Questions regarding the LLM search interface project

Hello,
I'm interested in the "Interface the database search engine to an AI
based LLM" project mentioned in the ideas list for GSOC 2026

A few questions before I draft the proposal:
- For running the model locally, I was planning to use llama.cpp. Is
that acceptable, or do you prefer integrating with the existing OpenCV
system that digiKam already uses for facial recognition and grouping?
- Is the main goal to map natural language phrases to params for the
existing search API, or is it to embed the photos for semantic
similarity search, or both?
- For vector storage if semantic search is in scope - would you prefer
extending the existing SQLite database with an extension like
sqlite-vec, or a separate store like qdrant?
- For machines that can't run a local model because of hardware
constraints, should the feature not be available, or should it be
strictly opt-in regardless of the hardware?
- digiKam already prompts users to download models for face
recognition at startup. Im assuming the LLM model download should
follow the same pattern?

Thanks,
Anamay Narkar
[email protected]
Gitlab: https://invent.kde.org/anamaynarkar