[pridemusvaire]

Why this PR exists

• I am a maintainer / this is internal Roomote work

Memory Settings currently warns "Memory has no embeddings provider, so it can only match keywords" on Cloudflare-only deployments. Those deployments already have a Workers AI token that speaks OpenAI-compatible embeddings, but Memory only treated an explicit embeddings upstream URL or an OpenAI/OpenRouter key as an embeddings provider.

What changed

Memory embeddings auto-wire to Cloudflare Workers AI when CLOUDFLARE_WORKERS_AI_API_TOKEN and a 32-hex CLOUDFLARE_WORKERS_AI_ACCOUNT_ID are set. Semantic recall no longer needs a separate OpenAI or OpenRouter key.

• Explicit R_BRAIN_EMBEDDINGS_UPSTREAM_URL still wins.
• The gateway base is https://api.cloudflare.com/client/v4/accounts/{account}/ai; Roomote appends /v1/embeddings.
• Chat synthesis stays on OpenAI, OpenRouter, or the helper model. Cloudflare is embeddings-only.
• Set R_BRAIN_EMBEDDING_MODEL=@cf/baai/bge-m3 and R_BRAIN_EMBEDDING_DIMENSIONS=1024 before Memory's first boot.

How it was tested

• SDK (19): isBrainEmbeddingAvailable and resolveBrainEmbeddingsUpstream (URL wins, CF /ai base, invalid account including 32-char non-hex, token/account missing, 30s cache on hits and misses)
• API gateway (32): embeddings route to Workers AI without an OpenAI key; chat stays on the provider when CF embeddings are wired
• Web (3): Settings status is incomplete without embeddings and connected when embeddings are available
• Pre-push gates: oxlint, turbo lint, check-types:fast (27/27), knip, gitleaks: pass

Checklist

• The PR title follows the repo convention: [Fix], [Feat], [Improve], [Refactor], [Docs], or [Chore] followed by a user-facing description
• This PR is small and scoped to one change
• pnpm lint and pnpm check-types pass locally
• I added tests or included a clear manual validation note above
• I removed secrets, tokens, private keys, and customer data from code, logs, and screenshots
• If this change should appear in the changelog, I ran pnpm changeset

[mintlify]

Preview approval required

This pull request comes from a fork, so Mintlify won't build a preview until someone with write access approves commit 7fc18b1.

Approve preview

[roomote-community]

No new code issues found. Both earlier routing findings are addressed: Cloudflare-model requests stay on Workers AI when a provider key is added, while OpenAI-model requests retain the provider path. An isolated resolver probe passed for both OpenAI and OpenRouter coexistence, cache isolation, and explicit-upstream precedence. CI tests and the app Docker build are still running. See session

• P1: Preserve the embedding route for existing OpenAI-model brains — packages/sdk/src/server/lib/brain-embeddings-upstream.ts:72-75: Cloudflare credentials unconditionally take over embeddings, forwarding the corpus's fixed OpenAI model to an incompatible provider and breaking writes and semantic queries.
• P2: Keep Cloudflare-model corpora on Cloudflare when a provider key is added — packages/sdk/src/server/lib/brain-embeddings-upstream.ts:58-59: Adding an OpenAI/OpenRouter task or chat key disables Cloudflare routing for a corpus pinned to @cf/baai/bge-m3, sending that unsupported model to the provider instead.

_{Reviewed 7fc18b1}

[pridemusvaire]

Closing this. Memory embeddings should be wired with the documented R_BRAIN_EMBEDDINGS_UPSTREAM_URL (any OpenAI-compatible embeddings server, plus R_BRAIN_INFERENCE_UPSTREAM_API_KEY when that server needs a bearer key). Auto-detecting Cloudflare Workers AI credentials is the wrong product: operators without Cloudflare never needed it, and operators with Cloudflare can point the documented upstream URL at Workers AI themselves.