@openassistant/core

The core package is built based on Vercel AI SDK and provides:

• a uniform interface for different AI providers
• allows you to integrate powerful tools for tasks like data analysis, visualization, mapping, etc.
• allows you to easily create your own tools by:
  • providing your own context (e.g. data, callbacks etc.) for the tool execution
  • providing your own UI component for rendering the tool result
  • passing the result from the tool execution to the tool UI component or next tool execution.

Getting Started

Installation

Install the core package:

npm install @openassistant/core

Usage

Then, you can use the OpenAssistant in your application. For example:

import { createAssistant } from '@openassistant/core';

// get the singleton assistant instance
const assistant = await createAssistant({
  name: 'assistant',
  modelProvider: 'openai',
  model: 'gpt-4o',
  apiKey: 'your-api-key',
  version: '0.0.1',
  instructions: 'You are a helpful assistant',
  // functions: {{}},
  // abortController: null
});

// now you can send prompts to the assistant
await assistant.processTextMessage({
  textMessage: 'Hello, how are you?',
  streamMessageCallback: ({ isCompleted, message }) => {
    console.log(isCompleted, message);
  },
});

See the source code of the example 🔗 here.

:::tip

If you want to use Google Gemini as the model provider, you can do the following:

Install vercel google gemini client:

npm install @ai-sdk/google

Then, you can use update the assistant configuration to use Google Gemini.

OpenAssistant also supports the following model providers:

| Model Provider | Models | Dependency | | -------------- | -------------------------------------------------------------------------------------------------- | ------------------ | | OpenAI | link | @ai-sdk/openai | | Google | models | @ai-sdk/google | | Anthropic | models | @ai-sdk/anthropic | | DeepSeek | models | @ai-sdk/deepseek | | xAI | models | @ai-sdk/xai | | Ollama | models | ollama-ai-provider-v2 |

:::

Common Tools

OpenAssistant provides a set of common tools that are useful for different tasks.

Think

The think tool is a tool that enables LLM to think about a problem before solving it.

Implementation

export const think = tool({
  parameters: z.object({
    question: z.string().describe('The question to think about'),
  }),
  execute: async ({ question }) => {
    return {
      llmResult: {
        success: true,
        result: {
          question,
          instruction: `
- Before executing the plan, please summarize the plan for using the tools.
- If the tools are missing parameters, please ask the user to provide the parameters.
- When executing the plan, please try to fix the error if there is any.
- After executing the plan, please summarize the result and provide the result in a markdown format.
`,
        },
      },
    };
  },
});

Usage

import { tool } from '@openassistant/utils';
import { z } from 'zod';

const weather =  tool({
      description: 'Get the weather in a city from a weather station',
      parameters: z
        .object({ cityName: z.string() })
        .describe('The city name to get the weather for'),
      execute: async ({ cityName }, options) => {
        const getStation = options.context?.getStation;
        const station = getStation ? await getStation(cityName) : null;
        return {
          llmResult: {
            success: true,
            result: `The weather in ${cityName} is sunny from weather station ${station}.`,
          },
          additionalData: {
            weather: 'sunny',
            station,
          },
        };
      },
      context: {
        getStation: async (cityName: string) => {
          const stations = {
            'New York': '123',
            'Los Angeles': '456',
            Chicago: '789',
          };
          return stations[cityName];
        },
      },
    });

const assistant = await createAssistant({
  name: 'assistant',
  modelProvider: 'openai',
  model: 'gpt-4o',
  apiKey: 'your-api-key',
  version: '0.0.1',
  instructions: 'You are a helpful assistant',
  tools: { {think, weather} },
});

Example output:

When you prompt the assistant with:

What is the weather in New York?

The assistant will think about the problem and then use the weather tool to get the weather in New York.