tabscreen_gpt.tools: Title and abstract screening with GPT API models using function calls via the tools argument

Description

This function supports the conduct of title and abstract screening with GPT API models in R. Specifically, it allows the user to draw on GPT-3.5, GPT-4, GPT-4o, GPT-4o-mini, and fine-tuned models. The function allows to run title and abstract screening across multiple prompts and with repeated questions to check for consistency across answers. All of which can be done in parallel. The function draws on the newly developed function calling which is called via the tools argument in the request body. This is the main different between tabscreen_gpt.tools() and tabscreen_gpt.original(). Function calls ensure more reliable and consistent responses to ones requests. See Vembye et al. (2024) for guidance on how adequately to conduct title and abstract screening with GPT models.

Usage

tabscreen_gpt.tools(data, prompt, studyid, title, abstract,
   model = "gpt-4o-mini", role = "user", tools = NULL, tool_choice = NULL, top_p = 1,
   time_info = TRUE, token_info = TRUE, api_key = get_api_key(), max_tries = 16,
   max_seconds = NULL, is_transient = gpt_is_transient, backoff = NULL,
   after = NULL, rpm = 10000, reps = 1, seed_par = NULL, progress = TRUE,
   decision_description = FALSE, messages = TRUE, incl_cutoff_upper = NULL,
   incl_cutoff_lower = NULL, force = FALSE, fine_tuned = FALSE, ...)
tabscreen_gpt(data, prompt, studyid, title, abstract,
   model = "gpt-4o-mini", role = "user", tools = NULL, tool_choice = NULL, top_p = 1,
   time_info = TRUE, token_info = TRUE, api_key = get_api_key(), max_tries = 16,
   max_seconds = NULL, is_transient = gpt_is_transient, backoff = NULL,
   after = NULL, rpm = 10000, reps = 1, seed_par = NULL, progress = TRUE,
   decision_description = FALSE, messages = TRUE, incl_cutoff_upper = NULL,
   incl_cutoff_lower = NULL, force = FALSE, fine_tuned = FALSE, ...)

Value

An object of class 'gpt'. The object is a list containing the following datasets and components:

answer_data: dataset of class 'gpt_tbl' with all individual answers.
price_dollar: numerical value indicating the total price (in USD) of the screening.
price_data: dataset with prices across all gpt models used for screening.
run_date: string indicating the date when the screening was ran. In some frameworks, time details are considered important to report (see e.g., Thomas et al., 2024).
...: some additional attributed values/components, including an attributed list with the arguments used in the function. These are used in screen_errors() to re-screen transient errors.

If the same question is requested multiple times, the object will also contain the following dataset with results aggregated across the iterated requests/questions.

answer_data_aggregated: dataset of class 'gpt_agg_tbl' with the summarized, probabilistic inclusion decision for each title and abstract across multiple repeated questions.

