20 questions game using several different LLMs

.env

OPENAI_API_KEY=<Paste API key here>
ANTHROPIC_API_KEY=<Paste API key here>

app.R

library(shiny)
library(shinychat)
library(ellmer)
library(bslib)
library(dotenv)
library(promises)

# Load environment variables for API keys
load_dot_env()

# Define UI
ui <- page_fluid(
  card(
    card_header(
      h2("20 Questions AI Game", class = "text-center"),
    ),

    # Chat interface
    chat_ui("chat"),
    div(
      style = "font-size: 0.8rem; width:min(680px, 100%); margin: 0 auto; display: flex; justify-content: flex-end;",
      fill = NA,
      checkboxInput("show_details", "Show conversation details", width = "auto")
    )
  ),

  conditionalPanel(
    "input.show_details === true",
    card(
      card_header(
        h2("Details", class = "text-center"),
      ),
      verbatimTextOutput("messages")
    )
  )
)

# Define server logic
server <- function(input, output, session) {
  # System prompt that instructs the AI how to play 20 questions
  system_prompt <- "You are playing the classic 20 Questions game with the user, but with reversed roles. 
  In this game, YOU will think of something and the USER will ask yes/no questions to guess what it is.
  
  Rules:
  1. At the start, secretly choose a common object, animal, person, or place. Don't reveal what it is.
  2. Keep track of how many questions the user has asked (maximum 20).
  3. Start by explaining the game and telling the user you've thought of something.
  4. Answer the user's questions honestly with 'Yes', 'No', or a very brief clarification if needed.
  5. If the user guesses correctly before 20 questions, congratulate them and offer to play again.
  6. If they reach 20 questions without guessing correctly, reveal your answer and offer to play again.
  7. Be friendly, enthusiastic, and make the game fun!
  8. If the user asks to play again or start over, think of a new object.
  9. IMPORTANT: You are part of a team of AIs taking turns answering. Maintain continuity with previous answers.
  10. Read the conversation history carefully to understand what object was chosen and what questions were already asked.
  
  Begin by inviting the user to play the game."

  # Initialize both chat models
  chat_models <- list(
    ellmer::chat_openai(
      system_prompt = system_prompt,
      model = "gpt-4o"
    ),
    ellmer::chat_anthropic(
      system_prompt = system_prompt,
      model = "claude-3-5-sonnet-latest"
    ),
    ellmer::chat_openai(
      system_prompt = system_prompt,
      model = "gpt-4o-mini"
    ),
    ellmer::chat_anthropic(
      system_prompt = system_prompt,
      model = "claude-3-5-haiku-latest"
    )
  )

  chat_idx <- 1

  # Send initial message when the app starts
  chat_models[[chat_idx]]$set_turns(
    list(
      Turn(
        "assistant",
        contents = list(ContentText(
          "Let's play 20 questions! I will think of an object and you ask the questions."
        ))
      )
    )
  )

  # Append that same initial message to the chat UI
  chat_append(
    "chat",
    "Let's play 20 questions! I will think of an object and you ask the questions.",
    role = "assistant"
  )

  # All the messages
  all_messages <- reactiveVal(chat_models[[chat_idx]]$get_turns())

  # Handle user input
  observeEvent(input$chat_user_input, {
    current_chat <- chat_models[[chat_idx]]

    # Grab the chat history from the previously-used LLM chat object
    current_chat$set_turns(all_messages())

    stream <- current_chat$stream_async(input$chat_user_input)

    chat_append("chat", stream)$then(function() {
      all_messages(current_chat$get_turns())
      # Increment the index and wrap around
      chat_idx <<- chat_idx %% length(chat_models) + 1
    })
  })

  output$messages <- renderText({
    msgs_simplified <- lapply(all_messages(), function(msg) {
      res <- list(
        role = msg@role,
        content = msg@contents[[1]]@text,
        model = msg@json$model
      )
      if (is.null(res$model)) {
        # Remove model if not present (will be the case for user messages)
        res$model <- NULL
      }
      res
    })

    jsonlite::toJSON(msgs_simplified, auto_unbox = TRUE, pretty = TRUE)
  })
}

# Run the application
shinyApp(ui = ui, server = server)