Skip to main content

Prompts

Prompts are the input provided to language models. A prompt typically contains instructions, examples, and a question or request for the model. LangChain provides tools to help construct and manage prompts.

Prompt Templates​

A prompt template is a reproducible way to generate a prompt dynamically. It contains a template string that can take in parameters and generate a full prompt.

For example:

template = "You are a {profession}. What is your advice about {topic}?"

prompt = PromptTemplate(template=template, input_variables=["profession", "topic"])

prompt.format(profession="doctor", topic="getting more sleep")

Would generate the prompt:

You are a doctor. What is your advice about getting more sleep?

Prompt templates allow prompts to be dynamically generated from parameters. They can contain:

  • Instructions to the model
  • Few shot examples
  • A question or request

See the prompt templates documentation for more details and examples.

Chat Prompt Templates​

Chat models take prompts in the form of a list of chat messages, rather than a single text string. LangChain provides ChatPromptTemplate to help construct prompts for chat models.

A ChatPromptTemplate consists of one or more MessagePromptTemplates that each generate a chat message. For example:

system_template = "You are an assistant that translates {input_language} to {output_language}."
system_prompt = SystemMessagePromptTemplate(system_template) 

human_template = "Hello! Can you translate this: {text}" 
human_prompt = HumanMessagePromptTemplate(human_template)

chat_prompt = ChatPromptTemplate(messages=[system_prompt, human_prompt])

chat_prompt.format(
   input_language="English", 
   output_language="Spanish",
   text="I love traveling"
)

Would generate:

[
  SystemMessage(You are an assistant that translates English to Spanish.),
  HumanMessage(Hello! Can you translate this: I love traveling)
]

See the chat prompt templates documentation for more details and examples.

LLMs vs Chat Models​

There are two main types of models in LangChain:

  • LLMs: Take a text string as input and return a text string. For example, GPT-3.
  • Chat Models: Take a list of chat messages as input and return a chat message. For example, GPT-4.

LLMs and Chat Models have slightly different interfaces, but both implement the base BaseLanguageModel interface in LangChain. This makes it possible to swap between LLMs and Chat Models.

When using a specific model, it is best to use the model-specific methods (i.e. predict for LLMs). But the base interface allows writing applications that are agnostic to model type.

See the models documentation for more details.