Skip to Content
remimem

Embedding providers

Use the default local model, a larger code-trained model, OpenAI or Voyage, or no embeddings at all.

Embeddings power search by meaning. The provider is set in .remimem/config.json under embedding.

ProviderWhere it runsCode leaves your machineDefault model
local (default)On your CPUNoXenova/all-MiniLM-L6-v2
openaiOpenAI's APIYestext-embedding-3-small
voyageVoyage AI's APIYesvoyage-code-3
noneNowhereNonone: keyword and exact-name search only

Changing the model

Vectors from different models cannot be mixed. After changing provider or model, run remimem index --full. Until you do, searches stop with a message telling you to.

The default local model

Xenova/all-MiniLM-L6-v2 is small (23 MB, 384 dimensions) and fast (about 200 chunks a second on an Apple M5). It is a general-purpose model, not one trained on code.

A code-trained local model

.remimem/config.json
{
  "embedding": { "provider": "local", "model": "jinaai/jina-embeddings-v2-base-code" }
}

This model is trained on code, but on the benchmark it found no more of the right files (92% against 93%) while indexing about five times slower (324 s against 62 s for the three benchmark repositories) and using a 163 MB download. It ranked the first right file slightly higher (MRR 0.65 against 0.61). Try it if your queries are mostly code-shaped and indexing time does not matter to you.

Other Hugging Face feature-extraction models can be used if they publish 8-bit quantized ONNX weights (a model_quantized.onnx file), which is what the local provider loads. Only these two have been measured.

OpenAI or Voyage

.remimem/config.json
{
  "embedding": { "provider": "openai", "model": "text-embedding-3-small" }
}
export OPENAI_API_KEY=sk-...
remimem index --full

API keys are read from the environment only, never from the config file, so the config stays safe to commit. Requests are batched (batchSize, 8 by default) and retried on rate limits and server errors.

Privacy

With a remote provider, the text of every indexed chunk is sent to that service when it is embedded, and every query when you search. remimem init says so when it runs. These providers are tested only against stubbed responses; they have not been exercised against the live services.

No embeddings

.remimem/config.json
{
  "embedding": { "provider": "none" }
}

Nothing is downloaded and indexing is fastest. Search uses the keyword index and exact symbol names. On the benchmark, keyword search alone found 80% of the right files against 93% for the full hybrid search.

Settings

KeyDefaultMeaning
embedding.providerlocallocal, openai, voyage, or none
embedding.modelprovider defaultModel name
embedding.batchSize8Texts per embedding call
embedding.maxTokens256Local models only: inputs are cut to this many model tokens

On this page