Remove models.R and assist.R #10
Description
Remove these two files since we transferred to use ellmer now in the back end. Keeping a record of the two files here for now.
models.R
#' Create a list of model and vendor
#'
#' @param model The name of the model
#' @param vendor The name of the vendor
#' @family large-lanugage-models
#' @export
model_vendor <- function(model, vendor, ...) {
get(paste0("model_", vendor))(model, ...)
}
#' Set the Ollama model
#'
#' @param port The port number of the Ollama server.
#' @param format The format to use. Default is "none" and currently only other option is "json".
#' @param temperature The temperature of the model where higher temperature means higher creativity.
#' @param stop A list of stop sequences to use.
#' @param seed The seed to use for the model.
#' @param top_k The top k tokens to consider.
#' @param top_p The nucleus sampling or penalty. It limits the cumulative probability of the most likely tokens.
#' Higher values allow more tokens and diverse responses, and while lower values are more focused and constrained answers.
#' @param num_predict The number of predictions to make.
#' @param keep_alive The time to keep the server alive.
#' @family large-lanugage-models
#' @export
model_ollama <- function(model = 'llama3.1:8b',
port = 11434,
format = c("none", "json"),
tools = NULL,
temperature = 0,
stop = NA,
seed = 0,
top_k = 1,
top_p = 0.1,
num_predict = 1000,
keep_alive = "5m",
.check_model_availability = TRUE) {
format <- match.arg(format)
if(.check_model_availability) {
available_model_list <- ollama_model_list()
# https://github.com/ollama/ollama#model-library
abort_if_model_not_available(model, "ollama", available_model_list)
}
list(model = model, vendor = "ollama",
args = list(port = port,
format = format,
tools = tools,
temperature = temperature,
stop = stop,
seed = seed,
top_k = top_k,
top_p = top_p,
num_predict = num_predict,
keep_alive = keep_alive))
}
#' Set an OpenAI model
#'
#' @param api The API key to use.
#' @param format The format
#' @param max_tokens The maximum number of tokens to generate. This limits the length of the response.
#' Tokens refers to a unit of text the model reads and can vary from one character to several words,
#' varying based on the model. As a rule of thumb 1 token is approximately 4 characters or 0.75 words
#' for English text.
#' @param n The number of completions/responses to generate.
#' @param frequency_penalty The frequency penalty to use. It discourages the model from repeating the same text.
#' A lower value results in the model more likely to repeat information.
#' @param presence_penalty Avoidance of specific topics in response provided in the user messages.
#' A lower value make the model less concerned about preventing these topics.
#' @inheritParams model_ollama
#' @family large-lanugage-models
#' @export
model_openai <- function(model = "gpt-4o-mini",
api = NULL,
seed = NA,
format = "none",
max_tokens = 100,
temperature = 0,
top_p = 0.1,
n = 1,
stop = NULL,
presence_penalty = 0.2,
frequency_penalty = 1) {
available_model_list <- openai_model_list()
abort_if_model_not_available(model, 'openai', available_model_list)
list(model = model, vendor = "openai",
args = list(api = api,
seed = seed,
format = format,
max_tokens = max_tokens,
temperature = temperature,
top_p = top_p,
n = n,
stop = stop,
presence_penalty = presence_penalty,
frequency_penalty = frequency_penalty))
}
#' Set a model from Mistral AI
#'
#' @param min_tokens The minimum number of tokens to generate.
#' @param safe_prompt Whether to use safe prompt or not.
#' @param tools The list of tools. Not implemented.
#' @param tool_choice The choice of tool. Not implemented.
