Skip to content

Latest commit

 

History

History
84 lines (61 loc) · 3.58 KB

tutorial.md

File metadata and controls

84 lines (61 loc) · 3.58 KB
Raw

How to generate a prompt:

The goal of Impromptu as a DSL is giving a way to split and structurize the process of generating prompts to a LLM.

The structure of the documents is:

  • Language that idicates in which langugae the propt will be written. It is optional (by default is English).

  • the set of assets: An asset is an structure that works as the fuction of a prgramming languge: the result of cumputing an asset by the DSL result in a generated prompt to use it in a LLM. There are different types of assets. Each one have a different syntax:

    • Composer
    composer <name>(<parameters>): <media>
<snippets>
separator=<string>              <- optional
    • Prompt
    prompt <name>(<parameters>): <media>
prefix=<snippets>               <- optional
core=<snippets>
suffix=<snippets>               <- optional
separator=<string>              <- optional
    • Chain (to be developed)
    chain <name>

The difference between a Composer and a Prompt it is mostly coneptual. A well-written prompt should consist of the description of the task plus some refiements such as set a role or give an examples. These parts make more sense in the suffix or prefix. On the other hand, a Composer is supposed to be used to concatenate several Asset

Parameters

Elements that are transmitted to the snippets of the asset. It can be a parameter (start wit @), or a multimodal (start with $).

Media

Indicates how the answer of the LLM would be (irrelevant right now). Media possible values:

text | image | audio | video | 3dobject

Snippets

A snippet is a subdivision of an asset and it can be

  • A TextLiteral: Text transmitted to the LLM as it is. Therefore, it is simply a string

  • A Trait: Keywords that helps to transmit a concept to the LLM. For example, The audience trait audience() transmits to the asset what the audicence target of the generated material is. One can see which traits are available in each LLM, and its poassible values in the follwing document.

  • A Input: It can also be a variable. A variable is a refrence to another snippet, typically a textLiteral

  • An AssetReuse: Similat to an imput,one can reutilize an asset in another asset as function, by declaring its name. By this method, one can create its custom traits.

Separator

Indicates which caracter is between an snippet and another. Each LLM has ha default saprator (for example, the dafault seprator for ChatGPT is " .")

Example

Suppose that we want to generate a poem about friendship for children using a LLP. The prompt that we would use will be similar to this one: Write an original poem about frendship. The content of this poem should be proper for children

Thus, three main ideas are transmitted to the LLM:

  1. Write a poem about friendship
  2. The poem has to be proper for children
  3. The poem should be original

Therefore, we can create an asset that generates a prompt with those ideas, each one represented by a snippet. As a result, we can transmit the second idea as a Trait (AudienceTrait):

prompt child_poem(): text 
core = "Write a poem about friendship",
       audience("children"),
       "The poem should be original"

We can go one more step, and generate the asset so that we can use it as well to generate poems with another topic and another possible audiences using parameters:

prompt write_poem(@topic,@target):text
core = "Write a poem about", @topic,
    audience(@target),
    "The poem should be original"