Skip to contents

Uses vector similarity search

Source: R/retrieve.R
ragnar_retrieve_vss.Rd

Computes a similarity measure between the query and the documents embeddings and uses this similarity to rank the documents.

Usage

ragnar_retrieve_vss(
  store,
  text,
  top_k = 3L,
  method = c("cosine_distance", "cosine_similarity", "euclidean_distance", "dot_product",
    "negative_dot_product")
)

Arguments

store

A RagnarStore object or a dplyr::tbl() derived from it. When you pass a tbl, you may use usual dplyr verbs (e.g. filter(), slice()) to restrict the rows examined before similarity scoring. Avoid dropping essential columns such as text, embedding, origin, and hash.

text

A string to find the nearest match too

top_k

Integer, maximum amount of document chunks to retrieve

method

A string specifying the method used to compute the similarity between the query and the document chunks embeddings store in the database.

Value

A dataframe of retrieved chunks. Each row corresponds to an individual chunk in the store. It always contains a column named text that contains the chunks.

Details

The supported methods are:

  • cosine_distance: Measures the dissimilarity between two vectors based on the cosine of the angle between them. Defined as \(1 - cos(\theta)\), where \(cos(\theta)\) is the cosine similarity.

  • cosine_similarity: Measures the similarity between two vectors based on the cosine of the angle between them. Ranges from -1 (opposite) to 1 (identical), with 0 indicating orthogonality.

  • euclidean_distance: Computes the straight-line (L2) distance between two points in a multidimensional space. Defined as \(\sqrt{\sum(x_i - y_i)^2}\).

  • dot_product: Computes the sum of the element-wise products of two vectors.

  • negative_dot_product: The negation of the dot product.

Pre-filtering with dplyr

The store behaves like a lazy table backed by DuckDB, so row‑wise filtering is executed directly in the database. This lets you narrow the search space efficiently without pulling data into R.

See also

Other ragnar_retrieve: ragnar_retrieve(), ragnar_retrieve_bm25(), ragnar_retrieve_vss_and_bm25()

Examples

if (FALSE) { # (rlang::is_installed("dbplyr") && nzchar(Sys.getenv("OPENAI_API_KEY")))
# Basic usage
store <- ragnar_store_create(
  embed = \(x) ragnar::embed_openai(x, model = "text-embedding-3-small")
)
ragnar_store_insert(store, data.frame(text = c("foo", "bar")))
ragnar_store_build_index(store)
ragnar_retrieve(store, "foo")

# More Advanced: store metadata, retrieve with pre-filtering
store <- ragnar_store_create(
  embed = \(x) ragnar::embed_openai(x, model = "text-embedding-3-small"),
  extra_cols = data.frame(category = character())
)

ragnar_store_insert(
  store,
  data.frame(
    category = "desert",
    text = c("ice cream", "cake", "cookies")
  )
)

ragnar_store_insert(
  store,
  data.frame(
    category = "meal",
    text = c("steak", "potatoes", "salad")
  )
)

ragnar_store_build_index(store)

# simple retrieve
ragnar_retrieve(store, "carbs")

# retrieve with pre-filtering
dplyr::tbl(store) |>
  dplyr::filter(category == "meal") |>
  ragnar_retrieve("carbs")
}