assistants.R
In samterfa/openai: Consumes the Open AI API
library(tidyverse)
library(openai)
list_tools <- function(){
tibble::tibble(
type = c('code_interpreter', 'function', 'retrieval')
)
}
run_code_screen_capture <- function(code_to_run, take_screenshot = TRUE){
if(take_screenshot) tmpfile <- tempfile(fileext = '.png')
#
# p <-
# promises::future_promise(
# expr = {
# Sys.sleep(4);
# screenshot::screenshot(file = tmpfile)
# })
# Clear Screen and run the code
rstudioapi::sendToConsole('cat("\\014")', execute = TRUE, focus = TRUE)
rstudioapi::sendToConsole(code_to_run, execute = TRUE, focus = TRUE)
if(take_screenshot){
Sys.sleep(2)
screenshot::screenshot(file = tmpfile)
return(tmpfile)
}else{
return(NULL)
}
#$ promises::promise_all(p)
}
create_function_parameters <- function(parameter, type){
}
create_run_code_function <- function(){
'{
"type": "function",
"function": {
"name": "run_r_code",
"description": "Run R code in RStudio",
"parameters": {
"type": "object",
"properties": {
"code": {
"type": "string",
"description": "The code to run in RStudio"
}
},
"required": [
"code"
]
}
}
}' %>%
stringr::str_remove_all("\n") %>%
stringr::str_squish() %>%
jsonlite::fromJSON(F, F, F, F)
}
make_session <- function(name = 'My Assistant', instructions = NULL, description = NULL, model = 'gpt-4-1106-preview', tools = list(list(type = 'retrieval'), list(type = 'code_interpreter'), create_run_code_function()), initial_messages = NULL, file_ids = NULL){
assistant <-
openai::create_assistant(model = model, name = name, description = description, instructions = instructions, tools = tools, file_ids = file_ids)
if(!is.null(file_ids)){
for(file_id in file_ids){
create_assistant_file(assistant_id = assistant$id, file_id = file_id)
}
}
thread <-
openai::create_thread(messages = initial_messages)
tibble::tibble(name = assistant$name, assistant_id = assistant$id, thread_id = thread$id)
}
use_assistant <- function(session = make_session(), new_message = NULL, additional_instructions = NULL, assistant = make_assistant(), thread = make_thread(), file_ids = NULL, role = 'user', timeout_mins = 3){
library(tidyverse)
library(openai)
prev_runs <-
session %>%
select(any_of(c('role', 'content', 'created_at')))
session <-
session %>%
select(-any_of(c('role', 'content', 'created_at'))) %>%
distinct()
assistant_id <- session$assistant_id[[1]]
thread_id <- session$thread_id[[1]]
if(!is.null(new_message)){
potential_error <-
tryCatch({
msg <-
create_message(content = new_message, thread_id = session$thread_id[[1]], role = role, file_ids = file_ids)
print(msg)
run <-
create_run(assistant_id = assistant_id, thread_id = thread_id, instructions = additional_instructions)
run
}, error = function(e) e)
if('error' %in% class(potential_error)){
stop("There was an error!")
runs <-
list_runs(thread_id = session$thread_id)
# Pick up where we left off.
run <-
runs$data %>% discard(~ .x$status %in% 'expired') %>% list_flatten()
}else{
run <- potential_error
}
is_finished_or_timed_out <- run$status == 'completed'
tm <- Sys.time();
