EmbeddingGemma 2 vs Mac Spotlight: 30/30 vs 5/30, offline on my Mac

Description and use

EmbeddingGemma 2 on Mac delivers 100 % file‑search accuracy and RAG answers, outperforming Spotlight, and supports face‑search with 94 % accuracy offline.

Problem

Need an offline, fast, accurate file‑search and question‑answering system on a Mac that outperforms macOS Spotlight.

Outcome

EmbeddingGemma 2 achieved 100 % top‑1 file retrieval, 100 % correct answers, and 94 % face‑search accuracy offline, with no cloud cost and ~34 ms per query.

Workflow

• Index 62 files into 768‑dim vectors with EmbeddingGemma 2
• Store vectors in memory (numpy) or lightweight DB
• For each question: embed query, retrieve top 3 files
• Run Gemma 4 E4B on retrieved files to answer
• For face‑search: detect faces with YuNet, embed with ArcFace R50, match against user‑labeled faces
• Filter by name, rank by EmbeddingGemma 2 relevance
• Move junk photos to _Sweep/ folder

Evidence

EmbeddingGemma 2 found the right file 30 times out of 30 on my Mac in October 2026, offline, in about 34 ms per question.

Source

Gemma 4 E4B answered from the files it found: 11/11, about 2 s each, nothing left the Mac.

Source

Find anyone by name (ArcFace, 5 clicks on 1 photo) 114/121 family appearances, 0 wrong names; 118/121 after 2 more clicks.

Source