Add support to transform assistant response and user input

[JamesHWade]

Allowing user to transform the user input or assistant response would be quite useful. Something similar to the python implementation would be fantastic.

A motivating example could be to recreate shiny assistant in R.

Very excited by this package!

[gadenbuie]

Here's an example that demos how to do both

1. user input → modify() → send to the LLM
2. text from the LLM → modify while streaming → show to the user

It's worth noting that there's a bunch more work to do to get from this example to the message transformation options on the Python side. But at the same time, we're on the fence with those transformers. Especially as we get into more detailed content types from ellmer and more advanced features like tool calling UI and bookmarking, each of which will move us a little more away from just sharing plain text, while introducing constraints that make it harder to support arbitrary content transformations.

Anyway, here's the example. Modifying the user input is easy, just change the text before you call chat$stream_async(). Modifying the assistant response (before it's shown in the UI) is harder and requires an adapting generator.

library(shiny)
library(bslib)
library(shinychat)

ui <- page_fillable(
  layout_columns(
    card(card_header("Chat"), chat_ui("chat")),
    card(card_header("Turns"), verbatimTextOutput("turns"))
  )
)

server <- function(input, output, session) {
  chat <- ellmer::chat_openai("You're a trickster who answers in riddles")

  turns <- reactiveVal()

  observeEvent(input$chat_user_input, {
    user <- toupper(input$chat_user_input)

    stream_to_lower <- coro::async_generator(function(stream) {
      for (msg in coro::await_each(stream)) {
        # Modify the message before sending it to the chat
        yield(tolower(msg))
      }
    })

    message("Sending user input: ", user)
    stream <- stream_to_lower(chat$stream_async(user))
    chat_append(id = "chat", stream)$then(function(value) {
      turns(chat$get_turns())
    })
  })

  output$turns <- renderPrint({
    validate(need(turns(), "No turns yet"))

    str(lapply(turns(), function(turn) {
      list(
        role = turn@role,
        content = ellmer::contents_markdown(turn)
      )
    }))
  })
}

shinyApp(ui, server)

Notice in the app that the LLM sees all caps user input, and the user will always see all lowercase assistant text.

[JamesHWade]

Thank you for the explanation!