Skip to main content

Prompts for Chat Models

Prompts are a key part of how chat models work. Unlike traditional language models that take raw text as input, chat models are designed to have a conversation and take a series of chat messages as input.

Constructing effective prompts is crucial to get the most out of chat models. LangChain provides several classes to make building prompts for chat models easy.

Overview​

There are 3 main classes:

  • PromptTemplate: Base class for creating parametrized prompt templates.
  • MessagePromptTemplate: Creates a prompt template associated with a specific role like system, assistant, or human.
  • ChatPromptTemplate: Creates a full prompt made of MessagePromptTemplates. Used to format prompts to pass to chat models.

MessagePromptTemplate​

A MessagePromptTemplate represents a single message in a conversation and is tied to a specific role.

For example, you can create a template for a system message:

template = "You are a helpful assistant that translates {input_language} to {output_language}."
system_message = SystemMessagePromptTemplate.from_template(template)

Or directly:

prompt = PromptTemplate(
  template="You are a helpful assistant that translates {input_language} to {output_language}.",
  input_variables=["input_language", "output_language"]
)
system_message = SystemMessagePromptTemplate(prompt=prompt)

Some common roles:

  • SystemMessagePromptTemplate
  • AIMessagePromptTemplate
  • HumanMessagePromptTemplate

Example​

Here is an example workflow creating a system message template:

# Template 
template = "You are a helpful assistant that translates {input_language} to {output_language}."

# Create template
system_message = SystemMessagePromptTemplate.from_template(template)

# Format template 
formatted_message = system_message.format(
  input_language="English",
  output_language="French"  
)

print(formatted_message)

# You are a helpful assistant that translates English to French.

ChatPromptTemplate​

A ChatPromptTemplate represents a full conversation prompt made of MessagePromptTemplates.

For example: 

system_message = SystemMessagePromptTemplate(...)
human_message = HumanMessagePromptTemplate(...)

chat_prompt = ChatPromptTemplate.from_messages([
  system_message,
  human_message
])

You can format the full prompt:

formatted_messages = chat_prompt.format_prompt(
  input_language="English",
  output_language="French",
  text="I love programming."  
).to_messages()

And pass to a chat model:

response = chat(formatted_messages)

Example​

Full prompt formatting example:

# System and human templates
system_template = "You are a helpful assistant that translates {input_language} to {output_language}."
human_template = "{text}"

system_message = SystemMessagePromptTemplate.from_template(system_template)  
human_message = HumanMessagePromptTemplate.from_template(human_template)

# Chat template
chat_prompt = ChatPromptTemplate.from_messages([
  system_message,
  human_message
])

# Format prompt
formatted_messages = chat_prompt.format_prompt(
  input_language="English", 
  output_language="French",
  text="I love programming."
).to_messages() 

# Pass to chat model 
response = chat(formatted_messages)

Summary​

The prompt classes in LangChain provide an easy way to construct prompts specifically for chat models. Key points:

  • MessagePromptTemplate represents a single chat message
  • ChatPromptTemplate represents a full conversation prompt
  • Format templates using input variables
  • Convert formatted templates to chat messages to pass to models

Using these templates can help you fully leverage chat capabilities compared to just passing raw text.