blob: 7414e35861ffca71ab556b9af92d1dd7e7d23ffc [file] [log] [blame]
#' Makes bunch completion requests to the GPT-3 API
#'
#' @description
#' `gpt3_completions()` is the package's main function for rquests and takes as input a vector of prompts and processes each prompt as per the defined parameters. It extends the `gpt3_single_completion()` function to allow for bunch processing of requests to the Open AI GPT-3 API.
#' @details
#' The easiest (and intended) use case for this function is to create a data.frame or data.table with variables that contain the prompts to be requested from GPT-3 and a prompt id (see examples below).
#' For a general guide on the completion requests, see [https://beta.openai.com/docs/guides/completion](https://beta.openai.com/docs/guides/completion). This function provides you with an R wrapper to send requests with the full range of request parameters as detailed on [https://beta.openai.com/docs/api-reference/completions](https://beta.openai.com/docs/api-reference/completions) and reproduced below.
#'
#' For the `best_of` parameter: The `gpt3_single_completion()` (which is used here in a vectorised manner) handles the issue that best_of must be greater than n by setting `if(best_of <= n){ best_of = n}`.
#'
#' If `id_var` is not provided, the function will use `prompt_1` ... `prompt_n` as id variable.
#'
#' Parameters not included/supported:
#' - `logit_bias`: [https://beta.openai.com/docs/api-reference/completions/create#completions/create-logit_bias](https://beta.openai.com/docs/api-reference/completions/create#completions/create-logit_bias)
#' - `echo`: [https://beta.openai.com/docs/api-reference/completions/create#completions/create-echo](https://beta.openai.com/docs/api-reference/completions/create#completions/create-echo)
#' - `stream`: [https://beta.openai.com/docs/api-reference/completions/create#completions/create-stream](https://beta.openai.com/docs/api-reference/completions/create#completions/create-stream)
#'
#' @param prompt_var character vector that contains the prompts to the GPT-3 request
#' @param id_var (optional) character vector that contains the user-defined ids of the prompts. See details.
#' @param param_model a character vector that indicates the [model](https://beta.openai.com/docs/models/gpt-3) to use; one of "text-davinci-003" (default), "text-davinci-002", "text-davinci-001", "text-curie-001", "text-babbage-001" or "text-ada-001"
#' @param param_output_type character determining the output provided: "complete" (default), "text" or "meta"
#' @param param_suffix character (default: NULL) (from the official API documentation: _The suffix that comes after a completion of inserted text_)
#' @param param_max_tokens numeric (default: 100) indicating the maximum number of tokens that the completion request should return (from the official API documentation: _The maximum number of tokens to generate in the completion. The token count of your prompt plus max_tokens cannot exceed the model's context length. Most models have a context length of 2048 tokens (except for the newest models, which support 4096)_)
#' @param param_temperature numeric (default: 0.9) specifying the sampling strategy of the possible completions (from the official API documentation: _What sampling temperature to use. Higher values means the model will take more risks. Try 0.9 for more creative applications, and 0 (argmax sampling) for ones with a well-defined answer. We generally recommend altering this or top_p but not both._)
#' @param param_top_p numeric (default: 1) specifying sampling strategy as an alternative to the temperature sampling (from the official API documentation: _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._)
#' @param param_n numeric (default: 1) specifying the number of completions per request (from the official API documentation: _How many completions to generate for each prompt. **Note: Because this parameter generates many completions, it can quickly consume your token quota.** Use carefully and ensure that you have reasonable settings for max_tokens and stop._)
#' @param param_logprobs numeric (default: NULL) (from the official API documentation: _Include the log probabilities on the logprobs most likely tokens, as well the chosen tokens. For example, if logprobs is 5, the API will return a list of the 5 most likely tokens. The API will always return the logprob of the sampled token, so there may be up to logprobs+1 elements in the response. The maximum value for logprobs is 5. If you need more than this, please go to [https://help.openai.com/en/](https://help.openai.com/en/) and describe your use case._)
#' @param param_stop character or character vector (default: NULL) that specifies after which character value when the completion should end (from the official API documentation: _Up to 4 sequences where the API will stop generating further tokens. The returned text will not contain the stop sequence._)
#' @param param_presence_penalty numeric (default: 0) between -2.00 and +2.00 to determine the penalisation of repetitiveness if a token already exists (from the official API documentation: _Number between -2.0 and 2.0. Positive values penalize new tokens based on whether they appear in the text so far, increasing the model's likelihood to talk about new topics._). See also: [https://beta.openai.com/docs/api-reference/parameter-details](https://beta.openai.com/docs/api-reference/parameter-details)
#' @param param_frequency_penalty numeric (default: 0) between -2.00 and +2.00 to determine the penalisation of repetitiveness based on the frequency of a token in the text already (from the official API documentation: _Number between -2.0 and 2.0. Positive values penalize new tokens based on their existing frequency in the text so far, decreasing the model's likelihood to repeat the same line verbatim._). See also: [https://beta.openai.com/docs/api-reference/parameter-details](https://beta.openai.com/docs/api-reference/parameter-details)
#' @param param_best_of numeric (default: 1) that determines the space of possibilities from which to select the completion with the highest probability (from the official API documentation: _Generates `best_of` completions server-side and returns the "best" (the one with the highest log probability per token)_). See details.
