Embedding models

The embedding models Aimogen Pro supports across OpenAI, Google and Ollama, their dimensions, and why changing model means rebuilding the index.

An embedding model turns text into a vector. Which one you use decides retrieval quality and the dimension your vector store must be created with.

Settings › Embeddings › Embeddings Model

Available models#

OpenAI#

Model	Note
`text-embedding-3-small`	The plugin default. Dimension 1536. Cheap and good
`text-embedding-3-large`	Higher quality, larger dimension, more expensive
`text-embedding-ada-002`	Previous generation, dimension 1536. Kept for existing indexes

Google#

Model
`gemini-embedding-2-preview`
`gemini-embedding-001`
`text-embedding-005`

Requires a Google AI Studio key.

Ollama#

A long list of open embedding models, including nomic-embed-text, mxbai-embed-large, snowflake-arctic-embed, all-minilm, bge-m3 and many community models. Only models you have actually pulled appear.

bash

ollama pull nomic-embed-text

Requires an Ollama URL.

Choosing#

Situation	Choose
Default	`text-embedding-3-small`
Quality matters more than cost	`text-embedding-3-large`, with a matching index dimension
Already on Google	`gemini-embedding-001`
Content must not leave your infrastructure	Ollama with `nomic-embed-text` or `bge-m3`
Existing index built on ada-002	Keep `text-embedding-ada-002` until you rebuild

Dimension matters#

Vector stores are created with a fixed dimension. The model output must match it.

Model	Dimension
`text-embedding-3-small`	1536
`text-embedding-ada-002`	1536
`text-embedding-3-large`	larger
Ollama models	varies by model

The plugin help text for Pinecone, Qdrant and Chroma all tell you to create the index with dimension 1536 and metric cosine, because that matches the default model.

If you choose a model with a different dimension, create the index to match.

Changing model means rebuilding#

This is the single most damaging mistake in RAG configuration, and it fails quietly — no error, just progressively worse answers.

The embedding provider is independent of the vector store#

Any combination works:

Embed with OpenAI, store in Pinecone
Embed with Ollama, store in Qdrant — nothing leaves your infrastructure
Embed with OpenAI, store in OpenAI Vector Stores — the simplest path

Choose the embedding model for quality and privacy; choose the store for operations and cost.

Cost#

Embedding is the cheapest AI operation. Indexing a thousand typical blog posts with text-embedding-3-small costs a fraction of generating one article.

The ongoing cost is not embedding but injection: retrieved chunks are input tokens on every affected request. See Retrieval and prompt injection.

Privacy#

Embedding sends your content to the provider. If your knowledge base contains confidential material, Ollama with a local model keeps it on your hardware — and pairing that with a self-hosted Qdrant or Chroma means the whole retrieval path stays internal.

See Data, GDPR and retention.

Still stuck? Open a support ticket and include the diagnostics from Aimogen Pro › System & Logs › System Info.