while(!is_finished_or_timed_out){
# cat('.')
Sys.sleep(1)
run_progress <-
retrieve_run(run_id = run$id, thread_id = thread_id)
elapsed_mins <- (Sys.time() - tm)/lubridate::dminutes(1)
if(run_progress$status == 'requires_action'){
required_action_details <-
run_progress$required_action$submit_tool_outputs %>%
map_dfr(~ .x %>% list_flatten() %>% as.data.frame() %>% tibble() %>% janitor::clean_names())
code_to_run <-
required_action_details %>% pull(function_arguments) %>% jsonlite::fromJSON() %>% pluck('code')
if(nrow(required_action_details) > 1) stop("Cannot handle multiple required actions")
submit_tool_outputs_to_run(run_id = run_progress$id, thread_id = session$thread_id, tool_outputs = list(list(tool_call_id = required_action_details$id, output = run_r_code(code_to_run))))
}
is_finished_or_timed_out <-
elapsed_mins > timeout_mins |
run_progress$status == 'completed'
}
}else{
run_progress <- list(status = 'retrieving messages only')
}
msgs <-
list_messages(thread_id = thread_id)$data %>%
map_dfr(~ tibble(created_at = .x$created_at %>% lubridate::as_datetime(tz = 'America/Chicago'),
role = .x$role,
content = .x$content %>% keep(~ .x$type == 'text') %>% compact() %>% pluck(1) %>% pluck('text') %>% pluck('value'))) %>%
arrange(created_at)
session <-
bind_cols(session, msgs)
cat(session %>% slice_max(created_at) %>% pull(content))
session
}
use_assistant_screenshot <- function(session = make_session(), new_message = NULL, additional_instructions = NULL, assistant = make_assistant(), thread = make_thread(), file_ids = NULL, role = 'user', timeout_mins = 3){
stop("GPT Vision does not work with assistants. Models cannot see screenshots.")
library(tidyverse)
library(openai)
prev_runs <-
session %>%
select(any_of(c('role', 'content', 'created_at')))
session <-
session %>%
select(-any_of(c('role', 'content', 'created_at'))) %>%
distinct()
assistant_id <- session$assistant_id[[1]]
thread_id <- session$thread_id[[1]]
if(!is.null(new_message)){
potential_error <-
tryCatch({
msg <-
create_message(content = new_message, thread_id = session$thread_id[[1]], role = role, file_ids = file_ids)
print(msg)
run <-
create_run(assistant_id = assistant_id, thread_id = thread_id, instructions = additional_instructions)
run
}, error = function(e) e)
if('error' %in% class(potential_error)){
print(potential_error)
runs <-
list_runs(thread_id = session$thread_id)
# Pick up where we left off.
run <-
runs$data %>% discard(~ .x$status %in% 'expired') %>% list_flatten()
}else{
run <- potential_error
}
is_finished_or_timed_out <- run$status == 'completed'
tm <- Sys.time();
while(!is_finished_or_timed_out){
# cat('.')
Sys.sleep(1)
run_progress <-
retrieve_run(run_id = run$id, thread_id = thread_id)
elapsed_mins <- (Sys.time() - tm)/lubridate::dminutes(1)
if(run_progress$status == 'requires_action'){
required_action_details <-
run_progress$required_action$submit_tool_outputs %>%
map_dfr(~ .x %>% list_flatten() %>% as.data.frame() %>% tibble() %>% janitor::clean_names())
code_to_run <-
required_action_details %>% pull(function_arguments) %>% jsonlite::fromJSON() %>% pluck('code')
if(nrow(required_action_details) > 1) stop("Cannot handle multiple required actions")
tmpfile <- run_code_screen_capture(code_to_run)
file <- upload_file(tmpfile, purpose = 'assistants')
submit_tool_outputs_to_run(run_id = run_progress$id, thread_id = session$thread_id, tool_outputs = list(list(tool_call_id = required_action_details$id, output = glue::glue("The following file contains a screenshot for the output of the code you ran. file_id: {file$id}"))))
}
is_finished_or_timed_out <-
elapsed_mins > timeout_mins |
run_progress$status == 'completed'
}
}else{
run_progress <- list(status = 'retrieving messages only')
}
msgs <-
list_messages(thread_id = thread_id)$data %>%
map_dfr(~ tibble(created_at = .x$created_at %>% lubridate::as_datetime(tz = 'America/Chicago'),
role = .x$role,
content = .x$content %>% keep(~ .x$type == 'text') %>% compact() %>% pluck(1) %>% pluck('text') %>% pluck('value'))) %>%
arrange(created_at)
session <-
bind_cols(session, msgs)
cat(session %>% slice_max(created_at) %>% pull(content))
session
}
vision_prompting <- function(iter = 5, initial_task = 'Recreate and iteratively improve this plot until it is publication ready.', final_task = '', img_path = 'sample_plot.png'){
model <- 'gpt-4-vision-preview'
if(!is.null(img_path)){
img <-
img_path %>%
base64enc::base64encode()
messages <-
list(
list(
role = 'system',
content = 'You are an R code generator. The user will run your code in an RStudio session and return to you a screenshot showing you the results of running your code. There is limited space in the screenshot so only return a little code at a time! Only return R code.'
