Why

The existing google recipe uses GOOGLE_GENERATIVE_AI_API_KEY (AI Studio public API key). This works for personal use but doesn't fit production GCP deployments that already have a service account / GCE metadata-based ADC chain.

This PR adds a parallel google-vertex recipe that authenticates through Vertex AI's standard ADC chain (via google-auth-library), supporting:

• GCE / GKE: automatic credentials via metadata service
• Local dev: gcloud auth application-default login
• Production: standard GOOGLE_APPLICATION_CREDENTIALS env

Use case: GCP-deployed gbrain instances (e.g. on a Compute Engine VM) can drop the AI Studio API key entirely and reuse the project's service account.

What

Five-file extension to the existing pluggable provider framework. No changes to embedding.ts (still a thin wrapper); all wiring goes through the documented "new native provider = recipe + factory in gateway.ts" pattern.

Dependency: @ai-sdk/google-vertex@^4.0.0 (matches @ai-sdk/google@^3.0.x peer family / ai@^6.0.x).

Total: 132 insertions, 1 deletion across 8 files (incl. package.json + bun.lock).

Bonus fix (last commit, fully independent — feel free to cherry-pick separately)

src/core/postgres-engine.ts insertChunks had a hardcoded 'text-embedding-3-large' fallback regardless of the active provider, so chunks recorded OpenAI metadata even when Vertex / Voyage / etc. did the embed. The fix mirrors the same await import('./ai/gateway.ts') pattern initSchema already uses (L153). Pre-existing bug, surfaced when running gbrain end-to-end on a non-OpenAI provider.

Usage

After this lands:

# config.json
{
  "engine": "postgres",
  "database_url": "postgresql://...",
  "embedding_model": "google-vertex:gemini-embedding-001"
}

# env (or systemd Environment=)
GBRAIN_VERTEX_PROJECT=your-gcp-project
GBRAIN_VERTEX_LOCATION=us-central1   # optional, defaults to us-central1

gbrain init
gbrain doctor   # → google-vertex:gemini-embedding-001 ✓ NNNNms, 1536 dims, DB aligned

Test plan

• ✅ bun run typecheck clean (no errors)
• ✅ gbrain doctor embedding_provider probe returns 1536d, dim-aligned with schema
• ✅ End-to-end put_page → search on Supabase Postgres, cross-machine consistent (write from Windows + read from Linux GCE, same RRF score 0.4489)
• ✅ Tested with @ai-sdk/google-vertex@4.0.122 (current 4.x stable)

Compatibility notes

dims.ts includes three providerOptions namespaces (vertex, google, googleVertex) because:

• @ai-sdk/google-vertex@4.x parses vertex first, then falls back to google
• @ai-sdk/google-vertex@5.x canary parses googleVertex first, then vertex

Passing all three keeps the dim-passthrough working across both major versions, at zero runtime cost (parseProviderOptions short-circuits on first match).

Rebased onto v0.28.12

Branch is rebased onto upstream/master HEAD (bca993e v0.28.12) so v0.28.11 multimodal embedding routing + v0.28.12 LongMemEval harness are preserved alongside the new Vertex AI provider.

^{Need help on this PR? Tag @codesmith with what you need.}

• Let Codesmith autofix CI failures and bot reviews