[Terr_]

IMO the best fusion for this kind of thing is:

1. There's a "normal" interface or query-language for searching.

2. The LLM suggests a query, based on what you said you wanted in English, possibly in conjunction with results of a prior submit.

3. The true query is not hidden from the user, but is made available so that humans can notice errors, fix deficiencies, and naturally--if they use it enough--learn how it works so that the LLM is no longer required.

[8n4vidtmkvmk]

Yessss! This is what I want. If there is a natural set of filters that can be applied, let me speak it in natural language, then the LLM can translate that as good as possible and then I can review it. E.g. searching photos between X and Y date, containing human Z, at location W. These are all filters that can be presented as separate UI elements so I can confirm the LLM interpreted correctly and I can adjust the dates or what have you without having to repeat the whole sentence again.

[Terr_]

Also, any additional LLM magic would be a separate layer with its own context, safely abstracted beneath the filter/search language. Not a post-processing step by some kind of LLM-shell.

For example, "Find me all pictures since Tuesday with pets" might become:

    type:picture after:2025-10-08 fuzzy-content:"with pets"

Then the implementation of "fuzzy-content" would generate a text-description of the photo and some other LLM-thingy does the hidden document-building like:

   Description: "black dog catching a frisbee"
   Does that "with pets"? 
   Answer Yes or No.
   Yes.