Assist: Get Source when Running a Script with an LLM?

Configuration Voice Assistant
, ,
1

I’ve run into what I think is a limitation with scripts and the Assist pipeline.

The whole goal I’m trying to accomplish is to write a HA script that a voice assistant can call, which will use Music Assistant to search for a track and play it on the device the request originated from (assuming the assist device has a companion media_player on the same device).

The core problem I’m having: I don’t see any way to identify the “source” of the script being run.

So if I have a script: script.my_script And this is invoked from a device like Voice PE, I get the following context:

this:
  entity_id: script.my_script
  state: 'off'
  attributes:
    last_triggered: '2025-03-10T13:56:10.221730+00:00'
    mode: single
    current: 0
    icon: mdi:bullhorn-variant
    friendly_name: My Script
  last_changed: '2025-03-10T13:56:17.617790+00:00'
  last_reported: '2025-03-10T13:56:17.617790+00:00'
  last_updated: '2025-03-10T13:56:17.617790+00:00'
  context:
    id: 01JP05F4JHHRABG3H5RXFHYGS2
    parent_id: null
    user_id: null
some_script_var: 'some value'
context:
  id: 01JP06VNDDE9S0A0ZRPWRNXE0H
  parent_id: null
  user_id: null

From this, how can I determine what the source is? For example, I want to know that this script was invoked by assist_satellite.office and not assist_satellite.kitchen.

Is this possible? If so, how?

Edit: Adjusted title for clarity.

2

What are the reasons to use a script if you are going to activate the process by voice? Сreate automation.

{{'media_player.'~device_attr(trigger.device_id, "name")~'_something_else'}}
3

Because it may not always be the same sentence trigger.

What Jake is doing allows it to be used more like a tool.

4

If going the hard way, I would try to make an auxiliary sensor that stored the value of the device that was last queried from. :upside_down_face:

5

Hm, that could work - assuming that nobody queries their Voice PE device to play music at (roughly) the same time. I didn’t think of this, I’ll give it a try and report back if it works!

Sorry, I wasn’t specific. I’m using an LLM to call the script so I can use natural language. The Custom Sentence trigger on intents doesn’t fit my needs.

6

It works! For those wondering, here’s what you have to do:

Create an input_text helper.

Then add the following automation:

alias: Set Last Used Voice Device
description: ""
conditions: []
mode: single
triggers:
  - trigger: state
    to: listening
    entity_id:
      # Replace with the list of your voice assistants.
      - assist_satellite.one
      - assist_satellite.two
      - assist_satellite.three
actions:
  - action: input_text.set_value
    target:
      # Replace with your helper entity ID.
      entity_id: input_text.last_queried_voice_entity_id
    data:
      # This template assumes you have an assist satellite and a 
      # media player with the same device name.
      # Tweak to map your entities accordingly.
      value: >-
        {{ trigger.entity_id | replace('assist_satellite.', 'media_player.') |
        replace('_assist_satellite', '_media_player') }}

To use in an action:

# In a service call / action:
target:
  entity_id: "{{ states('input_text.your_last_used_assist_id_helper') }}"
1 Like
7

Do you have any idea for getting who the last request came from, could that be available from some entity - if the sentence came from either a mobile app or from the web interface?

8

User currently isn’t in the stream. (who is the user from a given voice call on a VPE? who knows without voice identification) you can Make guesses based on states but that’s about it.

9

The mobile companion apps have their own respective users, so knowing that would be pretty handy. Web interface is also tied to user login. I have no idea how to deduct from any states from who the latest command came from. :confused: (So its not just voice id.)

10

If you launch a pipeline from script or the UI you can absolutely capture

 {{user}}

(i think it’s user maybe user_id)
Thats pretty easy actually. When you fire it from voice that will be blank and you’d have to fall back on which voice station ID

11

Thanks for the tip. I have this custom sentence:

intents:
  custom_intents:
    data:
      - sentences:
          - "{custom_text}"
lists:
  custom_text:
    wildcard: true

And this script to handle it (and pass it on to nodered):

intent_script:
  custom_intents:
    action:
      - action: mqtt.publish
        data:
          topic: /nodered/assist
          payload: "{{user}} {{user_id}}: {{custom_text}}"
      ...
    speech:
      text: ...

Unfortunately, none of the user or user_id get replaced by anything.

12

@NathanCu did say:

So that is expected. It’s not possible to know which user initiated a voice action. There is no “voice recognition” in Home Assistant currently.

13

In case anyone else comes across this thread, I solved a similar problem by using a variable in the esp firmware (similar scope as name or friendly name).

substitutions:
name: office-echo
default_speaker: media_player.office

Then using shared firmware, I send that variable to the scripts that I call. I use this to play music with the voice assist, but primarily to play a “ding” to acknowledge when micro_wake_word detects. My esp devices are hidden, so there is no feedback from the leds.