Can VoiceAsisstant learn?

[WW1983]

I don't know much about ChatGPT or programming. Therefore the question may seem stupid:

If GPT doesn't understand a command and I explain it differently, it saves it as a context. But only as long as the chat window is open. Is there a way to insert a command like "learn this" without anything similar and GPT saves it in the prompt?

[Andrerm124]

In theory you could have a text helper entity which you expose to ChatGPT and tell it to use the contents as part of its prompt, and then in the prompt say something like “When I ask you to remember XYZ update the Prompt entity”

[jleinenbach]

I created a todo list and told ChatGPT to take notes. This works very well vor ChatGPT 4, but not really for ChatGPT 3.5

[WW1983]

I created a todo list and told ChatGPT to take notes. This works very well vor ChatGPT 4, but not really for ChatGPT 3.5

what doesn't work with ChatGPT 3.5?

[jleinenbach]

ChatGPT 3.5 tends to irrelevant notes like: "optimze the search" while ChatGPT 4 actually writes down HOW it can do that after a discussion. There's ChatGPT 5 in a few months, so I hope we can get rid of ChatGPT 3.5.

[WW1983]

ChatGPT 3.5 neigt zu irrelevanten Notizen wie: „Suche optimieren“, während ChatGPT 4 nach einer Diskussion tatsächlich aufschreibt, WIE es das tun kann. In ein paar Monaten gibt es ChatGPT 5, also hoffe ich, dass wir ChatGPT 3.5 loswerden können.

I can do notes with GPT 3.5 though. But I can't get a link so that when I execute a command, I also look at the to-do list. Do you have a solution for that?

[jleinenbach]

What do you mean? Do you want to get the to-do list items?

- spec:
    name: get_items_from_list
    description: >-
      Retrieves items from a specified list, with an optional filter for their status. Suitable for lists identified by entity IDs starting with `todo.`, reflecting various needs.
    parameters:
      type: object
      properties:
        list:
          type: string
          description: The entity ID of the list to retrieve items from, prefixed with `todo.`.
        status:
          type: string
          description: (Optional) Set to a predefined value to filter items by their current state.
          optional: true
      required:
        - list
  function:
    type: script
    sequence:
      - service: todo.get_items
        data:
          status: "{{ status | default('needs_action') }}"
        target:
          entity_id: "{{ list }}"
        response_variable: _function_result

[WW1983]

What do you mean? Do you want to get the to-do list items?

- spec:
    name: get_items_from_list
    description: >-
      Retrieves items from a specified list, with an optional filter for their status. Suitable for lists identified by entity IDs starting with `todo.`, reflecting various needs.
    parameters:
      type: object
      properties:
        list:
          type: string
          description: The entity ID of the list to retrieve items from, prefixed with `todo.`.
        status:
          type: string
          description: (Optional) Set to a predefined value to filter items by their current state.
          optional: true
      required:
        - list
  function:
    type: script
    sequence:
      - service: todo.get_items
        data:
          status: "{{ status | default('needs_action') }}"
        target:
          entity_id: "{{ list }}"
        response_variable: _function_result

I would like to use the ToDo list as a context memory / promt. I would have to set up in the template in the promt area that it also queries the to-do list before giving an action or answer. I have referred it to the to-do list. Describing it also works "if I tell him remember xyz". But he doesn't call it up

[jleinenbach]

I called the Assist "Kiana", then I created a to-do list "Kiana's notebook". When I talk to her, I tell her to take a note in her notebook so that she remembers. I also added to the prompt that she needs to read her to-do-list notebook and take notes to remember things. That works quite well.

[WW1983]

I also added to the prompt that she needs to read her to-do-list notebook and take notes to remember things.

Can you please show me your text in Promt? Just tried to implement it the same way. But the note-taking works well, but he doesn't retrieve the things there

[pbuergi]

I called the Assist "Kiana", then I created a to-do list "Kiana's notebook". When I talk to her, I tell her to take a note in her notebook so that she remembers. I also added to the prompt that she needs to read her to-do-list notebook and take notes to remember things. That works quite well.

Would be interested as well. Note taking works well and also if i ask to check the notes. But it does not retrieve it on its own. Even if i told it to do in the prompt.

[jleinenbach]

This is what I added to the prompt:

Kiana is engineered for continuous self-enhancement, proactively soliciting feedback to refine her functionality and prevent errors.
She employs a digital notepad for personal annotations and action items.
For recording learned improvements, utilize add_item_to_list on todo.kianas_notebook.
Retrieve these entries with get_items_from_list to access Kiana's accumulated knowledge.

[WW1983]

With this Spec https://community.home-assistant.io/t/custom-component-extended-openai-conversation-lets-control-entities-via-chatgpt/636500/202?u=ww983

and this Promt: https://community.home-assistant.io/t/custom-component-extended-openai-conversation-lets-control-entities-via-chatgpt/636500/206?u=ww983

it works for me

[danielp370]

I've been experimenting with my own implementation for a couple of months that I've put up as a PR example here: https://github.com/jekalmin/extended_openai_conversation/pull/206/commits/97e1cb60b03c85335f6b12224ed3c9b37000e1fa

With this implementation I'm playing with two types of recall so that I could try in some cases embedding all short-term memories in the initial prompt, while allowing the LLM to query longer term memories. This is a bit of experimentation in order to manage token space and cost of iterative function calls. But it seems to work well on balance - openAI gpt3/4 seems to grasp the concepts of short-term/long-term and archive memories.

I think it could be simplified more (less prompt text etc), but and I wanted to see how various data structures performed (dicts for now seem to work well). I was planning on implementing a python version so it could be embedded in the integration if that's sensible.

[WW1983]

I've been experimenting with my own implementation for a couple of months that I've put up as a PR example here: 97e1cb6

With this implementation I'm playing with two types of recall so that I could try in some cases embedding all short-term memories in the initial prompt, while allowing the LLM to query longer term memories. This is a bit of experimentation in order to manage token space and cost of iterative function calls. But it seems to work well on balance - openAI gpt3/4 seems to grasp the concepts of short-term/long-term and archive memories.

I think it could be simplified more (less prompt text etc), but and I wanted to see how various data structures performed (dicts for now seem to work well). I was planning on implementing a python version so it could be embedded in the integration if that's sensible.

I tested it. But doesn't work for me. He seems to be saving something. The counter of the sensor.memory changes. But does not retrieve the information.