9753bccc71
- Add support for local build and linkchecking of docs - Add GitHub Action to automatically check links before prior to publication - Minor reformat of Contributing readme - Fix existing broken links Co-authored-by: Hunter Gerlach <hunter@huntergerlach.com> Co-authored-by: Hunter Gerlach <HunterGerlach@users.noreply.github.com> Co-authored-by: Hunter Gerlach <hunter@huntergerlach.com>
3.6 KiB
Key Concepts
Prompts
A prompt is the input to a language model. It is a string of text that is used to generate a response from the language model.
Prompt Templates
PromptTemplates are a way to create prompts in a reproducible way. They contain a template string, and a set of input variables. The template string can be formatted with the input variables to generate a prompt. The template string often contains instructions to the language model, a few shot examples, and a question to the language model.
PromptTemplates generically have a format method that takes in variables and returns a formatted string.
The most simple implementation of this is to have a template string with some variables in it, and then format it with the incoming variables.
More complex iterations dynamically construct the template string from few shot examples, etc.
To learn more about PromptTemplates, see Prompt Templates.
As an example, consider the following template string:
"""
Predict the capital of a country.
Country: {country}
Capital:
"""
Input Variables
Input variables are the variables that are used to fill in the template string. In the example above, the input variable is country.
Given an input variable, the PromptTemplate can generate a prompt by filling in the template string with the input variable. For example, if the input variable is United States, the template string can be formatted to generate the following prompt:
"""
Predict the capital of a country.
Country: United States
Capital:
"""
Few Shot Examples
Few shot examples refer to in-context examples that are provided to a language model as part of a prompt. The examples can be used to help the language model understand the context of the prompt, and as a result generate a better response. Few shot examples can contain both positive and negative examples about the expected response.
Below, we list out some few shot examples that may be relevant for the task of predicting the capital of a country.
Country: United States
Capital: Washington, D.C.
Country: Canada
Capital: Ottawa
To learn more about how to provide few shot examples, see Few Shot Examples.
Example selection
If there are multiple examples that are relevant to a prompt, it is important to select the most relevant examples. Generally, the quality of the response from the LLM can be significantly improved by selecting the most relevant examples. This is because the language model will be able to better understand the context of the prompt, and also potentially learn failure modes to avoid.
To help the user with selecting the most relevant examples, we provide example selectors that select the most relevant based on different criteria, such as length, semantic similarity, etc. The example selector takes in a list of examples and returns a list of selected examples, formatted as a string. The user can also provide their own example selector. To learn more about example selectors, see Example Selection.
Serialization
To make it easy to share PromptTemplates, we provide a serialize method that returns a JSON string. The JSON string can be saved to a file, and then loaded back into a PromptTemplate using the deserialize method. This allows users to share PromptTemplates with others, and also to save them for later use.
To learn more about serialization, see Serialization.