fix: replace deprecated Gemini embedding model (text-embedding-004 → gemini-embedding-001)

[vincenzopalazzo]

Problem

text-embedding-004 is no longer served by Google's Generative Language API. Calls to /v1beta/models/text-embedding-004:batchEmbedContents (and the single-contentvariant) return HTTP 404 for newly-issued API keys.

This means a fresh deployment of OpenMemory configured with OM_EMBEDDINGS=gemini falls back to synthetic embeddings on every call. Logs look like:

[EMBED] Provider: gemini, Tier: smart, Sector: semantic
[EMBED] Gemini error (1/3): Gemini: 404
[EMBED] Gemini error (2/3): Gemini: 404
[EMBED] Gemini batch failed, falling back to sequential: Error: Gemini failed after 3 attempts: Gemini: 404
[EMBED] gemini failed: Gemini failed after 3 attempts: Gemini: 404, trying synthetic
[EMBED] Fallback to synthetic succeeded for sector: semantic

The fallback is silent at the API surface (queries still return), but recall quality drops to the synthetic baseline — semantic search effectively stops working.

Verified against the live API on 2026-05-09 with a fresh AI Studio key:

GET  /v1beta/models?key=…
→ models/gemini-embedding-001        ✅ supports embedContent, asyncBatchEmbedContent
   models/gemini-embedding-2-preview ✅
   models/gemini-embedding-2         ✅
   (text-embedding-004 / embedding-001 are not listed)

POST /v1beta/models/text-embedding-004:embedContent     → 404
POST /v1beta/models/gemini-embedding-001:embedContent   → 200, returns vector
POST /v1beta/models/gemini-embedding-001:batchEmbedContents → 200

The older embedding-001 (which get_defaults() in core/models.ts falls back to) is also deprecated.

Fix

Replace both deprecated model names with gemini-embedding-001 in three places:

File	Change
packages/openmemory-js/src/memory/embed.ts	Hardcoded URL + request-body model field in emb_gemini() (lines 310, 314), and the getEmbeddingInfo() self-report (line 708).
models.yml	gemini: entry for all 5 sectors + the documentation comment block at the bottom.
packages/openmemory-js/src/core/models.ts	get_defaults() fallback for all 5 sectors.

gemini-embedding-001 exposes the same batchEmbedContents endpoint with the same request body shape (model, content, taskType); the response vectors get resized to env.vec_dim via the existing resize_vec helper, so no caller-side changes are needed.

Verification

Tested on a smart-tier deployment with 159 existing memories:

• Pre-fix: every Gemini call 404s, queries fall through to synthetic.
• Post-fix: queries return semantically-ranked matches in <1s, no errors in logs. Paraphrase queries ("travel plans next week" → matches a Pre-Miami plan on track memory with no shared keywords) confirm real semantic recall, not lexical fallback.

Out of scope (follow-up)

The emb_gemini() function does not call get_model(s, "gemini") like every other provider — it hardcodes the model name in the URL. That means models.yml's gemini: entries are loaded by models.ts but never read by the Gemini path. A separate PR should refactor emb_gemini to use get_model (and ideally also support a per-call outputDimensionality query param). I left that out of this PR to keep the scope to a model-name correction.

Also worth considering: an OM_GEMINI_MODEL env override (mirroring OM_OPENAI_MODEL / OM_OLLAMA_MODEL) so users can opt into gemini-embedding-2 without code changes.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 007e3cb72d

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Re-embed vectors when switching Gemini model

When upgrading a database that already contains Gemini vectors produced by text-embedding-004/embedding-001, this line changes query and newly-written embeddings to gemini-embedding-001 without invalidating the old stored vectors. The vector table stores only v/dim and no model identity (packages/openmemory-js/src/core/db.ts:161), while search compares the new query vector against every stored vector in the sector (packages/openmemory-js/src/core/vector/postgres.ts:58), so existing memories remain in the old embedding space and semantic ranking becomes unreliable until they are re-embedded. Please add a rebuild/migration path or a model-version guard before silently switching the default.

Useful? React with 👍 / 👎.