How to make llama.cpp return control to add additional context?

[simplejackcoder]

I want to be able to tell the model that if it can't reply something useful to return control so I can give more information.

Similarly, how do I add more context so that it can reason about a full conversation or say a specific set of documents?

For example, I ask it something and it should say I don't know can you provide me more information? And then I give it a document. Then I can add another document to the prompt, so it can understand from that and so on.

I've heard this is some sort of chaining, but I don't understand.

[Carnildo]

It's not possible. Large-language models are incapable of identifying things they don't know. They work by predicting the most likely continuation of the input text based on the training data they've seen. This is a remarkably good approximation to actual thinking in many cases, but it also means an LLM is unable to say "I don't know".

(If the training data says that something isn't known, an LLM is able to give the correct response, but this is different from being able to identify gaps in the training data.)

[github-actions]

This issue was closed because it has been inactive for 14 days since being marked as stale.