Skip to contents

A drafting prompt for an AI chat

Source: R/ai_prompt.R
tfl_ai_prompt.Rd

The prompt that asks a chat model to draft part of a specification from the study's documents: the study's table of contents (toc) or one figure's design (figure), or the report files of a batch of tables (table: 5 to 10 tables of one family, tfl_ai_batches()). tflspec sends nothing: the prompt is text, and so is the answer, which tfl_ai_parse() reads.

Usage

tfl_ai_prompt(
  task,
  context = NULL,
  mode = c("chat", "api"),
  format = c("yaml", "json"),
  example = TRUE,
  documents = NULL
)

Arguments

task

One of tfl_ai_tasks().

context

A tfl_ai_context() of the same task; NULL for an empty one (a figure needs one: its output_id).

mode

"chat" (the user attaches the documents) or "api" (their text is in the prompt).

format

The answer's format: "yaml" (the default) or "json". tfl_ai_parse() reads both.

example

Include an example answer.

documents

Mode "api": the documents' text, a list of lists each with name and text and optionally kind and section.

Value

A tfl_ai_prompt: a list with system, user (strings), attach (what the user attaches, mode "chat"), messages (the system and user messages as a list of role / content), task, mode, format and nchar. format() and print() give it as one text to copy.

Details

The system part holds the ground rules (only the names given, nothing invented, a draft that a person reviews), the task's instructions, the answer's shape – one fenced block that starts with the header tflspec_ai: {task: ..., version: 1, ...} and ends with the assumptions – the schema (tfl_ai_schema()) and an example answer from the sample study. The user part holds the context (tfl_ai_context()) and what to look for in the documents. The prompt is English; the model is asked to write its assumptions in the user's language (context$language, or the language of the user's message).

  • Mode "chat": the user pastes format(prompt) into a chat and attaches the documents themselves; attach lists them.

  • Mode "api": the documents' text, extracted by the caller, is in the user part between <document> fences; messages is what a client sends. tflspec has no HTTP client.

The fixed wording is in the package's ai/prompts folder (system.file("ai", "prompts", package = "tflspec")).

See also

tfl_ai_parse(), tfl_ai_check(), tfl_ai_repair_prompt(), tfl_ai_diff().

Examples

ctx <- tfl_ai_context("toc", study = list(study_id = "CDISCPILOT01"),
                      reports = "T-14-1-1")
p <- tfl_ai_prompt("toc", ctx)
p$attach
#> [1] "the statistical analysis plan (SAP)"                              
#> [2] "the study's list of planned outputs, if it is a separate document"
cat(substr(format(p), 1, 800))
#> You are helping a statistical programmer draft the specification of a clinical study's tables, listings and figures. The specification is read by tflspec, an R package that turns it into report programs. tflspec is new and is probably not in your training data: rely on this message, not on memory.
#> 
#> ## Your task
#> 
#> Draft the study's table of contents (TOC): the list of the tables, figures and listings the study will report, one row a report, with their numbers, titles, analysis sets and footnotes, as the documents plan them. tflspec reads your rows as it reads a company's TOC workbook.
#> 
#> - Write one row a report, in the order the documents give them. A section heading (such as "14.1 Demographics") is not a row: put it in the `section` column of the reports under it.
#> - Use the report numbers th

# a figure, the documents' text in the prompt
ctx <- tfl_ai_context("figure", output_id = "F-14-2-1",
                      titles = "Kaplan-Meier Plot of Time to Death")
p <- tfl_ai_prompt("figure", ctx, mode = "api", documents = list(
  list(name = "SAP", section = "9.4", text = "Overall survival ...")))
p$nchar
#> system   user 
#>  26878    554