Aid: TypeScript Library for Typed LLM Interactions
A.I. :D
Aid is a TypeScript library designed for developers working with Large Language Models (LLMs) such as OpenAI's GPT-4 (including Vision) and GPT-3.5. The library focuses on ensuring consistent, typed outputs from LLM queries, enhancing the reliability and usability of LLM responses. Advanced users can leverage few-shot examples for more sophisticated use cases. It provides a structured and type-safe way to interact with LLMs.
Features
- Typed Response: Aid leverages TypeScript and JSON Schema to ensure consistent, reliable outputs from LLMs, adheres to the predefined schema.
- Task Based: Easily define custom tasks with specific input and output types, streamlining the process of LLM interactions.
- Few-Shot Learning Support: Allows for the provision of few-shot prompt examples to guide the LLM in producing the desired output.
- Visual Task Support: Includes support for visual tasks with image inputs, harnessing the power of OpenAI's GPT-4 Vision. Example
- OpenAI Integration: Integrates with OpenAI's official library to provide a seamless experience.
- Customizable: Allows for customization LLM models, just implement the
QueryEngine function. Example
Installation
pnpm install @ai-d/aid
Usage
Basic Setup
First, import the necessary modules and set up your OpenAI instance:
import { OpenAI } from "openai";
import { Aid } from "@ai-d/aid";
const openai = new OpenAI({ apiKey: "your-api-key" });
const aid = Aid.from(openai, { model: "gpt-4-1106-preview" });
Using GPT-4 Vision
import { OpenAI } from "openai";
import { Aid, OpenAIQuery } from "@ai-d/aid";
const openai = new OpenAI({ apiKey: "your-api-key" });
const aid = Aid.vision(
OpenAIQuery(openai, { model: "gpt-4-vision-preview", max_tokens: 2048 }),
);
Using Other LLM
For example, Cohere's Command.
import { Aid, CohereQuery } from "@ai-d/aid";
const aid = Aid.chat(
CohereQuery(COHERE_TOKEN, { model: "command" }),
);
You can implement your own QueryEngine function.
Creating a Custom Task
Define a custom task with expected output types:
import { z } from "zod";
const analyze = aid.task(
"Summarize and extract keywords",
z.object({
summary: z.string().max(300),
keywords: z.array(z.string().max(30)).max(10),
}),
);
Visual Task Example
const analyze = aid.task(
"Analyze the person in the image",
z.object({
gender: z.enum(["boy", "girl", "other"]),
age: z.enum(["child", "teen", "adult", "elderly"]),
emotion: z.enum(["happy", "sad", "angry", "surprised", "neutral"]),
clothing: z.string().max(100),
background: z.string().max(100),
}),
);
Executing a Task
Execute the task and handle the output:
const { result } = await analyze("Your input here, e.g. a news article");
console.log(result); // { summary: "...", keywords: ["...", "..."] }
Visual Task Example
const datauri = `data:image/png;base64,${fs.readFileSync("path/to/image.png" "base64")}`;
const { result } = await analyze({ images: [{ url: datauri }] });
console.log(result); // { "gender": "boy", "age": "teen", ... }
Advanced Usage with Few-Shot Examples
For more complex scenarios, you can use few-shot examples:
const run_advanced_task = aid.task(
"Some Advanced Task",
z.object({
// Define your output schema here
}),
{
examples: [
// Provide few-shot examples here
],
}
);
Formulation
Case Parameter -> (join) Task Defination -> (join) Format Constraint -> (perform) Query
Query and Format Constraint are defined and implemented by the QueryEngine and FormatEngine.
Task Defination is defined by the user with task method. Task Goal, Expected Schema, Examples, etc.
Case Parameter is defined by the user on each single call. Text, Image, etc.
Contributing
Contributions are welcome! Please submit pull requests with any bug fixes or feature enhancements.
