Skip to content

Ollama Embedding Engine

Source: R/sentence_embeddings-ollama.R
embedding_ollama.Rd

embedding_ollama() creates an engine for step_sentence_embeddings() that gets embeddings from a running Ollama server. No Python environment or API key is needed, the engine sends HTTP requests to the server, which is usually running on your own machine.

Usage

embedding_ollama(
  model,
  host = "http://localhost:11434",
  text_prefix = NULL,
  normalize = FALSE,
  batch_size = 64
)

Arguments

model

A single string giving the name of the Ollama model, such as "all-minilm". The model must already be available on the server, which you can do with ollama pull <model>. There is no default, so the model is always visible in your code.

host

A single string giving the address of the Ollama server. Defaults to "http://localhost:11434".

text_prefix

A single string added to the start of every text before it is embedded, or NULL for no prefix. Some models are trained with prefixes, for example "query: " or "passage: ". Defaults to NULL.

normalize

A single logical. If TRUE, each embedding is scaled to have unit length. Defaults to FALSE, which returns the vectors as the server returns them.

batch_size

A single positive whole number giving the maximum number of texts sent in one request. Defaults to 64.

Value

An object of class sentence_embedding_engine.

Details

Embeddings are requested from the /api/embed endpoint. Requests that fail with a transient error (too many requests, or the server being unavailable) are retried a few times. Any other failure stops bake() with an error, so partial results are never returned. Texts longer than the context length of the model are an error rather than being silently truncated.

Before any text is embedded, the engine checks that the server can be reached and that model is available on it, and gives an error that says how to fix it if not. The text is sent to the server at host. For a server you don't control, keep in mind that this is where your data goes.

The engine stores the model name and the server address, not the model. A prepped recipe therefore needs the same model to be available on the server when it is baked, and Ollama model names, such as "all-minilm", can point to different model versions over time. The engine doesn't pin a version.

Requires the httr2 package.

Setting up Ollama

The engine needs an Ollama server that is running and has the model.

  1. Install Ollama from https://ollama.com/download. See the Ollama documentation for your system.

  2. Make sure the server is running. The desktop apps start it when you open them. Otherwise run ollama serve in a terminal and leave it running.

  3. Pull the model once, for example ollama pull all-minilm. Models are listed at https://ollama.com/library.

You can check which models are on the server with ollama list.

See also

step_sentence_embeddings()

Other Embedding Engines: embedding_custom(), embedding_huggingface()

Examples

if (FALSE) {
# Needs a running Ollama server with the model pulled:
#   ollama pull all-minilm
library(recipes)

sample_data <- tibble(
  text = c("The cat sat.", "A dog ran.", "Stocks fell sharply today.")
)

rec <- recipe(~text, data = sample_data) |>
  step_sentence_embeddings(
    text,
    engine = embedding_ollama("all-minilm")
  )

rec |>
  prep() |>
  bake(new_data = NULL)
}