Bind an ingestion, segmentation and reading configuration together

A recipe is the unit gr_compare() treats as one experiment. Fixing two axes and varying the third is what makes a difference in the answer attributable to the change you made. That is exactly what the previous release could not do, because its modes shared one chunk object.

Usage

gr_recipe(name = NULL, ingest = NULL, segment = NULL, read = NULL)

Arguments

name

Label used in results and traces.

ingest

A gr_ingest_spec, preset name, or named list.

segment

A gr_segment_spec, segmenter name, or named list.

read

A gr_read_spec, reader name, or named list.

Value

A gr_recipe: a list of name, ingest, segment and read, each a validated spec. It has a print() method that shows all three at once.

See also

gr_recipes() for the built-ins, answer_document() to run one, gr_compare() to run several, gr_ingest_spec(), gr_segment_spec(), gr_read_spec()

Examples

gr_recipe("semantic_topk", segment = list(method = "semantic", max_tokens = 800),
          read = list(reader = "retrieve", top_k = 6))
#> <gr_recipe 'semantic_topk'>
#>   ingest  : clean=standard ocr=auto
#>   segment : semantic (max 800 tok, overlap 0, min 0)
#>   read    : retrieve [topk|1|none] model=gpt-5.6-terra

# Start from a built-in and change one axis. The other two stay fixed, so
# any difference in the answer is attributable to the change.
r <- gr_recipes("thorough")
r$segment <- gr_segment_spec(method = "structural", max_tokens = 1200)
r$name <- "thorough_structural"
r
#> <gr_recipe 'thorough_structural'>
#>   ingest  : clean=standard ocr=auto
#>   segment : structural (max 1200 tok, overlap 0, min 0)
#>   read    : map_reduce [all|N+logN|tree] model=gpt-5.6-terra