Skip to contents

Ready-made recipes

Source: R/pipeline.R
gr_recipes.Rd

Named starting points that pair a segmentation strategy with a reader that suits it. Each is a plain gr_recipe, so you can modify any field.

Usage

gr_recipes(name = NULL)

Arguments

name

Optional recipe name, or a character vector of names; omit to list them all. A single name returns a gr_recipe; several return a named list of them.

Value

A gr_recipe when name is a single string, otherwise a named list of gr_recipes.

"auto"

answer_document() defaults to recipe = "auto". It is not a recipe but a choice between two of these, made once the document is ingested: "fast" for a document of at most 50,000 tokens (less on a model with a small context window), "thorough" for anything longer. "Choosing the recipe" in answer_document() gives the whole rule. Functions that read several documents or compare recipes take one fixed recipe, and refuse "auto".

See also

gr_recipe() to build your own, answer_document(), gr_compare(), gr_segmenters() and gr_readers() for the pieces they are made of

Examples

# What each built-in actually is, in one table.
do.call(rbind, lapply(names(gr_recipes()), function(n) {
  r <- gr_recipes(n)
  data.frame(recipe = n, clean = paste(as.character(r$ingest$clean), collapse = "+"),
             segment = r$segment$method, max_tokens = r$segment$max_tokens,
             reader = r$read$reader)
}))
#>       recipe    clean    segment max_tokens       reader
#> 1       fast standard  paragraph       4000        stuff
#> 2    precise standard   sentence        600         skim
#> 3     needle standard   semantic        500     retrieve
#> 4   thorough standard  paragraph       1200   map_reduce
#> 5     survey standard structural       1500 hierarchical
#> 6  narrative standard  paragraph       1500       refine
#> 7    scanned     scan       page       2000       rerank
#> 8   research academic structural        900    iterative
#> 9  consensus standard  recursive       1000     ensemble
#> 10    legacy   legacy  paragraph       3000   map_reduce

gr_recipes("precise")
#> <gr_recipe 'precise'>
#>   ingest  : clean=standard ocr=auto
#>   segment : sentence (max 600 tok, overlap 60, min 0)
#>   read    : skim [all|N+1|none] model=gpt-5.6-terra