JSON Format for Smart Tools

A portable way to share SkyDeck.AI smart tools written in Python code

Overall

To create a smart tool on SkyDeck.AI, you need to upload a set of files according to the specifications mentioned in the File Structure section. Once uploaded, our platform will perform the initial setup, which may take a few minutes. Afterward, the tool will be available in the GenStudio Workspace.

File Structure

<tool_name>.json

The tool's behavior is also configured through a JSON file. Here is a brief overview of the key fields in the configuration:

version: The current version of the tool.
tool_name: The name of the tool. This name should be unique in your workspace.
tool_code: Contains the Python code to be executed. More details on this field will be on the next section.
description: A brief description of what the tool does.
usage_notes: Instructions on how to use the tool.
model_version: Specify the models that are allowed to use for follow-up questions. To specify all models, use ["gpt-4", "gpt-3.5", "claude", "chat-bison"].
creator: Information about the creator of the tool, including name, email, and organization.
variables: An array of variables used by the tool. Each variable has a name, description, and default value. The order of variables in the UI follows the order in this array.
expected_output: The type of output produced by the tool. During the development stage, the value should always be text.
avatar_type: The format of the avatar used in the tool's UI.
timestamp: The date and time when the tool was last updated.
requirements: specifies the required packages to run the script in tool_code.
avatar: the string representing the logo of this tool

tool_code convention:

This script outlines the functioning of your tool. The main component of this script is the execute function, which has the following requirements:

The function should have a single input parameter called variables, which is a dictionary. Each key in this dictionary corresponds to a field that the user would input into your tool.
The function should return a string, which will be displayed as the response on the GenStudio UI.

Example Tools

Image generation using DALL-E 2

Description:

This tool accepts an image description as input and generates a corresponding URL for the image. The output includes the URL along with an expiration note. The tool functions by sending the query to the OpenAI DALL-E API and retrieving the response.

Input:

Description: Image description, e.g., "A white furry cat"

Output:

A message with the generated URL for the image along with the expiration note.

Python script (which would be a field inside image_generation.json):

import openai
def execute(variables):
    openai.api_key = '<USER API KEY>'
    description = variables['Description']
    response = openai.Image.create(
        prompt=description,
        n=1,
        size="1024x1024"
    )
    image_url = response['data'][0]['url']
    return f"Here is the link to your image:{image_url}. The link will expire in 1 hour."

image_generation.json

{
    "version": "0.1",
    "metadata": {
        "tool_name": "Image generation",
        "tool_code": "import openai\n\ndef execute(variables):\n openai.api_key = ''\n description = variables['Description']\n response = openai.Image.create(\n prompt=description,\n n=1,\n size=\"1024x1024\"\n )\n\n image_url = response['data'][0]['url']\n return f\"Here is the link to your image:{image_url}. The link will be expired in 1 hour.\"\n",
        "description": "Generated an image based on the description with OpenAI's DALL-E model.",
        "usage_notes": "Describe the image in detail and put it in the description field. An URL of the image will be returned. The lifetime of the URL is about 1 hour, so make sure to download it before expired",
        "model_version": ["gpt-3.5","gpt-3.5-turbo", "gpt-4", "claude"],
        "creator": {
            "name": "SkyDeck AI",
            "email": "[email protected]",
            "organization": "East Agile"
        },
        "variables": [
            {
                "name": "Description",
                "description": "Image description",
                "default": "a white siamese cat"
            }
        ],
        "expected_output": {
            "type": "text"
        },
        "avatar_type": "base64",
        "timestamp": "2023-05-23T10:00:00Z",
        "requirements": "openai>=0.27.4",
        "avatar": ""
    }
}

Real-time weather report with Open-Meteo API

Description:

This tool leverages the Open-Meteo API to provide real-time weather information based on users' questions. By asking a question about the weather, such as temperature, precipitation, or wind conditions, the tool retrieves the most relevant and up-to-date data.

The functioning of this tool relies on APIChain - a feature from the LangChain library - to access the Open-Meteo API documentation. This enables the tool to learn how to make the correct API calls and retrieve the required information seamlessly.

Input:

Question: Ask a specific question about the weather, e.g., "What is the current temperature in New York City?"

Output:

A response providing the requested weather information.

Python script (which would be a field inside weather_reporter.json):

from langchain.chains.api import open_meteo_docs
from langchain.chat_models import ChatOpenAI
from langchain.chains import APIChain
def execute(variables):
    question = variables['Question']
    llm = ChatOpenAI(
        model_name='gpt-3.5-turbo',
        openai_api_key='<USER API KEY>'
    )
    api_chain = APIChain.from_llm_and_api_docs(
        llm, open_meteo_docs.OPEN_METEO_DOCS, verbose=False
    )
    result = api_chain.run(question)
    return result

Note that Langchain is now integrated with Rememberizer. See https://python.langchain.com/docs/integrations/retrievers/rememberizer/

weather_reporter.json

{
  "version": "0.1",
  "metadata": {
    "tool_name": "Weather Reporter",
    "tool_code": "from langchain.chains.api import open_meteo_docs\nfrom langchain.chat_models import ChatOpenAI\nfrom langchain.chains import APIChain\n\n\ndef execute(variables):\n question = variables['Question']\n llm = ChatOpenAI(model_name='gpt-3.5-turbo',\n openai_api_key='')\n api_chain = APIChain.from_llm_and_api_docs(\n llm, open_meteo_docs.OPEN_METEO_DOCS, verbose=False)\n result = api_chain.run(question)\n return result\n",
    "description": "Leverage the Open-Meteo API to retrieve real-time weather details",
    "usage_notes": "Enter your weather-related question in the provided field",
    "model_version": ["gpt-3.5", "gpt-3.5-turbo", "gpt-4", "claude"],
    "creator": {
      "name": "SkyDeck AI",
      "email": "[email protected]",
      "organization": "East Agile"
    },
    "variables": [
      {
        "name": "Question",
        "description": "Inquire about the weather conditions",
        "default": "What is the current temperature in Munich, Germany, expressed in degrees Celsius?"
      }
    ],
    "expected_output": {
      "type": "text"
    },
    "avatar_type": "base64",
    "timestamp": "2023-07-13T10:00:00Z",
    "requirements": "openai>=0.27.4\nlangchain>=0.0.229",
    "avatar": ""
  }
}

Limitations

AWS Lambda only allows a function to run for a maximum of 15 minutes and 10GB of RAM. Therefore, the tools should finish their execution within this constraint.

Last updated 2 months ago