Embeddings

Tool id	Vendor	Credentials	Status	Divinci usage
@cf/baai/bge-base-en-v1.5 transforms any given text into a 768-dimensional vector	BAAI	Divinci-managed	available	—
@cf/google/embeddinggemma-300m multilingual Gemma embedding model producing 768-dimensional vectors	Google	Divinci-managed	available	—
gemini-embedding-001@1536 Google's gemini-embedding-001 model truncated to 1536 dimensions — recommended for qdrant-cosine and other 1536d collections.	Google	Divinci-managed	available	—
gemini-embedding-001@3072 Google's gemini-embedding-001 at its full 3072 dimensions — highest retrieval quality. Not compatible with Cloudflare Vectorize (1536 max) or Couchbase (2048 max).	Google	Divinci-managed	available	—
gemini-embedding-2-preview@1536 Google's gemini-embedding-2-preview model truncated to 1536 dimensions via Matryoshka. 8K token input, multimodal capable. Best balance for qdrant-cosine and other 1536d collections.	Google	Divinci-managed	available	—
vertex-ai-text-embedding-005 transforms any given text into a 768-dimensional vector using Google Vertex AI	Google	Divinci-managed	available	—

Usage figures for this category are not collected yet — the weekly job populates them.

Never mix embedding models inside one vector

Vectors produced by two different models live in incompatible spaces, even at identical dimensionality. Mixing them does not error — similarity search just quietly gets worse, in a way that looks like "the corpus isn't very good". The vector's embeddingModel is the source of truth; do not override it per request.

Picking by ingestion shape, not by quality alone

The two Gemini embedding models differ far more in rate limit than in output quality, and that is usually what decides:

| Model | Requests/min | API mode | Use for | | --- | --- | --- | --- | | gemini-embedding-001@1536 | ~100K | batched | Catalog imports and any batch over ~100 files. Survives parallel ingestion. | | gemini-embedding-2-preview@1536 | ~4K | unbatched | Individual uploads and quality-sensitive corpora. Saturates immediately under batch load. |

Practical rule:

• A single PDF, transcript or page — either model finishes inside the window, even with hundreds of chunks.
• A catalog of 30 or fewer items — either works; the preview model will hit its cap briefly and the retry chain absorbs it.
• A bulk import, or a fresh corpus reseed — use gemini-embedding-001, or ingest through a temporary vector on it and re-embed later.

The failure signature of getting this wrong is a pile of Temporary API error (429) / Quota exceeded for … online_prediction_requests entries in the vector's failed sources. That is a rate cap, not a bug: retries do eventually succeed, at the cost of hours of wall-clock.

Dimensions

1536 and 3072 variants of gemini-embedding-001 are separate tools because dimensionality is part of the vector's contract — the store is created at that width. Larger is not free: it costs proportionally more storage and query time for a gain that is corpus-dependent.

See also

• RAG & Vector Stores
• RAG Knowledge Base