now passing in a dictionary with a messages key where that contains a list of messages.

response = chain.invoke({"messages": [HumanMessage(content="hi! I'm bob")]})

response.content
'Hello Bob! How can I assist you today?'

We can now wrap this in the same Messages History object as before

with_message_history = RunnableWithMessageHistory(chain, get_session_history)
config = {"configurable": {"session_id": "abc5"}}
response = with_message_history.invoke(
[HumanMessage(content="Hi! I'm Jim")],
config=config,
)

response.content
Parent run 51e624b3-19fd-435f-b580-2a3e4f2d0dc9 not found for run b411f007-b2ad-48c3-968c-aa5ecbb58aea. Treating as a root run.
'Hello Jim! How can I assist you today?'
response = with_message_history.invoke(
[HumanMessage(content="What's my name?")],
config=config,
)

response.content
Parent run a30b22cd-698f-48a1-94a0-1a172242e292 not found for run 52b0b60d-5d2a-4610-a572-037602792ad6. Treating as a root run.
'Your name is Jim.'

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"),
]
)

chain = prompt | model

Note that we have added a new language input to the prompt. We can now invoke the chain and pass in a language of our choice.

response = chain.invoke(
{"messages": [HumanMessage(content="hi! I'm bob")], "language": "Spanish"}
)

response.content
'Β‘Hola, Bob! ΒΏEn quΓ© puedo ayudarte hoy?'

Let's now wrap this more complicated chain in a Message History class. This time, because there are multiple keys in the input, we need to specify the correct key to use to save the chat history.

with_message_history = RunnableWithMessageHistory(
chain,
get_session_history,
input_messages_key="messages",
)
config = {"configurable": {"session_id": "abc11"}}
response = with_message_history.invoke(
{"messages": [HumanMessage(content="hi! I'm todd")], "language": "Spanish"},
config=config,
)

response.content
Parent run d02b7778-4a91-4831-ace9-b33bb456dc90 not found for run ee0a20dd-5b9e-4862-b3c9-8e2e72b8eb82. Treating as a root run.
'Β‘Hola Todd! ΒΏEn quΓ© puedo ayudarte hoy?'
response = with_message_history.invoke(
{"messages": [HumanMessage(content="whats my name?")], "language": "Spanish"},
config=config,
)

response.content
Parent run 12422d4c-6494-490e-845e-08dcc1c6a4b9 not found for run a82eb759-f51d-4488-871b-6e2d601b4128. Treating as a root run.
'Tu nombre es Todd.'

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"),
HumanMessage(content='whats 2 + 2'),
AIMessage(content='4'),
HumanMessage(content='thanks'),
AIMessage(content='no problem!'),
HumanMessage(content='having fun?'),
AIMessage(content='yes!')]

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

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

from operator import itemgetter

from langchain_core.runnables import RunnablePassthrough

chain = (
RunnablePassthrough.assign(messages=itemgetter("messages") | trimmer)
| prompt
| model
)

response = chain.invoke(
{
"messages": messages + [HumanMessage(content="what's my name?")],
"language": "English",
}
)
response.content
API Reference:RunnablePassthrough
"I'm sorry, I don't have access to personal information. How can I assist you today?"

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

response = chain.invoke(
{
"messages": messages + [HumanMessage(content="what math problem did i ask")],
"language": "English",
}
)
response.content
'You asked "what\'s 2 + 2?"'

Let's now wrap this in the Message History

with_message_history = RunnableWithMessageHistory(
chain,
get_session_history,
input_messages_key="messages",
)

config = {"configurable": {"session_id": "abc20"}}
response = with_message_history.invoke(
{
"messages": messages + [HumanMessage(content="whats my name?")],
"language": "English",
},
config=config,
)

response.content
Parent run e1bb2af3-192b-4bd1-8734-6d2dff1d80b6 not found for run 0c734998-cf16-4708-8658-043a6c7b4a91. Treating as a root run.
"I'm sorry, I don't have access to your name. How can I assist you today?"

As expected, the first message where we stated our name has been trimmed. Plus there's now two new messages in the chat history (our latest question and the latest response). This means that even more information that used to be accessible in our conversation history is no longer available! In this case our initial math question has been trimmed from the history as well, so the model no longer knows about it:

response = with_message_history.invoke(
{
"messages": [HumanMessage(content="what math problem did i ask?")],
"language": "English",
},
config=config,
)

response.content
Parent run 181a1f04-9176-4837-80e8-ce74866775a2 not found for run ad402c5a-8341-4c62-ac58-cdf923b3b9ec. Treating as a root run.
"You haven't asked a math problem yet. Feel free to ask any math question you have, and I'll do my best to help you with it."

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 function chatbot. However, one really important UX consideration for chatbot application is streaming. LLMs can sometimes take a while to respond, and so in order to improve the user experience one thing that most application 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!

All chains expose a .stream method, and ones that use message history are no different. We can simply use that method to get back a streaming response.

config = {"configurable": {"session_id": "abc15"}}
for r in with_message_history.stream(
{
"messages": [HumanMessage(content="hi! I'm todd. tell me a joke")],
"language": "English",
},
config=config,
):
print(r.content, end="|")
Parent run e0ee52b6-1261-4f2d-98ca-f78c9019684b not found for run 0f6d7995-c32c-4bdb-b7a6-b3d932c13389. Treating as a root run.
``````output
|Sure|,| Todd|!| Here|'s| a| joke| for| you|:

|Why| don|'t| scientists| trust| atoms|?

|Because| they| make| up| everything|!||

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 leave detailed feedback on GitHub.