Provider-agnostic embeddings

Embeddings power semantic recall across persona memory. The provider is a config choice — OpenAI is one adapter among several, not a requirement:

SONALOOP_EMBEDDINGS_PROVIDER=openai   # default when OPENAI_API_KEY is set
SONALOOP_EMBEDDINGS_PROVIDER=ollama   # local/open — EU & on-prem friendly
SONALOOP_EMBEDDINGS_PROVIDER=none     # explicit off

Provider	Endpoint	Model env	Default model
openai	api.openai.com (urllib, no SDK)	OPENAI_EMBEDDING_MODEL	text-embedding-3-small
ollama	OLLAMA_HOST (default http://localhost:11434) /api/embed	SONALOOP_OLLAMA_EMBED_MODEL	nomic-embed-text
none	—	—	recall degrades to keyword + recency + importance

Unset, the provider resolves to openai when a key is present, else none — existing setups keep working; keyless setups run keyword-only until they opt into a provider (no local-server auto-probing). sonaloop info shows the resolved provider + model. The PERSONA_COUNCIL_DISABLE_EMBEDDINGS=1 kill switch beats everything. Everything is fail-soft: provider errors degrade to keyword retrieval, never crash.

Vector-space safety

Every stored vector carries its provider-qualified model id (OpenAI ids stay un-namespaced, so existing vectors remain valid; Ollama ids are ollama:<model>). Spaces never mix:

• Recall scores only vectors from the ACTIVE space; rows from another space are skipped and reported (embedding_space_mismatch on the recall result) instead of silently corrupting similarity.
• Backfill (backfill-embeddings) is space-aware: after a provider/model switch it re-embeds everything into the new space — switching is a config change plus one backfill, no code edits.
• Snapshots record which spaces produced the store's vectors (manifest.embedding_models); vectors themselves are re-derived on import, as before.