Using OpenAI chat completion with function calls to SMHI api

The API is instructed to use the current time in Europe/Stockholm timezone. If the message can not be handled with the weather API endpoint, the Open AI assistant will reply instead.

• SMHI, forecast documentation
• OPEN AI, GPT function calling documentation

/ask?q=What is the weather like in Stockholm today

{ 
answer?: string // If the message could not be answered with the SMHI  API
error?:string // 
data?: unknown // the actual data returned from SMHI, if the API is called
summary?: string // a summary of the data, by GPT API
 }

• How is the weather in the Capital of Sweden tomorrow.
• How is the weather at Liseberg on Friday.

1. Handle the API request and count the tokens in the question.
2. Send the question to Open AI API moderation
3. Create tool calling by converting schema to JSON schema
4. Send the question to Open AI Chat Completion and expose tool calling
5. Make the API call to the SMHI API with parameters from OPEN AI
6. Add additional information to the SMHI response (converting numerical categories to text)
7. Remove data from the response that was not requested by the Open AI parameters.
8. Send the modified response to Open AI API together with the SMHI response JSON Schema.
9. Send a summary of the forcast (by Open AI API) together with the forecast data as a response to the initial request.

• OPENAI_CHAT: Needs to be authorized to write chat completions and to the moderation API.

• openai: For typesafe API request and responses
• valibot: for describing the SMHI API response and function API input
• valibot/to-json-schema: Transform the schema to json schema (readable by the GPT API)
• gpt-tokenizer: count the number of tokens
• formkit/temporal: To handle dates in a specific timezone (Europe/Stockholm)