#'
#' @return A list with two data tables (if `param_output_type` is the default "complete"): [[1]] contains the data table with the columns `n` (= the mo. of `n` responses requested), `prompt` (= the prompt that was sent), `gpt3` (= the completion as returned from the GPT-3 model) and `id` (= the provided `id_var` or its default alternative). [[2]] contains the meta information of the request, including the request id, the parameters of the request and the token usage of the prompt (`tok_usage_prompt`), the completion (`tok_usage_completion`), the total usage (`tok_usage_total`), and the `id` (= the provided `id_var` or its default alternative).
#'
#' If `output_type` is "text", only the data table in slot [[1]] is returned.
#'
#' If `output_type` is "meta", only the data table in slot [[2]] is returned.
#' @examples
#' # First authenticate with your API key via `gpt3_authenticate('pathtokey')`
#'
#' # Once authenticated:
#' # Assuming you have a data.table with 3 different prompts:
#' dt_prompts = data.table::data.table('prompts' = c('What is the meaning if life?', 'Write a tweet about London:', 'Write a research proposal for using AI to fight fake news:'), 'prompt_id' = c(LETTERS[1:3]))
#'gpt3_completions(prompt_var = dt_prompts$prompts
#' , id_var = dt_prompts$prompt_id)
#'
#' ## With more controls
#'gpt3_completions(prompt_var = dt_prompts$prompts
#' , id_var = dt_prompts$prompt_id
#' , param_max_tokens = 50
#' , param_temperature = 0.5
#' , param_n = 5)
#'
#' ## Reproducible example (deterministic approach)
#'gpt3_completions(prompt_var = dt_prompts$prompts
#' , id_var = dt_prompts$prompt_id
#' , param_max_tokens = 50
#' , param_temperature = 0.0)
#'
#' ## Changing the GPT-3 model
#'gpt3_completions(prompt_var = dt_prompts$prompts
#' , id_var = dt_prompts$prompt_id
#' , param_model = 'text-babbage-001'
#' , param_max_tokens = 50
#' , param_temperature = 0.4)
#' @export
gpt3_completions = function(prompt_var
, id_var
, param_output_type = 'complete'
, param_model = 'text-davinci-003'
, param_suffix = NULL
, param_max_tokens = 100
, param_temperature = 0.9
, param_top_p = 1
, param_n = 1
, param_logprobs = NULL
, param_stop = NULL
, param_presence_penalty = 0
, param_frequency_penalty = 0
, param_best_of = 1){
data_length = length(prompt_var)
if(missing(id_var)){
data_id = paste0('prompt_', 1:data_length)
} else {
data_id = id_var
}
empty_list = list()
meta_list = list()
for(i in 1:data_length){
print(paste0('Request: ', i, '/', data_length))
row_outcome = gpt3_single_completion(prompt_input = prompt_var[i]
, model = param_model
, output_type = 'complete'
, suffix = param_suffix
, max_tokens = param_max_tokens
, temperature = param_temperature
, top_p = param_top_p
, n = param_n
, logprobs = param_logprobs
, stop = param_stop
, presence_penalty = param_presence_penalty
, frequency_penalty = param_frequency_penalty
, best_of = param_best_of)
row_outcome[[1]]$id = data_id[i]
row_outcome[[2]]$id = data_id[i]
empty_list[[i]] = row_outcome[[1]]
meta_list[[i]] = row_outcome[[2]]
}
bunch_core_output = try(data.table::rbindlist(empty_list), silent = T)
if("try-error" %in% class(bunch_core_output)){
bunch_core_output = data.table::rbindlist(empty_list, fill = T)
}
bunch_meta_output = try(data.table::rbindlist(meta_list), silent = T)
if("try-error" %in% class(bunch_meta_output)){
bunch_meta_output = data.table::rbindlist(meta_list, fill = T)
}
if(param_output_type == 'complete'){
output = list(bunch_core_output
, bunch_meta_output)
} else if(param_output_type == 'meta'){
output = bunch_meta_output
} else if(param_output_type == 'text'){
output = bunch_core_output
}
return(output)
}