Trouble Converting Custom GPT to Assistant

API
1

I have tried to reproduce the “actions” I created while building my GPT in an “Assistant” and it was impossible for me to make it access to a URL to read its data (the url is a data api that updates constantly their data, so I can’t just download it and upload it as a file). Does anyone know if it can be done and how?

2

Have you copied the instructions as well to match with the custom GPT ?

1 Like
3
  1. Basically, You need to convert your action to function/tool.
  2. Then add reference to the function/tool in your Instructions.
  3. When the function is invoked in the run, get the parameters from the tools and use it to fetch your external API endpoint.
  4. Submit the response back to the run.

If you have your Action schema, you can use a function like below to automatically convert it to function definition

function convertSchema(inputSchema) {
  var outputSchema = {}

  for (var path in inputSchema) {
    var methods = inputSchema[path]
    for (var method in methods) {
      var operation = methods[method]
      outputSchema.name = operation.operationId
      outputSchema.description = operation.summary
      outputSchema.parameters = {
        type: "object",
        properties: {}
      }
      operation.parameters.forEach(function(param) {
        outputSchema.parameters.properties[param.name] = {
          type: param.schema.type,
          description: param.description
        }
      })
    }
  }
  
  return outputSchema

}

For example, you have this schema:

...
"get": {
        "summary": "Get current weather information",
        "operationId": "checkWeatherUsingGET",
        "parameters": [
          {
            "name": "location",
            "in": "query",
            "required": true,
            "description": "Location for which to retrieve weather information.",
            "schema": {
              "type": "string"
            }
          }
        ],
...

It will output like this:

{
      "name": "checkWeatherUsingGET",
      "description": "Get current weather information",
      "parameters": {
        "type": "object",
        "properties": {
          "location": {
            "type": "string",
            "description": "Location for which to retrieve weather information."
          }
        },
        "required": ["location"]
      }
    }

Now, when this is invoked in the run, you’ll get something like this:

{
    id: 'call_3Bf62kh5wpgzE8wjn0TvVxeA',
    type: 'function',
    function: {
      name: 'checkWeatherUsingGET',
      arguments: '{"location":"Tokyo"}'
    }
  }

Using the location parameter, call your endpoint

https://.../api/weather?location=Tokyo

You will receive something like this

{"location":{"name":"Tokyo","region":"Tokyo","country":"Japan","lat":35.69,"lon":139.69,"tz_id":"Asia/Tokyo","localtime_epoch":1695556336,"localtime":"2023-09-24 20:52"},"current":{"last_updated_epoch":1695555900,"last_updated":"2023-09-24 20:45","temp_c":23,"temp_f":73.4,"is_day":0,"condition":{"text":"Clear","icon":"//cdn.weatherapi.com/weather/64x64/night/113.png","code":1000},"wind_mph":20.6,"wind_kph":33.1,"wind_degree":50,"wind_dir":"NE","pressure_mb":1020,"pressure_in":30.12,"precip_mm":0,"precip_in":0,"humidity":73,"cloud":0,"feelslike_c":24.9,"feelslike_f":76.9,"vis_km":10,"vis_miles":6,"uv":1,"gust_mph":17.8,"gust_kph":28.6},"infoLink":"https://weathergpt.vercel.app/tokyo"}

Submit that as your tool output and continue the run

{
    tool_call_id: 'call_3Bf62kh5wpgzE8wjn0TvVxeA',
    output: JSON.stringify(api_response)
  }

That should work.

4 Likes
4

quick follow up question @supershaneski - I assume there is no way to use something like your function (convertSchema) dynamically/programmatically, rather you were suggesting to generate the functions offline and hard code the function definitions, correct?

We are also struggling with 2 major issues around actions, custom GPT does not allow more than 30 operations and we have a very large API spec (300+), and converting that to functions seems very kuldgy and every time an API changes we’ll have to regenerate the schema/functions all over and redeploy a new assistant API - If anyone has other suggestions around this topic we’d be grateful.

1 Like
5

You can probably deploy or include a similar function in your workflow. Every time you add/update an action schema that you update the associated JSON file/string of the function/tool definition by calling such convertSchema function. Then you can perhaps also include some dynamic testing to check if it works.

1 Like