#' @inheritParams model_openai
#' @family large-lanugage-models
#' @export
model_mistral <- function(model = "open-mistral-7b",
temperature = 0,
top_p = 0.1,
max_tokens = 100,
min_tokens = 0,
stop = NULL,
seed = NA,
format = "none",
safe_prompt = FALSE,
tools = NULL,
tool_choice = "auto") {
abort_if_model_not_available(model, 'mistral', mistral_model_list())
list(model = model, vendor = "mistral",
args = list(temperature = temperature,
top_p = top_p,
max_tokens = max_tokens,
min_tokens = min_tokens,
format = format,
stop = stop,
random_seed = seed,
tools = tools,
stream = FALSE,
safe_prompt = safe_prompt,
tool_choice = tool_choice,
response_format = switch(format,
none = list(type = "text"),
json = list(type = "json_object"))))
}
#' Get the list of ollama models available locally
#'
#' @param port The port number.
#' @export
ollama_model_list <- function(port = emend_get_option("model")$args$port) {
request <- httr2::request(base_url = paste0('http://localhost:', port, '/api/tags'))
response <- httr2::req_perform(request)
json <- httr2::resp_body_json(response)
model_names <- map_chr(json$models, function(x) x$name)
gsub(":latest$", "", model_names)
}
#' Get the list of Mistral AI models
#'
#' @param api The Mistral AI API key.
#' @export
mistral_model_list <- function(api = NULL) {
api <- api %||% Sys.getenv("MISTRAL_API_KEY")
request <- httr2::request(base_url = 'https://api.mistral.ai/v1/models')
requesth <- httr2::req_headers(request,
'Authorization' = paste('Bearer', api))
response <- httr2::req_perform(requesth)
json <- httr2::resp_body_json(response)
unique(map_chr(json$data, function(x) x$name))
}
#' Get the list of OpenAI models
#'
#' @param api The OpenAI API key. Provide "none" to get the list without an API key.
#' @export
openai_model_list <- function(api = NULL) {
# Last extracted 2024-10-05
models <- c("gpt-4o-realtime-preview-2024-10-01", "chatgpt-4o-latest",
"dall-e-2", "tts-1", "tts-1-1106", "gpt-4-0125-preview", "gpt-3.5-turbo-0125",
"gpt-4o-2024-05-13", "gpt-4-turbo-preview", "gpt-3.5-turbo",
"whisper-1", "gpt-4-turbo", "gpt-3.5-turbo-16k", "text-embedding-3-small",
"gpt-3.5-turbo-1106", "tts-1-hd", "tts-1-hd-1106", "gpt-3.5-turbo-instruct-0914",
"gpt-4-0613", "gpt-4o-2024-08-06", "gpt-4", "gpt-3.5-turbo-instruct",
"gpt-4-1106-preview", "babbage-002", "davinci-002", "dall-e-3",
"gpt-4o", "gpt-4-turbo-2024-04-09", "text-embedding-ada-002",
"gpt-4o-mini", "gpt-4o-realtime-preview", "text-embedding-3-large",
"gpt-4o-mini-2024-07-18")
if(!is.null(api) && api == "none") {
cli::cli_alert("The list of OpenAI models below was last extracted in 2024-10-05 so may not be up-to-date.")
return(models)
}
api <- api %||% Sys.getenv("OPENAI_API_KEY")
request <- httr2::request(base_url = 'https://api.openai.com/v1/models')
requesth <- httr2::req_headers(request,
'Authorization' = paste('Bearer', api))
response <- httr2::req_perform(requesth)
json <- httr2::resp_body_json(response)
map_chr(json$data, function(x) x$id)
}
assist.R
#' emend the general assistant
#'
#' @param prompt A list of prompts or text with a prompt.
#' @param model A list of model and vendor name. Use model helpers like `model_ollama()` and `model_openai()`.