Arguments

data: Dataset containing the titles and abstracts.
prompt: Prompt(s) to be added before the title and abstract.
studyid: Unique Study ID. If missing, this is generated automatically.
title: Name of the variable containing the title information.
abstract: Name of variable containing the abstract information.
model: Character string with the name of the completion model. Can take multiple models. Default is the latest "gpt-4o-mini". Find available model at https://platform.openai.com/docs/models/model-endpoint-compatibility.
role: Character string indicating the role of the user. Default is "user".
tools: This argument allows this user to apply customized functions. See https://platform.openai.com/docs/api-reference/chat/create#chat-create-tools. Default is NULL. If not specified the default function calls from AIscreenR are used.
tool_choice: If a customized function is provided this argument 'controls which (if any) tool is called by the model' (OpenAI). Default is NULL. If set to NULL when using a customized function, the default is "auto". See https://platform.openai.com/docs/api-reference/chat/create#chat-create-tool_choice.
top_p: 'An alternative to sampling with temperature, called nucleus sampling, where the model considers the results of the tokens with top_p probability mass. So 0.1 means only the tokens comprising the top 10% probability mass are considered. We generally recommend altering this or temperature but not both.' (OpenAI). Default is 1. Find documentation at https://platform.openai.com/docs/api-reference/chat/create#chat/create-top_p.
time_info: Logical indicating whether the run time of each request/question should be included in the data. Default is TRUE.
token_info: Logical indicating whether token information should be included in the output data. Default is TRUE. When TRUE, the output object will include price information of the conducted screening.
api_key: Numerical value with your personal API key. Default setting draws on the get_api_key() to retrieve the API key from the R environment, so that the key is not compromised. The API key can be added to the R environment via set_api_key() or by using usethis::edit_r_environ(). In the .Renviron file, write CHATGPT_KEY=INSERT_YOUR_KEY_HERE. After entering the API key, close and save the .Renviron file and restart RStudio (ctrl + shift + F10). Alternatively, one can use httr2::secret_make_key(), httr2::secret_encrypt(), and httr2::secret_decrypt() to scramble and decrypt the API key.
max_tries, max_seconds: 'Cap the maximum number of attempts with max_tries or the total elapsed time from the first request with max_seconds. If neither option is supplied (the default), httr2::req_perform() will not retry' (Wickham, 2023). The default of max_tries is 16.
is_transient: 'A predicate function that takes a single argument (the response) and returns TRUE or FALSE specifying whether or not the response represents a transient error' (Wickham, 2023). This function runs automatically in the AIscreenR but can be customized by the user if necessary.
backoff: 'A function that takes a single argument (the number of failed attempts so far) and returns the number of seconds to wait' (Wickham, 2023).
after: 'A function that takes a single argument (the response) and returns either a number of seconds to wait or NULL, which indicates that a precise wait time is not available that the backoff strategy should be used instead' (Wickham, 2023).
rpm: Numerical value indicating the number of requests per minute (rpm) available for the specified model. Find more information at https://platform.openai.com/docs/guides/rate-limits/what-are-the-rate-limits-for-our-api. Alternatively, use rate_limits_per_minute().
reps: Numerical value indicating the number of times the same question should be send to the server. This can be useful to test consistency between answers, and/or can be used to make inclusion judgments based on how many times a study has been included across a the given number of screenings. Default is 1 but when using gpt-3.5-turbo models or gpt-4o-mini, we recommend setting this value to 10 to catch model uncertainty.
seed_par: Numerical value for a seed to ensure that proper, parallel-safe random numbers are produced.
progress: Logical indicating whether a progress line should be shown when running the title and abstract screening in parallel. Default is TRUE.
decision_description: Logical indicating whether a detailed description should follow the decision made by GPT. Default is FALSE. When conducting large-scale screening, we generally recommend not using this feature as it will substantially increase the cost of the screening. We generally recommend using it when encountering disagreements between GPT and human decisions.
messages: Logical indicating whether to print messages embedded in the function. Default is TRUE.
incl_cutoff_upper: Numerical value indicating the probability threshold for which a studies should be included. ONLY relevant when the same questions is requested multiple times (i.e., when any reps > 1). Default is 0.5, indicating that titles and abstracts should only be included if GPT has included the study more than 50 percent of the times.
incl_cutoff_lower: Numerical value indicating the probability threshold above which studies should be check by a human. ONLY relevant when the same questions is requested multiple times (i.e., when any reps > 1). Default is 0.4, meaning that if you ask GPT the same questions 10 times and it includes the title and abstract 4 times, we suggest that the study should be check by a human.
force: Logical argument indicating whether to force the function to use more than 10 iterations for gpt-3.5 models and more than 1 iteration for gpt-4 models other than gpt-4o-mini. This argument is developed to avoid the conduct of wrong and extreme sized screening. Default is FALSE.
fine_tuned: Logical indicating whether a fine-tuned model is used. Default is FALSE.
...: Further argument to pass to the request body. See https://platform.openai.com/docs/api-reference/chat/create.

References

Vembye, M. H., Christensen, J., Mølgaard, A. B., & Schytt, F. L. W. (2024) GPT API Models Can Function as Highly Reliable Second Screeners of Titles and Abstracts in Systematic Reviews: A Proof of Concept and Common Guidelines https://osf.io/preprints/osf/yrhzm

Thomas, J. et al. (2024). Responsible AI in Evidence SynthEsis (RAISE): guidance and recommendations. https://osf.io/cn7x4

Wickham H (2023). httr2: Perform HTTP Requests and Process the Responses. https://httr2.r-lib.org, https://github.com/r-lib/httr2.

Examples

Run this code

if (FALSE) {

library(future)

set_api_key()

prompt <- "Is this study about a Functional Family Therapy (FFT) intervention?"

plan(multisession)

tabscreen_gpt(
  data = filges2015_dat[1:2,],
  prompt = prompt,
  studyid = studyid,
  title = title,
  abstract = abstract
  )

plan(sequential)

 # Get detailed descriptions of the gpt decisions.

 plan(multisession)

 tabscreen_gpt(
   data = filges2015_dat[1:2,],
   prompt = prompt,
   studyid = studyid,
   title = title,
   abstract = abstract,
   decision_description = TRUE
 )

plan(sequential)

}

Run the code above in your browser using DataLab