Return code generated by the assistant to evaluate it in a shiny app?

[useReconomist]

How could I return the code generated by the assistant and evaluate it in a shiny application for a plotoutput and textoutput?

I'm trying to return the code generated by assistant inside shinyapp so that I can view the plot and code but I have not been successful, I would appreciate any observations.

library(shiny)
library(shinychat)
library(tidyverse)
library(elmer)

prompt_chat = "You are an expert in data manipulation and visualization in the R programming language and the use of tidyverse libraries and with a lot of experience in ggplot. 
Return only and exclusively the R code that answers the question you are asked without giving any other explanation, 
the database you have at your disposal is called df_base, the ggplot library is already in the environment so it is not necessary to call it."

df_base = as_tibble(iris)

ui <- fluidPage(
  theme = bs_theme(version = 5),
  chat_ui("chat"),
  textOutput("code_ia"),
  plotOutput("plot_ia")
)

server <- function(input, output, session) {
  chat <- chat_github( model = "gpt-4o-mini",
                       system_prompt = prompt_chat,
                       api_key = token_github,
                       echo = TRUE
  )
  
  observeEvent(input$chat_user_input, {
    # Add user message to the chat history
    chat_append("chat", chat$stream_async(input$chat_user_input))
  })
  
  output$code_ia = renderText({
  })
  
  output$plot_ia = renderPlot({
    
  })
  
}

shinyApp(ui, server)

[jcheng5]

Since you're not asking for any explanation, only code, you wouldn't call chat_append, or maybe not even chat$stream_async(); instead, you'd call chat$chat() (or chat$chat_async() if it needs to be scalable).

The main trick is to set the response as a reactiveVal (code in this case) and then have the rest of the app access that code.

I had to make some other changes to make the model give sensible results (like giving it access to str(df_base)).

Finally, this is not something you'd want to deploy--malicious users would probably have an easy time getting this to generate harmful code and execute it on your server.

library(shiny)
library(bslib)
library(shinychat)
library(tidyverse)
library(elmer)

df_base = as_tibble(iris)

prompt_chat = "You are an expert in data manipulation and visualization in the R programming language and the use of tidyverse libraries and with a lot of experience in ggplot. 
Return only and exclusively the R code that answers the question you are asked without giving any other explanation, 
the database you have at your disposal is called df_base, the ggplot library is already in the environment so it is not necessary to call it.
Do not include any R Markdown formatting in your answer, only pure R code.

The df_base dataframe has the following structure:\n\n" |>
  paste0(paste(collapse = "\n", capture.output(str(df_base))))

ui <- page_sidebar(
  sidebar = sidebar(
    chat_ui("chat"),
    width = "50%"
  ),
  card(
    card_title("Plot"),
    plotOutput("plot_ia")
  ),
  card(
    card_title("Code"),
    verbatimTextOutput("code_ia"),
    fill = FALSE
  ),
)

server <- function(input, output, session) {
  code <- reactiveVal(NULL)

  chatbot_eval_env <- new.env(parent = globalenv())
  chatbot_eval_env$df_base <- df_base

  chat <- chat_github( model = "gpt-4o-mini",
                       system_prompt = prompt_chat,
                       api_key = token_github,
                       echo = TRUE
  )
  
  observeEvent(input$chat_user_input, {
    # Add user message to the chat history
    response <- chat$chat(input$chat_user_input)
    code(response)
    chat_append("chat", paste0("```r\n", response, "\n```"))
  })
  
  output$code_ia = renderText({
    req(code())
  })
  
  output$plot_ia = renderPlot({
    req(code())
    parsed <- parse(text = code())
    eval(parsed, envir = chatbot_eval_env)
  })
}

shinyApp(ui, server)

[useReconomist]

Thank you so much for the modifications and suggestions. It worked amazing!

[zhaohongxin0]

In this case, response <- chat$chat(input$chat_user_input) works fine.
However, response <- chat$stream_async(input$chat_user_input) does not work.

Is there any mthod to obtain the code , and still keep the chat as stream?