relay 🤖

A framework for agents

Features

• Simple Convo class to manage conversation history (user/model turns).
• GeminiModel implementation for interacting with Google Gemini.
• Supports basic text generation.
• Supports requesting JSON output via schemas.
• Type-safe interactions with Deno/TypeScript.

Installation

This library is intended for use with Deno and can be imported directly from JSR.

import { Convo, GeminiModel } from "jsr:@xfth/relay@^0.1.0";

Usage

Setup

1. Make sure you have Deno installed.
2. Create a .env file in your project root:

   GEMINI_API_KEY="your_actual_gemini_api_key"
   
   
   

3. Ensure your script has environment variable access (--allow-env) and network access (--allow-net) when running.

Basic Example

import { Convo, GeminiModel } from "jsr:@xfth/relay@^0.1.0";

// This will automatically load the API key from .env
const model = new GeminiModel();

const c = new Convo("You are a helpful assistant.");
c.user("What is the capital of France?");

const response = c.model(await model.gen(c));

console.log(response); // Output: Paris
console.log(c.messages); // Shows the full conversation history

JSON Output Example

import { Convo, GeminiModel } from "jsr:@xfth/relay@^0.1.0";

const model = new GeminiModel();

const c = new Convo("Extract the city and country from the user query.");
c.user("I'm visiting Berlin in Germany next month.");

type Location = {
  city: string;
  country: string;
};

const locationSchema = {
  type: "object",
  properties: {
    city: { type: "string", description: "The name of the city mentioned." },
    country: { type: "string", description: "The name of the country mentioned." },
  },
  required: ["city", "country"],
};

const extractedData = c.model(
  await model.gen<Location>(c, locationSchema)
);

console.log(extractedData); // Output: { city: "Berlin", country: "Germany" }

API Overview

• Convo: Manages the conversation history.
  • constructor(systemPrompt?: string)
  • user(content: string | object): Adds a user message.
  • model(content: string | object): Adds a model message.
  • clone(): Creates a deep copy of the conversation.
  • messages: Array of Message objects ({ role: string, mime: string, content: string }).
  • system: The system prompt string.
• GeminiModel: Implementation for Google Gemini.
  • constructor(baseModel?: string): Optionally specify a Gemini model name (defaults to "gemini-1.5-flash").
  • gen<T>(convo: Convo, schema?: object): Generates a response. If schema is provided, attempts to return parsed JSON matching the schema (T), otherwise returns a string (T).

License

This project is licensed under the CC0 1.0 Universal license.