feat: optional MLX embedder backend (torch-free, Apple-silicon native)

[ry-ops]

Summary

Adds an optional MLX embedding backend so aiana can embed on Apple silicon without PyTorch, selectable via AIANA_EMBEDDER_BACKEND=mlx. Sentence-transformers stays the default — no behavior change unless opted in.

Motivation: running the sentence-transformers/torch embedder alongside a local LLM is heavy (it was the dependency that OOM'd a 16 GB M1 in practice). An MLX embedder shares the same unified-memory runtime as MLX-served models, so the embedding layer becomes lighter and fully Apple-silicon-native.

Changes

• embeddings/mlx_embedder.py — new MLXEmbedder mirroring the Embedder interface (embed, dimension, similarity), running on mlx-embeddings. Optional import with a clear install hint; defaults to all-MiniLM-L6-v2 (384-dim) — the same model family and vector space as the sentence-transformers default, so existing aiana_memories vectors stay searchable (no migration).
• embeddings/embedder.py — get_embedder() selects the backend from AIANA_EMBEDDER_BACKEND (sentence-transformers default, or mlx).
• pyproject.toml — new optional extra mlx (mlx-embeddings), added to all.

Usage

pip install "aiana[mlx]"
export AIANA_EMBEDDER_BACKEND=mlx

Verification

On Python 3.13 / Apple silicon:

AIANA_EMBEDDER_BACKEND=mlx → get_embedder() -> MLXEmbedder
model: mlx-community/all-MiniLM-L6-v2-bf16
dimension: 384 | embedding L2-norm: 1.0
semantic separation (related > unrelated): True

Notes

• Default backend and behavior are unchanged; this is purely additive/opt-in.
• Switching an existing deployment from one model family to another implies re-embedding; defaulting the MLX backend to all-MiniLM-L6-v2 avoids that for the common case.
• Independent of fix: make aiana installable and compatible with qdrant-client >= 1.12 #17 (the install/qdrant-client-compat fixes); this branch is cut from main and stacks cleanly with it.

🤖 Generated with Claude Code