Auto-tagging and semantic search are heuristic and lexical without a Groq or Cohere key (see keepstack/ai.py). A small in-process model would make them genuinely learned with no cloud dependency.
What to do: add an optional local model path (for example a small CLIP) for tagging and embeddings, behind the same pluggable interface. Keep the current deterministic fallbacks as the zero-dependency default.
Acceptance: with the local model enabled, semantic search ranks by visual and semantic similarity rather than lexical overlap, and no external API is called.
Auto-tagging and semantic search are heuristic and lexical without a Groq or Cohere key (see
keepstack/ai.py). A small in-process model would make them genuinely learned with no cloud dependency.What to do: add an optional local model path (for example a small CLIP) for tagging and embeddings, behind the same pluggable interface. Keep the current deterministic fallbacks as the zero-dependency default.
Acceptance: with the local model enabled, semantic search ranks by visual and semantic similarity rather than lexical overlap, and no external API is called.