#' @export
emend_assist <- function(prompt = NULL,
model = emend_get_option("model"),
format = model$args$format,
.content_envir = rlang::caller_env(),
...) {
if(is.null(prompt)) cli::cli_alert("To get started, provide a prompt.")
if(is.list(prompt) & !is_prompt(prompt)) prompt <- lapply(prompt, function(x) as_prompt(x, .content_envir = .content_envir))
if(is.character(prompt)) prompt <- lapply(prompt, function(x) prompt_user(x, .content_envir = .content_envir))
if(is_prompt(prompt)) prompt <-list(prompt)
msgs <- list(messages = c(prompt,
if(format == "json") list(prompt_user("Respond using JSON."))))
# https://github.com/ollama/ollama/blob/main/docs/api.md#request-reproducible-outputs
# https://github.com/ollama/ollama/blob/main/docs/modelfile.md#valid-parameters-and-values
dots <- list(...)
if(model$vendor == "ollama") {
params <- c(list(model = model$model,
format = ifelse(format == "none", NA, format),
stop = model$args$stop,
keep_alive = model$args$keep_alive,
tools = model$args$tools,
options = list(if(!is.na(model$args$seed)) seed = model$args$seed,
temperature = model$args$temperature,
num_predict = model$args$num_predict,
top_k = model$args$top_k,
top_p = model$args$top_p,
dots),
raw = TRUE,
stream = FALSE))
request <- httr2::request(base_url = paste0('http://localhost:', model$args$port,'/api/chat'))
request_with_data <- httr2::req_body_json(request, data = c(params, msgs))
response <- httr2::req_perform(request_with_data)
json <- httr2::resp_body_json(response)
content <- json$message$content
if(format == "json") {
tryCatch({
content <- jsonlite::fromJSON(content)
}, error = function(e) {
cli::cli_alert("The response is not in JSON format.")
})
}
} else if(model$vendor == "openai") {
api <- model$args$api %||% Sys.getenv("OPENAI_API_KEY")
if(api == "") cli::cli_abort("Please set the OPENAI_API_KEY environment variable.")
request <- httr2::request(base_url = "https://api.openai.com/v1/chat/completions")
requesth <- httr2::req_headers(request,
'Authorization' = paste('Bearer', api),
'Content-Type' = 'application/json')
args_parse <- model$args
args_parse$format <- NULL
args_parse$api <- NULL
args_parse$seed <- NULL
params <- c(list(model = model$model), args_parse, dots)
request_with_data <- httr2::req_body_json(requesth, data = c(params, msgs))
response <- httr2::req_perform(request_with_data)
json <- httr2::resp_body_json(response)
content <- json$choices[[1]]$message$content
} else if(model$vendor == "mistral") {
api <- model$args$api %||% Sys.getenv("MISTRAL_API_KEY")
if(api == "") cli::cli_abort("Please set the MISTRAL_API_KEY environment variable.")
request <- httr2::request(base_url = "https://api.mistral.ai/v1/chat/completions")
requesth <- httr2::req_headers(request,
'Authorization' = paste('Bearer', api),
'Content-Type' = 'application/json',
'Accept' = 'application/json')
args_parse <- model$args
args_parse$format <- NULL
args_parse$api <- NULL
args_parse$tool_choice <- NULL
args_parse$tools <- NULL
params <- c(list(model = model$model), args_parse, dots)
request_with_data <- httr2::req_body_json(requesth, data = c(params, msgs))
response <- httr2::req_perform(request_with_data)
json <- httr2::resp_body_json(response)
content <- json$choices[[1]]$message$content
} else {
cli::cli_abort("Currently the only vendor supported is Ollama and OpenAI.")
}
structure(content, class = c("emend", class(content)))
}
#' @export
print.emend <- function(x, ...) {
if(is.list(x)) {
print(unclass(x))
} else if(is.character(x)) {
cat(x)
}
x
}
#' Answer with yes or no
#'
#' @param prompt A prompt.
#' @export
emend_yes_no <- function(prompt) {
if(is.character(prompt)) prompt <- list(prompt_user(prompt))
out <- emend_assist(c(prompt, list(prompt_user("Answer with 'yes' or 'no'."))))
grepl("yes", tolower(out))
}