),
list(
role = 'user',
content = list(
list(
type = 'text',
text = initial_task
),
list(
type = 'image_url',
image_url = list(
url = glue::glue("data:image/jpeg;base64,{img}")
)
)
)
)
)
}else{
messages <-
list(
list(
role = 'system',
content = 'You are an R code generator. The user will run your code in an RStudio session and return to you a screenshot showing you the results of running your code. Screen space is limited for each screenshot; only try one thing at a time so that you can see the full output. Only return R code.'
),
list(
role = 'user',
content = list(
list(
type = 'text',
text = initial_task
)
)
)
)
}
for(i in 1:iter){
completion <-
create_chat_completion(messages = messages, model = model, max_tokens = 1000)
return_message <-
completion$choices[[1]]$message
messages <-
messages %>%
append(list(return_message))
code_to_run <-
return_message$content %>%
extract_r_code()
if(is.na(code_to_run)) break()
tmpfile <-
run_code_screen_capture(code_to_run)
new_message <-
list(
role = 'user',
content = list(
list(
type = 'text',
text = 'Here is a screenshot of your RStudio session.'
),
list(
type = 'image_url',
image_url = list(
url = glue::glue("data:image/jpeg;base64,{base64enc::base64encode(tmpfile)}")
)
)
)
)
messages <-
messages %>%
append(list(new_message))
}
new_message <-
list(
role = 'user',
content = list(
list(
type = 'text',
text = final_task
)
)
)
messages <-
messages %>%
append(list(new_message))
completion <-
create_chat_completion(messages = messages, model = model, max_tokens = 1000)
return_message <-
completion$choices[[1]]$message
messages <-
messages %>%
append(list(return_message))
code_to_run <-
return_message$content %>%
extract_r_code()
if(!is.na(code_to_run)){
run_code_screen_capture(code_to_run, take_screenshot = FALSE)
}
readr::write_rds(messages, '~/Desktop/openai_testing/messages.rds')
messages
}
extract_r_code <- function(txt){
txt %>%
stringr::str_replace_all("\n", "asdfdareasrads") %>%
stringr::str_extract("(?<=```).*(?=```)") %>%
stringr::str_remove("^[rR]") %>%
stringr::str_replace_all(stringr::fixed("asdfdareasrads"), "\n")
}
extract_json <- function(txt){
txt %>%
stringr::str_replace_all("\n", "asdfdareasrads") %>%
stringr::str_extract("(?<=```).*(?=```)") %>%
stringr::str_remove("^json|^JSON") %>%
stringr::str_replace_all(stringr::fixed("asdfdareasrads"), "\n")
}
run_r_code <- function(code_to_run){
suppressMessages(
suppressWarnings(
reprex::reprex(input = paste0("suppressMessages({x <- readr::read_csv('~/Desktop/data.csv')}); ", code_to_run %>% str_replace_all("\n", "; "), "\n"),
session_info = FALSE,
advertise = FALSE,
tidyverse_quiet = TRUE,
html_preview = FALSE) %>%
paste(collapse = '\n') %>%
stringr::str_remove(stringr::fixed("suppressMessages({x <- readr::read_csv('~/Desktop/data.csv')}); "))
)
)
}
###### vision_prompting(iter = 10, initial_task = 'There is a dataset x in memory. Iteratively explore the dataset using tidyverse functions. After some exploration, return 3 significant, meaningful, complex insights from the dataset.', final_task = 'Summarize the most significant insights you found including helpful plots in a file called "insights.Rmd".')
samterfa/openai documentation built on Oct. 9, 2024, 5:18 p.m.
R Package Documentation
Browse R Packages
We want your feedback!
