Get text embeddings with optional caching — get

A thin wrapper around fuzzylink::get_embeddings() that adds file-based caching. If cache points to an existing file, embeddings for documents already present in the cache are loaded from disk. Embeddings for any documents not found in the cache are fetched via the API and appended to the cache before returning.

Usage

get_embeddings(
  documents,
  cache = NULL,
  type = c("text", "image"),
  model = NULL,
  ...
)

Arguments

documents: A character vector of texts to embed. When type = "image", a character vector of paths to local image files.
cache: Optional path to an .rds file. If the file exists, cached embeddings matching documents (by row name) are reused. Any new documents are fetched and written back to the cache. If NULL, embeddings are always computed fresh without caching.
type: One of "text" (default) or "image". "text" embeds documents via fuzzylink::get_embeddings() (OpenAI/Mistral text embeddings). "image" treats documents as local image file paths and embeds them via OpenRouter's multimodal embeddings endpoint, which requires an OPENROUTER_API_KEY environment variable.
model: Character string naming the embedding model to use. When NULL (the default), each type uses its own built-in default: "text-embedding-3-large" for type = "text", or "google/gemini-embedding-2" for type = "image".
...: Additional arguments passed to fuzzylink::get_embeddings() (for type = "text") or to the internal OpenRouter request body (for type = "image").

Value

A numeric matrix with one row per document and one column per embedding dimension. Row names are set to documents.