feat(embedding): pluggable EmbeddingProvider layer (OpenAI + Ollama)

[nbzy1995]

Summary

Promotes the env-var shim in src/core/embedding.ts to a proper EmbeddingProvider interface so embedding backends slot in without touching call sites or the schema. Ships OpenAIProvider (default, identical behavior to before) and OllamaProvider (any model on a local Ollama daemon, e.g. nomic-embed-text @ 768d). Templates the vector(N) schema dim by the resolved provider so gbrain init --provider=ollama --model=nomic-embed-text actually creates a vector(768) brain.

Why

Today embedding.ts reads EMBEDDING_MODEL / EMBEDDING_DIMENSIONS / EMBEDDING_BASE_URL env vars and gates the Matryoshka dimensions param via model.startsWith('text-embedding-3'). The schema, however, hardcodes vector(1536) and 'text-embedding-3-large', so setting those env vars produces a brain whose vector column doesn't match what the embedder writes — a silent footgun.

This patch finishes the abstraction: provider quirks (Matryoshka, error normalization, dim inference) live in concrete provider classes; schema dim flows from the resolved provider at gbrain init time and is persisted to ~/.gbrain/config.json. After init, connectEngine() hydrates the persisted provider before any command runs, so embed, query, import, serve all use the brain's frozen choice — env-var changes can't silently corrupt vectors mid-life.

Scope (4 commits, ~870 LOC, all behavior-preserving by default)

#	Commit	What
1	8b206b8	Extract EmbeddingProvider interface, OpenAIProvider, OllamaProvider, factory, service. embedding.ts becomes a backward-compat shim. 15 new provider tests.
2	0abd57c	Template pglite-schema.ts and schema-embedded.ts by (dimensions, defaultModel). engine.initSchema(opts?) plumbs them through. init.ts parses --provider/--model/--dimensions/--base-url, persists embedding: {...} to config, refuses re-init on dim mismatch. cli.ts --version prints the active provider. 11 new schema-templating tests.
3	7081e12	parseEmbeddingFlags accepts both --flag value and --flag=value.
4	41469e1	cli.ts connectEngine() hydrates the provider from config.embedding before any command — so embed/query don't fall back to OpenAI defaults when the brain was inited for Ollama.

Tests

• 883 pass / 0 fail (was 861) — added 26 new tests across test/embedding/provider.test.ts and test/schema-templating.test.ts
• All default behavior preserved: gbrain init with no flags → identical SQL as before, OpenAIProvider instantiated, all existing tests unchanged
• Ran end-to-end on a real personal corpus: 447 pages, 925 chunks, 742 embeddings via Ollama nomic-embed-text. Verified gbrain query returns 1.0 cosine scores on topic-aligned questions

Backward compatibility

• embedding.ts still exports embed and embedBatch with unchanged signatures (re-exports from the new embedding/service.ts). No call sites need editing.
• PGLITE_SCHEMA_SQL and SCHEMA_SQL const aliases preserved (evaluate the schema function with default opts → identical SQL).
• engine.initSchema() with no args defaults to (1536, 'text-embedding-3-large') — existing test harnesses keep working.
• EMBEDDING_* env vars still honored as the resolution fallback when no CLI flag and no persisted config.

Example usage

# Default (unchanged): OpenAI text-embedding-3-large at 1536d
gbrain init

# Local Ollama, free, offline, 768d
ollama pull nomic-embed-text
gbrain init --provider=ollama --model=nomic-embed-text --base-url=http://localhost:11434/v1

# Any OpenAI-compatible endpoint (vLLM, LiteLLM, etc.)
gbrain init --provider=openai --model=text-embedding-3-large --base-url=https://my-proxy.example/v1

Notes for maintainer

• Two follow-ups not in this PR (happy to send separately if you want):
  • gbrain config show renders nested objects as [object Object] — predates this PR but newly visible because embedding is the first nested config field
  • Some chunks failed to embed via Ollama with "input length exceeds the context length" — nomic-embed-text's native context is ~2K tokens, but the chunker can produce larger chunks. Worth gating chunker max size by provider.maxInputChars or the model's context window
• This PR is independent of feat(expansion): OpenAI-compat JSON mode replaces Anthropic tool use #165 (query expansion / Anthropic SDK swap) — different code path
• Happy to split into two PRs (provider-layer first, schema-templating second) if that's easier to review

🤖 Generated with Claude Code