Note that we can't provide technical support on individual packages. You should contact the package authors for that.
library(tidyverse)
library(openai)
list_tools <- function(){
tibble::tibble(
type = c('code_interpreter', 'function', 'retrieval')
)
}
run_code_screen_capture <- function(code_to_run, take_screenshot = TRUE){
if(take_screenshot) tmpfile <- tempfile(fileext = '.png')
#
# p <-
# promises::future_promise(
# expr = {
# Sys.sleep(4);
# screenshot::screenshot(file = tmpfile)
# })
# Clear Screen and run the code
rstudioapi::sendToConsole('cat("\\014")', execute = TRUE, focus = TRUE)
rstudioapi::sendToConsole(code_to_run, execute = TRUE, focus = TRUE)
if(take_screenshot){
Sys.sleep(2)
screenshot::screenshot(file = tmpfile)
return(tmpfile)
}else{
return(NULL)
}
#$ promises::promise_all(p)
}
create_function_parameters <- function(parameter, type){
}
create_run_code_function <- function(){
'{
"type": "function",
"function": {
"name": "run_r_code",
"description": "Run R code in RStudio",
"parameters": {
"type": "object",
"properties": {
"code": {
"type": "string",
"description": "The code to run in RStudio"
}
},
"required": [
"code"
]
}
}
}' %>%
stringr::str_remove_all("\n") %>%
stringr::str_squish() %>%
jsonlite::fromJSON(F, F, F, F)
}
make_session <- function(name = 'My Assistant', instructions = NULL, description = NULL, model = 'gpt-4-1106-preview', tools = list(list(type = 'retrieval'), list(type = 'code_interpreter'), create_run_code_function()), initial_messages = NULL, file_ids = NULL){
assistant <-
openai::create_assistant(model = model, name = name, description = description, instructions = instructions, tools = tools, file_ids = file_ids)
if(!is.null(file_ids)){
for(file_id in file_ids){
create_assistant_file(assistant_id = assistant$id, file_id = file_id)
}
}
thread <-
openai::create_thread(messages = initial_messages)
tibble::tibble(name = assistant$name, assistant_id = assistant$id, thread_id = thread$id)
}
use_assistant <- function(session = make_session(), new_message = NULL, additional_instructions = NULL, assistant = make_assistant(), thread = make_thread(), file_ids = NULL, role = 'user', timeout_mins = 3){
library(tidyverse)
library(openai)
prev_runs <-
session %>%
select(any_of(c('role', 'content', 'created_at')))
session <-
session %>%
select(-any_of(c('role', 'content', 'created_at'))) %>%
distinct()
assistant_id <- session$assistant_id[[1]]
thread_id <- session$thread_id[[1]]
if(!is.null(new_message)){
potential_error <-
tryCatch({
msg <-
create_message(content = new_message, thread_id = session$thread_id[[1]], role = role, file_ids = file_ids)
print(msg)
run <-
create_run(assistant_id = assistant_id, thread_id = thread_id, instructions = additional_instructions)
run
}, error = function(e) e)
if('error' %in% class(potential_error)){
stop("There was an error!")
runs <-
list_runs(thread_id = session$thread_id)
# Pick up where we left off.
run <-
runs$data %>% discard(~ .x$status %in% 'expired') %>% list_flatten()
}else{
run <- potential_error
}
is_finished_or_timed_out <- run$status == 'completed'
tm <- Sys.time();
