@isdk/ai-tool-prompt

ai-tool-prompt

• Management of various prompts
• Collection The LLM System template prompts With Enhancement (guessing corresponding system templates based on the model file name)
• Multiple versions support under the same system prompt template file
• Recommendation of parameters for LLM models
• Extensible multiple template formats to write system prompt template: jinjia2 (hf), golang (ollama), python fstring (langchain). Here jinjia2 is used as the default template format.
• Extensible file format: yaml, json. Here yaml format is used as the default file format.

Installation

npm install @isdk/ai-tool @isdk/ai-tool-prompt

Prompt Structure

• _id: string (unique) - id of the prompt
• templateFormat: 'hf'|'ollama'|'langchain' - template format of the prompt, defaults to 'hf'
• template: string - template of the prompt, eg, '{{messages[0].content}}'
• type: 'system'|'tool'|'char' - type of the prompt
• prompt: object - default prompt template data
• modelPattern: string|RegExpr|(string|RegExpr)[]|{[version:string]: string|RegExpr} - pattern(s) to match the LLM model file name
• parameters: {[ruleMatched: string]: object]} - the default parameters of LLM models
  • the key is the matched rule named group: id or group[1], it is lowercase always.
  • the value is parameters of the LLM model
• extends: string - the id of the parent prompt to inherit
• version: {[name: string]: object} - the version/feature info of the prompt
  • the key is the name of the version/feature
  • the value is the prompt to overwrite.

The Current Default System Prompt Templates: here

The YAML(.yaml|.yml) and JSON(.json) file format are supported.

Usage

The following code snippet shows how to use the prompts by AIPromptsFunc sqilteDB.

import { AIPromptsFunc } from '@isdk/ai-tool-prompt';

const promptsDB = new AIPromptsFunc(AIPromptsName, {dbPath: ':memory:'})
// load default system template prompts automatically into sqliteDB
await promptsDB.initData()

const promptInfo = await promptsDB.$getPrompt({ model: 'qwen2.5-7b-it.gguf', type: 'system' })
console.log(promptInfo)
// {prompt: ..., version: 'qwen'}

The following code snippet shows how to use the prompts by findPrompt function.

import { parse as parseIni } from 'ini'
import { ConfigFile, getConfigs } from '@isdk/ai-tool';
import { findPrompt } from '@isdk/ai-tool-prompt';

// add your prompt file format parser
ConfigFile.register(['.ini'], parseIni)

// load all prompts from your prompts directory, searh all .ini|json|yaml|yml files
const prompts = getConfigs('dir/your-prompts-dir')
const promptInfo = await findPrompt(prompts, 'qwen2.5-7b-it.gguf', { type: 'system' })