db11fe391d
Previous to this change our results were ungrounded and simpler models like GPT 3.5 and claude had a real tough time figuring out how to run commands This splits responses into 2 phases: Phase 1: figure out if you need to run a command Phase 2: respond to user This seems to produce better results on both Claude and GPT 3.5 but still needs a fair bit of tuning. |
||
|---|---|---|
| .. | ||
| database | ||
| inference | ||
| tokenizer | ||
| chat_message_classificator.rb | ||
| classificator.rb | ||
| post_classificator.rb | ||