Embedding resolution: local-default + light + cache

Refs #1.

ir.embed.resolve_embedder(spec):

'default'/'local'/'minilm' -> ef.embedder_adapters.sentence_transformers_embedder('all-MiniLM-L6-v2'), set USE_TF=0 before import (avoids TF/numpy ABI crash), wrapped in ef.CachedEmbedder over a dol cache at ~/.cache/ir/embeddings/<model>/ keyed by (model, ef.content_hash).
'light'/'hashing' -> ef.HashingEmbedder (numpy-only) for fast tests.
callable / Embedder -> ef.as_embedder.
Graceful fallback to hashing with a warning if sentence-transformers unavailable.

Embedder must carry an input_type distinction (query vs document) — ef already supports this.

Decisions (done in PR #14):

Default = all-MiniLM-L6-v2 (384-d) via ef.embedder_adapters.sentence_transformers_embedder(..., normalize=True).
USE_TF=0 is set on import ir.embed — without it, transformers auto-imports TensorFlow, which crashes on this stack's numpy ABI (numpy.core.umath failed to import). Import ir before anything that imports transformers.
Wrapped in ef.CachedEmbedder over a dol numpy store at ~/.cache/ir/embeddings/<model_id> — each model gets its own dir, so the cache key is effectively (model, content_hash) with no cross-model collisions.
"light"/"hashing" → ef.HashingEmbedder (numpy only) for fast tests; graceful fallback to it (with a warning) if sentence-transformers is missing.
First call loads the model (~18s) then is warm; the disk cache makes re-indexing near-instant. A persistent embedding server is a later optimization if cold CLI latency matters.

Navigation Menu