Write a study's table requests, one file a request

Splits the TOC's tables into requests (tfl_ai_batches()) and writes the prompt of each (tfl_ai_prompt()) to its own file, to paste into a chat as one message: no attachment is needed. A request longer than max_chars is split again, until it fits or holds one report. The requests' contexts are kept beside them (_contexts.rds), so tfl_ai_collect_answers() checks each answer against its own request.

Usage

tfl_ai_write_requests(
  study_dir,
  shell,
  adam,
  size = 8L,
  dir = file.path("ai", "requests"),
  max_chars = NULL,
  families = NULL,
  types = "table",
  language = NULL,
  examples = NULL,
  instructions = c("each", "once"),
  workbook = c("text", "attach", "both"),
  overwrite = FALSE
)

Arguments

study_dir

The study folder (tfl_ai_study_init()).

shell

The TOC and shells (tfl_ai_read_shell()).

adam

The ADaM summary (tfl_ai_adam_summary()).

size

The reports a request holds at most (1 or more).

dir

The folder the requests go to, under study_dir.

max_chars

A request's length at most, in characters; NULL for no limit.

families, types

As tfl_ai_batches().

language

As tfl_ai_context(): the language of the AI's assumptions.

examples

Reviewed report files for the AI to follow (tfl_ai_context()'s examples). Named by family (c(ae = "spec/outputs/T-14-3-1.yml")), each goes only to its family's requests; unnamed, to every request.

instructions

"each" (default): every request holds the instructions, one paste a request. "once": the instructions in 00-instructions.md, the requests without them.

workbook

How the TOC / SHELL workbook reaches the chat: "text" (default) the shells as text in each request (tfl_ai_read_shell()); "attach" the workbook attached to each chat, the request saying where each report's shell is in it (its sheet, cells and TOC row) instead of the text; "both" the places and the text, the attachment right where the two differ – for a workbook whose text is not read right. A chat that takes no Excel file needs "text".

overwrite

TRUE to write over the requests already there. The folder the answers go to (ai/answers, beside dir) is made too.

Value

A data frame, one row a request: request (the file's name, ae-1), family, outputs (the report ids, comma-separated), n, chars, file. The same is written to index.md in dir.

Details

Start small (size = 2 or 3), check the answers, then raise size while the chat keeps up.

A request holds the instructions and schema (about 23,000 characters, the same in every request) and its own reports (the TOC rows, shells and ADaM summary). For a chat that cannot take that in one message, instructions = "once" writes the instructions to 00-instructions.md and each request without them: paste 00-instructions.md first in a new chat, then the request.

See also

tfl_ai_collect_answers() for the answers.

Examples

f <- system.file("ai", "example", "onco_toc_shell.xlsx", package = "tflspec")
d <- file.path(tempdir(), "req-study")
dir.create(file.path(d, "adam"), recursive = TRUE, showWarnings = FALSE)
if (requireNamespace("haven", quietly = TRUE)) {
  haven::write_xpt(as.data.frame(lapply(adsl, function(x)
    if (is.factor(x)) as.character(x) else x)), file.path(d, "adam", "adsl.xpt"))
  tfl_ai_study_init(d, overwrite = TRUE)
  r <- tfl_ai_write_requests(d, tfl_ai_read_shell(f),
                             tfl_ai_adam_summary(file.path(d, "adam")),
                             size = 2, overwrite = TRUE)
  r[c("request", "outputs", "chars")]
}
#> 4 requests (4 reports) in ai/requests; the longest 29341 characters.
#> Next: paste ai/requests/demographics-1.md into a new chat; save the whole answer as ai/answers/demographics-1.md, then run tfl_ai_collect_answers().
#>          request  outputs chars
#> 1 demographics-1 T-14-1-1 29341
#> 2           ae-1 T-14-3-1 29222
#> 3         labs-1 T-14-3-5 29043
#> 4     efficacy-1 T-14-2-1 29214