The catalogue for axis 3, and the cheapest way to choose one: the signature
column tells you how a reader traverses the chunks and cost_calls tells you
what that costs, both before you spend anything.
Value
A data frame with one row per registered reader: name, signature
(see gr_reader_signature()), cost_calls (a formula in N, the number
of chunks, not a number), description, and calls, the stems of the
labels its calls carry, joined by commas: the names an effort map can
use.
See also
gr_read(), gr_reader_signature(), gr_register_reader(),
gr_read_spec(), gr_compare() to run several and compare
Other reading functions:
gr_answer,
gr_check_reader(),
gr_read(),
gr_read_spec(),
gr_reader_signature(),
gr_register_reader(),
is_not_found(),
new_answer()
Examples
# Grouped by how they select chunks, which is the real taxonomy:
# `all|...` readers see every chunk, `topk|...` readers see a selection.
r <- gr_readers()
r[order(r$signature), c("name", "signature", "cost_calls")]
#> name signature
#> 13 stuff all|1|none
#> 12 skim all|N+1|none
#> 2 extract all|N+conflicts|none
#> 6 map_reduce all|N+logN|tree
#> 3 hierarchical all|N+tree+1|tree
#> 8 refine all|N|forward
#> 1 ensemble ensemble|sum+1|none
#> 11 screen head|1|none
#> 5 loose one|1|none
#> 7 preview planned|1+s+1|none
#> 10 retrieve topk|1|none
#> 9 rerank topk|m+1|none
#> 4 iterative topk|rounds*2|forward
#> cost_calls
#> 13 1
#> 12 N + 1
#> 2 N (+1 per conflict with resolve = 'model')
#> 6 N + merges
#> 3 N + fan-in levels + 1
#> 8 N
#> 1 sum of members + 1
#> 11 1
#> 5
#> 7 1 + skimmed sections + 1
#> 10 1 + embeddings
#> 9 m + 1
#> 4 up to 2 x max_rounds