Initial CRAN submission.
Deciders: decider_typesafe(),
decider_openai(), decider_foundry(),
decider_openrouter() and
decider_systemone().
Questions: question_noul(),
question_choice() and
question_score().
decide() calls one model for many inputs. It returns
a tibble in the tidymodels style, with one row for each input in input
order. A failed call keeps its row, with NA answers and the
reason in .error.
Setups: decision_setup(), write_setup()
and read_setup() keep the question, the descriptions, the
references and the calibrator in one JSON file.
Scoring: score_setup() scores a setup on labeled
rows. evaluate_once() runs the one-time test.
Six vignettes: getting started, choosing a model, open models, grading with a rubric, images, and the industrial image inspection case study.
Four datasets: ohio_injuries,
ohio_recalls, textiles and
rubric_answers.
An agent skill, in
system.file("skills", "decisions", package = "decisions").
Requirements: R 4.3.0 or newer (for @ on S7
objects), httr2 1.2.0 or newer, and rlang 1.3.0 or newer. rlang 1.3.0
changed the values of rlang::hash(), and the
path files, the run hash and the calibrator hash of a setup
use it. With an older rlang, a path file or a setup from a
newer rlang does not match.