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

1pnpm install @ai-d/aid

Usage

Basic Setup

First, import the necessary modules and set up your OpenAI instance:

1import { OpenAI } from "openai";
2import { Aid } from "@ai-d/aid";
3
4const openai = new OpenAI({ apiKey: "your-api-key" });
5const aid = Aid.from(openai, { model: "gpt-4-1106-preview" });

1import { OpenAI } from "openai";
2import { Aid, OpenAIQuery } from "@ai-d/aid";
3
4const openai = new OpenAI({ apiKey: "your-api-key" });
5const aid = Aid.vision(
6    OpenAIQuery(openai, { model: "gpt-4-vision-preview", max_tokens: 2048 }),
7);

1import { Aid, CohereQuery } from "@ai-d/aid";
2
3const aid = Aid.chat(
4    CohereQuery(COHERE_TOKEN, { model: "command" }),
5);

Creating a Custom Task

Define a custom task with expected output types:

1import { z } from "zod";
2
3const analyze = aid.task(
4    "Summarize and extract keywords",
5    z.object({
6        summary: z.string().max(300),
7        keywords: z.array(z.string().max(30)).max(10),
8    }),
9);

1const analyze = aid.task(
2    "Analyze the person in the image",
3    z.object({
4        gender: z.enum(["boy", "girl", "other"]),
5        age: z.enum(["child", "teen", "adult", "elderly"]),
6        emotion: z.enum(["happy", "sad", "angry", "surprised", "neutral"]),
7        clothing: z.string().max(100),
8        background: z.string().max(100),
9    }),
10);

Executing a Task

Execute the task and handle the output:

1const { result } = await analyze("Your input here, e.g. a news article");
2console.log(result); // { summary: "...", keywords: ["...", "..."] }

1const datauri = `data:image/png;base64,${fs.readFileSync("path/to/image.png" "base64")}`;
2
3const { result } = await analyze({ images: [{ url: datauri }] });
4console.log(result); // { "gender": "boy", "age": "teen", ... }

Advanced Usage with Few-Shot Examples

For more complex scenarios, you can use few-shot examples:

1const run_advanced_task = aid.task(
2    "Some Advanced Task",
3    z.object({
4        // Define your output schema here
5    }),
6    {
7        examples: [
8            // Provide few-shot examples here
9        ],
10    }
11);

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.