@glennjones/prompt

Prompt

At the moment this code is more of a sketched idea than a library, but it is usable if you are interested in tring it out. It may be well your time worth looking at the Langchain JS which is the most heavily used tool of this type at the moment (Apr 2023). It is more advanced and has a lot of additional/useful features this tool does not have.

What is prompt

Its a light warpper around LLM inetrfaces such as OpenAI's ChatGPT. It provides:

• Prompt templating
• Simple interface prompt building in JavaScript
• Switching between model providers - OpenAI built-in, expanable to others
• Fine-tuned model selection for OpenAI
• Structured output for coding aginst
• Safe conversion output text to JSON
• Automatic API retries if service is busy
• Caching archecture

I have been using it to benchmark NER and Classification tasks with OpenAI prompts against other approcahes.

Install library

npm i @glennjones/prompt

Setup the objects

const Prompt = require('prompt');
const {Prompter, OpenAI} = Prompt;

Example of Image result for Named entity recognition (NER)

import {OpenAI, Prompter} from prompt

let model = new OpenAI(apiKey);
let prompt = new Prompter(model);

let result = await prompt.fit('ner', {
    domain: 'ux design',
    textInput: 'Senior UX Researcher, part-time 6 month contract, Edinburgh - Hybrid',
  });
console.log(result);

### Output

{
  promptTokens: 152,
  completionTokens: 99,
  totalTokens: 251,
  data: [
    { label: 'Position', text: 'Senior UX Researcher' },
    { label: 'Contract Type', text: 'part-time' },
    { label: 'Contract Duration', text: '6 month' },
    { label: 'Location', text: 'Edinburgh' },
    { label: 'Work Type', text: 'Hybrid' },
    { branch: 'UX Design', group: 'Research' }
  ]
}

TODO

• Get list of models from rest API
• Selection of fine tuned models works
• Request retry after fixed period
• Add Prompt hash
• Add Cache interface
• Add simple json file cache
• Add examples for all templates
• Add JSDocs to support typescript
• Add memory cache
• Parse templates to give user list all variables
• Check chat gtp 3.5 works
  • Check user context is working
  • Check prompt-chains are working
  • Consider extending template structure to have prompt-chains
• Look at allowing the upload and storage of templates

Based on the ideas/code of Promptify - https://github.com/promptslab/promptify