Local RAG for a personal second brain: embedder and hybrid retrieval picks in 2026?

Building a fully local RAG setup for personal notes (life-logging second brain, single user, privacy is the whole point so no hosted APIs for the data). Stack is SQLite + sqlite-vec + FTS5, Ollama for embeddings and generation, all on an Apple Silicon Mac.

Two questions where I'd love real 2026 experience:

1. Embedder pick. I was defaulting to nomic-embed-text out of habit, but recent chatter favors qwen3-embedding:0.6b or embeddinggemma at similar sizes. Anyone benchmarked these head-to-head for English personal-notes retrieval (not BEIR)? Does it actually matter at ~50k chunks, or am I bikeshedding?

2. Hybrid fusion. FTS5 (BM25) + vector cosine, per-query score normalization. FTS5 has no typo tolerance — anyone wired in trigram + spellfix and found it worth it? Any fusion gotchas at small corpus sizes (e.g. vector signal drowning out keyword on exact-match queries)?

Corpus is Markdown notes + JSON records, chunked ~512 tokens. Query load is one human. Not chasing SOTA, chasing "correct answers on my own data."

What would you change?