kercre123 / wire-pod

Free, fully-featured server software for the Anki Vector robot.
MIT License
435 stars 148 forks source link

[BUG] Many words can't be used on a sentence when trying to talk to the LLM #363

Closed MicaelJarniac closed 2 months ago

MicaelJarniac commented 2 months ago

Describe the bug Trying to ask the LLM for information about a game seems to always get Vector to start playing Blackjack.

Many other words have similar behavior, like "bad" and "start".

https://github.com/kercre123/wire-pod/blob/daf374cc9febd6c2ea8c70146eed58035a85aef2/chipper/intent-data/en-US.json

To Reproduce Steps to reproduce the behavior:

  1. Say "Hey Vector, what can you tell me about the Minecraft game?"
  2. Vector starts Blackjack

Expected behavior The prompt shouldn't match any known intents, thus forwarding it to the LLM.

Screenshots

Device you are running wire-pod on (please complete the following information):

Additional context

2024.07.08 20:56:55: Intent matched: intent_play_blackjack, transcribed text: 'what can you talk about the minecraft game', device: 00706f36
2024.07.08 20:57:07: Intent matched: intent_play_blackjack, transcribed text: 'what can you tell me about the mind curve game', device: 00706f36
MicaelJarniac commented 2 months ago

A possible solution would be to have something similar to "I have a question", but for talking directly to the LLM, thus skipping intent matching.

Zomboy4313 commented 2 months ago

WirePod is intent matching since you said 'game', saying I have a question is the method to always send responses to the LLM. This is not a bug

MicaelJarniac commented 2 months ago

Solved by #368.