while(!is_finished_or_timed_out){
# cat('.')
Sys.sleep(1)
run_progress <-
retrieve_run(run_id = run$id, thread_id = thread_id)
elapsed_mins <- (Sys.time() - tm)/lubridate::dminutes(1)
if(run_progress$status == 'requires_action'){
required_action_details <-
run_progress$required_action$submit_tool_outputs %>%
map_dfr(~ .x %>% list_flatten() %>% as.data.frame() %>% tibble() %>% janitor::clean_names())
code_to_run <-
required_action_details %>% pull(function_arguments) %>% jsonlite::fromJSON() %>% pluck('code')
if(nrow(required_action_details) > 1) stop("Cannot handle multiple required actions")
submit_tool_outputs_to_run(run_id = run_progress$id, thread_id = session$thread_id, tool_outputs = list(list(tool_call_id = required_action_details$id, output = run_r_code(code_to_run))))
}
is_finished_or_timed_out <-
elapsed_mins > timeout_mins |
run_progress$status == 'completed'
}
}else{
run_progress <- list(status = 'retrieving messages only')
}
msgs <-
list_messages(thread_id = thread_id)$data %>%
map_dfr(~ tibble(created_at = .x$created_at %>% lubridate::as_datetime(tz = 'America/Chicago'),
role = .x$role,
content = .x$content %>% keep(~ .x$type == 'text') %>% compact() %>% pluck(1) %>% pluck('text') %>% pluck('value'))) %>%
arrange(created_at)
session <-
bind_cols(session, msgs)
cat(session %>% slice_max(created_at) %>% pull(content))
session
}
use_assistant_screenshot <- function(session = make_session(), new_message = NULL, additional_instructions = NULL, assistant = make_assistant(), thread = make_thread(), file_ids = NULL, role = 'user', timeout_mins = 3){
stop("GPT Vision does not work with assistants. Models cannot see screenshots.")
library(tidyverse)
library(openai)
prev_runs <-
session %>%
select(any_of(c('role', 'content', 'created_at')))
session <-
session %>%
select(-any_of(c('role', 'content', 'created_at'))) %>%
distinct()
assistant_id <- session$assistant_id[[1]]
thread_id <- session$thread_id[[1]]
if(!is.null(new_message)){
potential_error <-
tryCatch({
msg <-
create_message(content = new_message, thread_id = session$thread_id[[1]], role = role, file_ids = file_ids)
print(msg)
run <-
create_run(assistant_id = assistant_id, thread_id = thread_id, instructions = additional_instructions)
run
}, error = function(e) e)
if('error' %in% class(potential_error)){
print(potential_error)
runs <-
list_runs(thread_id = session$thread_id)
# Pick up where we left off.
run <-
runs$data %>% discard(~ .x$status %in% 'expired') %>% list_flatten()
}else{
run <- potential_error
}
is_finished_or_timed_out <- run$status == 'completed'
tm <- Sys.time();
while(!is_finished_or_timed_out){
# cat('.')
Sys.sleep(1)
run_progress <-
retrieve_run(run_id = run$id, thread_id = thread_id)
elapsed_mins <- (Sys.time() - tm)/lubridate::dminutes(1)
if(run_progress$status == 'requires_action'){
required_action_details <-
run_progress$required_action$submit_tool_outputs %>%
map_dfr(~ .x %>% list_flatten() %>% as.data.frame() %>% tibble() %>% janitor::clean_names())
code_to_run <-
required_action_details %>% pull(function_arguments) %>% jsonlite::fromJSON() %>% pluck('code')
if(nrow(required_action_details) > 1) stop("Cannot handle multiple required actions")
tmpfile <- run_code_screen_capture(code_to_run)
file <- upload_file(tmpfile, purpose = 'assistants')
submit_tool_outputs_to_run(run_id = run_progress$id, thread_id = session$thread_id, tool_outputs = list(list(tool_call_id = required_action_details$id, output = glue::glue("The following file contains a screenshot for the output of the code you ran. file_id: {file$id}"))))
}
is_finished_or_timed_out <-
elapsed_mins > timeout_mins |
run_progress$status == 'completed'
}
}else{
run_progress <- list(status = 'retrieving messages only')
}
msgs <-
list_messages(thread_id = thread_id)$data %>%
map_dfr(~ tibble(created_at = .x$created_at %>% lubridate::as_datetime(tz = 'America/Chicago'),
role = .x$role,
content = .x$content %>% keep(~ .x$type == 'text') %>% compact() %>% pluck(1) %>% pluck('text') %>% pluck('value'))) %>%
arrange(created_at)
session <-
bind_cols(session, msgs)
cat(session %>% slice_max(created_at) %>% pull(content))
session
}
vision_prompting <- function(iter = 5, initial_task = 'Recreate and iteratively improve this plot until it is publication ready.', final_task = '', img_path = 'sample_plot.png'){
model <- 'gpt-4-vision-preview'
if(!is.null(img_path)){
img <-
img_path %>%
base64enc::base64encode()
messages <-
list(
list(
role = 'system',
content = 'You are an R code generator. The user will run your code in an RStudio session and return to you a screenshot showing you the results of running your code. There is limited space in the screenshot so only return a little code at a time! Only return R code.'
