Avoid hallucinations and information from trainning data

Hi @EduardoDominguez,

You are correct that it is often a good idea to include a fallback phrase in your Instructions, such as "If the required information is not available, you must respond only with 'There is no data to score'".

These kinds of imperatives can give the model a phrase to latch onto when it's uncertain, and can help to reduce hallucinations.

I think this is what you were asking for clarification on, but if I misunderstood your question, let us know.

Best,

-J

Thanks a lot for your super-fast advice! My results have increased quality.

I only need a bit more of information about when is a best practique incorporate a description for a Guided attribute.

@Generable struct SearchSuggestions { @Guide(description: "A list of suggested search terms", .count(4)) var searchTerms: [String] }

Listening the WWDC25 session "Deep dive into Foundation Generation Models" from min 9:00 to 9:15:

Multiline

Foundation Models will even automatically include details about your Generable type in the prompt, in a specific format that the model has been trained on. You don’t have to tell it about what fields your Generable type has

BlockQuote

I interpret that with an auto-descriptive name to the attribute like "SearchSuggestions" write a description is overwork because the model already may infer what must be the content. Am I right?

If my state is true, go deeper and learn what must be set in the description. Some examples would be the best to set clearly best practiques 😀

@EduardoDominguez If I understand correctly, you're noting that Foundation Models is able to see the structure of your generable type, and you are asking if this means that guides are redundant.

Our general advice is this: First, create your @Generable type with just descriptive property names, and see what kinds of values are created. Then, in areas where the model isn't consistently generating relevant data, try adding @Guide to provide additional guidance to the model for generating those values.

With on-device models, reducing complexity and keeping the context window small with succinct instructions is generally preferably, but through experimentation you may find that property-specific guides lead to more consistent results. Remember, building with large language models is all about testing and iterating!

Best,

-J

Thanks -J for your advices.