Using Chat GPT json response in an automation

Configuration
, , ,
1

After trying to use chat gpt to help me build an automation, I’ve come to the end of my rope and back to the real experts.

I have a well running automation which announces events from my calendar 45 min ahead of their start time. It has random phrases and uses amazon polly. But, its become a little repetitive, even with the randomness, and I wanted to see if chat gpt could help.
Turns out I can send a request to the OpenAI API and get a response back, but I’m struggling to parse the json and feed the text into my TTS script.

So far I have this Rest Command:

rest_command:
  generate_calendar_prompt:
    url: "https://api.openai.com/v1/chat/completions"
    method: POST
    headers:
      Authorization: "Bearer !secret openai_api_key"
      Content-Type: "application/json"
    payload: >
      {
        "model": "gpt-3.5-turbo",
        "messages": [
          {
            "role": "system",
            "content": "You are a friendly assistant providing natural responses to calendar events."
          },
          {
            "role": "user",
            "content": "Create a varied, natural announcement for an event called '{{ event_title }}', starting at '{{ event_start }}'. Make it fun and a little snarky but do not include any emojis. Keep it less than 200 characters"
          }
        ],
        "temperature": 0.7
        }
    content_type: "application/json; charset=utf-8"

When I try this in the developer tools I get a good response (also tried it with curl in terminal and works well)

content:
  id: chatcmpl-AeXHZSe5oE8RYXkOiWbcVu5Ppc88g
  object: chat.completion
  created: 1734224421
  model: gpt-3.5-turbo-0125
  choices:
    - index: 0
      message:
        role: assistant
        content: >-
          Alright everyone, it's time for Kids Bedtime! Get your little ones
          ready to wind down and enjoy a cozy evening at 9:52 PM on December 6,
          2024. Let's make bedtime fun!
        refusal: null
      logprobs: null
      finish_reason: stop
  usage:
    prompt_tokens: 85
    completion_tokens: 45
    total_tokens: 130
    prompt_tokens_details:
      cached_tokens: 0
      audio_tokens: 0
    completion_tokens_details:
      reasoning_tokens: 0
      audio_tokens: 0
      accepted_prediction_tokens: 0
      rejected_prediction_tokens: 0
  system_fingerprint: null
status: 200

If I plug the json into the template demo, I am able to extract the text properly with:

{{ value_json["choices"][0].message.content }}

BUT, when I try to bring it all together in the final automation, it’s just not working. I’m stuck.

- id: "1733272058982"
  alias: Family Reminders ChatGPT
  description: remind family of events using Chat GPT
  triggers:
    - trigger: calendar
      event: start
      entity_id: calendar.family
      offset: -0:45:0
  conditions:
    - condition: time
      before: "23:00:00"
      after: 07:30:00
      weekday:
        - sun
        - mon
        - tue
        - wed
        - thu
        - fri
        - sat
    - condition: state
      entity_id: input_boolean.vacation_mode
      state: "off"
  actions:
    - action: rest_command.generate_calendar_prompt
      metadata: {}
      data:
        event_title: "{{ trigger.calendar_event.summary }}"
        event_start: "{{ trigger.calendar_event.start }}"
      response_variable: gpt_response
    - delay:
        hours: 0
        minutes: 0
        seconds: 5
        milliseconds: 0
    - alias: Parse GPT Response
      variables:
        content: " {{ gpt_response['choices'][0].message.content }} "
    - action: script.voice_announcement_gpt
      metadata: {}
      data:
        message: "{{ content }}"
  mode: queued

Thanks all for the help in advance!

2

What does the trace show for the value of gpt_response in the “Changed Variables” section?

3

I get an error.
Error: UndefinedError: ‘dict object’ has no attribute ‘choices’

BUT, in the trace for the rest command itself (within the automation), I do get the response as follows:

context:
  id: 01JF5DAPC3884DKJJEKHY7A0X2
  parent_id: null
  user_id: null
gpt_response:
  content:
    id: chatcmpl-AekO8rUPwOnLvXmIeeFdZqCpHbqVU
    object: chat.completion
    created: 1734274800
    model: gpt-3.5-turbo-0125
    choices:
      - index: 0
        message:
          role: assistant
          content: >-
            Get ready to celebrate Toad's Birthday on December 15, 2024, at
            10:45 AM EST! Join us for a fun-filled event to make Toad's special
            day unforgettable.
          refusal: null
        logprobs: null
        finish_reason: stop
    usage:
      prompt_tokens: 86
      completion_tokens: 41
      total_tokens: 127
      prompt_tokens_details:
        cached_tokens: 0
        audio_tokens: 0
      completion_tokens_details:
        reasoning_tokens: 0
        audio_tokens: 0
        accepted_prediction_tokens: 0
        rejected_prediction_tokens: 0
    system_fingerprint: null
  status: 200
4

Hello Cowboy-Ry,

I’m not good at this JSON stuff, but I’m thinking this maybe?

5

OK I figured it out with some additional help from ChatGPT. Here is the final automation!

- id: "1733272058982"
  alias: Family Reminders ChatGPT
  description: remind family of events using Chat GPT
  triggers:
    - trigger: calendar
      event: start
      entity_id: calendar.family
      offset: -0:45:0
  conditions:
    - condition: time
      before: "23:00:00"
      after: 07:30:00
      weekday:
        - sun
        - mon
        - tue
        - wed
        - thu
        - fri
        - sat
    - condition: state
      entity_id: input_boolean.vacation_mode
      state: "off"
  actions:
    - action: rest_command.generate_calendar_prompt
      metadata: {}
      data:
        event_title: "{{ trigger.calendar_event.summary }}"
      response_variable: gpt_response
    - delay:
        hours: 0
        minutes: 0
        seconds: 3
        milliseconds: 0
    - alias: Parse GPT Response
      variables:
        announcement: >
          {% if gpt_response and gpt_response['content'] and 'choices' in gpt_response['content'] %}
            {{ gpt_response['content']['choices'][0]['message']['content'] }}
          {% else %}
            "Error: Invalid or empty response from GPT."
          {% endif %}
    - action: script.voice_announcement_gpt
      metadata: {}
      data:
        message: "{{ announcement }}"
  mode: queued