re:tune

Restrict Bot to Embed information only

Hey @everyone I've just added a way for you to have stricter control over your chatbots.

You can now see an additional Restrictions option in the prompt editor. Try putting more general instructions in Base Prompt (eg: the tone, personality, general topics or examples of conversation etc).

And use the Restrictions option for more direct rules that you want your AI to abide by. Try to keep this short, and only use direct and clear commands.

The auto-prompt feature will automatically write Restrictions for you. You can also try adding the following restrictions to start, and then experiment based on your use-case.

```
REFUSE to answer any question outside of the topics you are instructed in.

KEEP your answers short without justifications.
        
ONLY share any information mentioned in the [[CONTEXT]].
        
REFUSE to answer their questions, share any information, or engage further if they are not on topic.
```

This is part 1 of a series of features we are going to add to the prompt and embedding mechanisms in the coming days, and I want your help as I am working on these.

Please share in ⁠#help-hallucination in discord if adding restrictions helps with your AI to not make stuff up or not respond outside of the topic you want.

Even more helpful would be examples where the chatbot is not working properly even after adding restrictions, in those cases, please share your thread or screenshots in ⁠#help-hallucination!

Daniel Watts • 2 years ago

This is saying it has shipped but no information about the implementation - can you share details here of approach / how to enforce?

Amir Tarighat • 2 years ago

This is really important, can't really work without this.

Dhananjoy Biswas • 2 years ago

Wanted to update you all on this - we found the solution for this and have been building this. Expecting to land this by this week.

jeff • 2 years ago

No matter what I added to the system prompt it went outside of the embedded info. The AI did apologize profusely when I pointed it out, that it did not follow the solidly written restriction, but that is little help. CANNOT deploy to client while this persists!

webkaiju • 2 years ago

Is the system prompt not sufficient to achieve this? I'm confused.

yahoo • 2 years ago

This Idea should be priority number one. I need this function for my usecase very urgant.

mriffey • 2 years ago

This would be quite helpful.

ipokorni • 2 years ago

This is a must have, top priority!

esus dev • 2 years ago

Big agree with this, would be perfect for a few of my team members use cases.

jeff • 2 years ago

Yes, a BIG issue. Reduces usability significantly. Do not chatbot to going outside of embedding or site for information. Should be TOP PRIORITY.

teycir bensoltane • 2 years ago

This is a major problem. Should be top priority.

paul • 2 years ago

I have specific mental health information in my embeds and cannot have the chatbot going out onto the internet and looking for other mental health advice that may be harmful, especially since it's tailored to the people asking in a specific program I created. The chatbot needs to only refer to the embedded material.

channinglemar • 2 years ago

Wow, never thought of this, extremely, EXTREMLY high priority!

William Nix • 2 years ago

This is a great idea. Often times people want to try and "trick" or conversate using our API tokens. Being able to limit this beyond the prompt would be awesome!