Skip to contents

Check a drafting answer in its context

Source: R/ai_check.R
tfl_ai_check.Rd

Runs tflspec's own checks on a read answer (tfl_ai_parse()), with the study's context (tfl_ai_context()), and adds their problems to the answer's. Errors block applying the answer; warnings do not.

Usage

tfl_ai_check(answer, context = NULL, adam = NULL)

Arguments

answer

A tfl_ai_answer (tfl_ai_parse()).

context

The tfl_ai_context() the prompt was made with; NULL for none (only what needs no context is checked).

adam

figure: the study's data (tfl_read_adam()), for the checks that look into it; NULL for none.

Value

A data frame of the problems: where (toc row 2 field type, layers[3] point field size ...), problem, severity (error or warning); no rows when there are none.

Details

  • toc: the rows are written to a temporary .csv and read back with tfl_read_toc() (its errors – a row with no output id but other cells filled, an id twice – become problems on their rows); a kind it had to guess, and a row with no id it passes over as a heading, are warnings; a column outside the context's toc_fields is an error.

  • table: each report against the schema – a key a report file does not have, a content function that is not a table's, an argument the function does not have (tfl_schema_check(), an error here: the AI's invention), a value of the wrong type, a required argument missing; its ard: fields and method; the analysis sets of spec/data.yml (with study_dir); with the context's datasets (the ADaM summary), every dataset and variable an analysis names (made by derive or add aside); a report asked for and missing. Then the answer is put in a copy of the study folder and its ARD program (tfl_ard_code()) and each table's code (tfl_content_code()) are written: what fails there is a problem too. Nothing is run.

  • figure: tfl_check_fig_design() on the design (with adam, the data-aware checks too: the variables and PARAMCDs it names); an output_id other than the context's is an error; a template not known, or not among the context's templates, and a dataset not in the context's datasets are warnings.

See also

tfl_ai_repair_prompt() to send them back.

Examples

answer <- c(
  "```yaml",
  "tflspec_ai: {task: toc, version: 1}",
  "toc:",
  "- {output_id: T-14-1-1, type: table, title: Demographics}",
  "- {output_id: T-14-1-1, type: tbl, title: Disposition}",
  "- {output_id: L-16-2-1, type: list, title: Deaths}",
  "assumptions: []",
  "```")
tfl_ai_check(tfl_ai_parse(answer, "toc"), tfl_ai_context("toc"))
#>                       where                   problem severity
#> 1 toc row 1 field output_id 'T-14-1-1' is given twice    error
#> 2 toc row 2 field output_id 'T-14-1-1' is given twice    error