),
list(
role = 'user',
content = list(
list(
type = 'text',
text = initial_task
),
list(
type = 'image_url',
image_url = list(
url = glue::glue("data:image/jpeg;base64,{img}")
)
)
)
)
)
}else{
messages <-
list(
list(
role = 'system',
content = 'You are an R code generator. The user will run your code in an RStudio session and return to you a screenshot showing you the results of running your code. Screen space is limited for each screenshot; only try one thing at a time so that you can see the full output. Only return R code.'
),
list(
role = 'user',
content = list(
list(
type = 'text',
text = initial_task
)
)
)
)
}
for(i in 1:iter){
completion <-
create_chat_completion(messages = messages, model = model, max_tokens = 1000)
return_message <-
completion$choices[[1]]$message
messages <-
messages %>%
append(list(return_message))
code_to_run <-
return_message$content %>%
extract_r_code()
if(is.na(code_to_run)) break()
tmpfile <-
run_code_screen_capture(code_to_run)
new_message <-
list(
role = 'user',
content = list(
list(
type = 'text',
text = 'Here is a screenshot of your RStudio session.'
),
list(
type = 'image_url',
image_url = list(
url = glue::glue("data:image/jpeg;base64,{base64enc::base64encode(tmpfile)}")
)
)
)
)
messages <-
messages %>%
append(list(new_message))
}
new_message <-
list(
role = 'user',
content = list(
list(
type = 'text',
text = final_task
)
)
)
messages <-
messages %>%
append(list(new_message))
completion <-
create_chat_completion(messages = messages, model = model, max_tokens = 1000)
return_message <-
completion$choices[[1]]$message
messages <-
messages %>%
append(list(return_message))
code_to_run <-
return_message$content %>%
extract_r_code()
if(!is.na(code_to_run)){
run_code_screen_capture(code_to_run, take_screenshot = FALSE)
}
readr::write_rds(messages, '~/Desktop/openai_testing/messages.rds')
messages
}
extract_r_code <- function(txt){
txt %>%
stringr::str_replace_all("\n", "asdfdareasrads") %>%
stringr::str_extract("(?<=```).*(?=```)") %>%
stringr::str_remove("^[rR]") %>%
stringr::str_replace_all(stringr::fixed("asdfdareasrads"), "\n")
}
extract_json <- function(txt){
txt %>%
stringr::str_replace_all("\n", "asdfdareasrads") %>%
stringr::str_extract("(?<=```).*(?=```)") %>%
stringr::str_remove("^json|^JSON") %>%
stringr::str_replace_all(stringr::fixed("asdfdareasrads"), "\n")
}
run_r_code <- function(code_to_run){
suppressMessages(
suppressWarnings(
reprex::reprex(input = paste0("suppressMessages({x <- readr::read_csv('~/Desktop/data.csv')}); ", code_to_run %>% str_replace_all("\n", "; "), "\n"),
session_info = FALSE,
advertise = FALSE,
tidyverse_quiet = TRUE,
html_preview = FALSE) %>%
paste(collapse = '\n') %>%
stringr::str_remove(stringr::fixed("suppressMessages({x <- readr::read_csv('~/Desktop/data.csv')}); "))
)
)
}
###### vision_prompting(iter = 10, initial_task = 'There is a dataset x in memory. Iteratively explore the dataset using tidyverse functions. After some exploration, return 3 significant, meaningful, complex insights from the dataset.', final_task = 'Summarize the most significant insights you found including helpful plots in a file called "insights.Rmd".')
R Package Documentation
Browse R Packages
We want your feedback!
Note that we can't provide technical support on individual packages. You should contact the package authors for that.
Embedding an R snippet on your website
Add the following code to your website.
For more information on customizing the embed code, read Embedding Snippets.