Choose an embedding model

Set embedder.type in the configuration. Whichever you choose, embedder.dimensions must equal the model’s output and the vector index’s dimension; Keyvector checks this at startup.

	tei	bedrock	openai
What	Hugging Face text-embeddings-inference running any Sentence-Transformers model	Titan Text Embeddings V2, Cohere Embed	OpenAI or any server speaking its /embeddings API (vLLM, Infinity, Ollama)
Where your text goes	Nowhere: it runs beside Keyvector	Bedrock, in your account	That API
Per-token cost	None, you pay for the compute	Yes	Yes
Throughput limit	Your hardware	Titan V2: 60 requests a minute, one text each, not adjustable. Cohere takes 96 texts per request	The provider’s

Titan V2’s quota

Measured on a new account in November 2026: on-demand Titan Text Embeddings V2 allows 60 requests a minute, and AWS marks the quota as not adjustable. Titan takes one text per request, so that’s one row per second, about 3,600 rows an hour. Keyvector paces itself to it by default, so it doesn’t spend its time being throttled. To go faster, set embedder.requests-per-minute only if your account’s quota is higher.

For a large backfill or a busy table:

• use TEI;
• or use Cohere Embed on Bedrock (96 texts per request);
• or embed the backfill with Bedrock batch inference and leave only live changes to Keyvector.

Prefixes

Some models expect the document and the query to be marked: E5 wants passage: and query: . Set embedder.document-prefix and embedder.query-prefix.

Changing models

Vectors from two models can’t be compared. To change models:

1. create a new index;
2. run a second deployment that writes to it;
3. backfill;
4. switch your queries over.