--- title: "Getting started with gander" output: rmarkdown::html_vignette vignette: > %\VignetteIndexEntry{Getting started with gander} %\VignetteEngine{knitr::rmarkdown} %\VignetteEncoding{UTF-8} --- ```{r, include = FALSE} knitr::opts_chunk$set( collapse = TRUE, comment = "#>" ) ``` gander is a high performance and low friction chat experience for data scientists in RStudio and Positron--sort of like completions with Copilot, but it knows how to talk to the objects in your R environment. The package supplies an addin that can be registered to a keybinding. Once that's done, just trigger the addin while you're working on documents and type a request; along with your request, the addin will automatically inject context about the document you're working in and the objects in your R environment. ```{r} #| label: gander-demo #| echo: false #| fig-alt: "A screencast of an RStudio session. A script called example.R is open in the editor with lines library(ggplot2), data(stackoverflow), and stackoverflow. After highlighting, I trigger the addin and ask to plot the data in plain language, at which code to plot the data using ggplot2 is streamed from an LLM into the source file that uses the correct column names and a minimal style. From there, I iteratively call the addin to refine the output." knitr::include_graphics("https://github.com/user-attachments/assets/4aead453-c2f2-446b-8e81-9154fd3baab0") ``` This interface excels beyond a typical LLM chat in two notable ways: * **Higher-performance**: Including information about your R environment and relevant code ensures that the model has all of the context it needs to supply an answer that's as accurate as possible _and_ aligns with your taste and style preferences in that it can mirror the code you've written already. * **Lower-friction**: gander eliminates the need to copy and paste relevant code into the chat window and its result back into your document. There's also no need to type anything other than your request, e.g. descriptions of the style you'd like the response to take or bits of exposition like "Here's some additional context:". Further, iterating on pieces of content is easier in that the model's output will automatically be selected after it's streamed into your document; to iterate on a response, just trigger the addin again. ## Choosing a model The gander addin supports any model supported by [ellmer](https://ellmer.tidyverse.org/). When choosing a model for use with gander, you'll want to the use the most performant model possible that satisfies your privacy needs; gander automatically passes along your code and the objects in your R environment to your chosen model, so it's especially important to consider data privacy when using LLMs with gander. gander uses the `.gander_chat` option to configure which model powers the addin; just set the option to whatever your usual ellmer setup is. For example, to use Anthropic's Claude, you might write `options(.gander_chat = ellmer::chat_claude())`. Paste that code in your `.Rprofile` via `usethis::edit_r_profile()` to always use the same model every time you start an R session. If you're using ellmer inside a organization, you'll be limited to what your IT department allows, which is likely to be one provided by a big cloud provider, e.g. `chat_azure()`, `chat_bedrock()`, `chat_databricks()`, or `chat_snowflake()`. If you're using ellmer for your own exploration, you'll have a lot more freedom, so we have a few recommendations to help you get started: - As of early 2025, Anthropic's **Claude Sonnet 3.5** is a very powerful model for code assistance and is the model I've used while developing the package. If you want to use Claude, you'll need to register an [API key](https://console.anthropic.com/) to the environment variable `ANTHROPIC_API_KEY` and then set `options(.gander_chat = ellmer::chat_claude())`. * Regarding OpenAI's models, `chat_openai()` defaults to **GPT-4o**, but you can use `model = "gpt-4o-mini"` for a cheaper, lower-quality model, or `model = "o1-mini"` for more complex reasoning; to use an OpenAI model, you'll need to set the options `options(.gander_chat = ellmer::chat_openai(model = "gpt-4o"))` and register your OpenAI API key with the `OPENAI_API_KEY` environment variable. - You can use a **local model** with `chat_ollama()`, which uses [Ollama](https://ollama.com) and allows you to run models on your own computer. While the biggest models you can run locally aren't as good as the state of the art hosted models, they don't share your data and are effectively free. To use an ollama model, run the model locally and then set `options(.gander_chat = ellmer::chat_ollama(model = "model-name"))`. Instead of `model-name`, you'd substitute in the name of the model that appears when you run `ollama list` at the console. ## What is gander actually doing? gander automatically incorporates two kinds of context about your analyses under the hood: * **Code context**: The assistant will automatically inform the model about the type of file you're working in and relevant lines of code surrounding your cursor. * **Environment context**: The assistant also interfaces directly with your global environment to describe the relevant data frames you have loaded, including their column names and type. This allows you to describe the analyses you'd like to carry out in plain language while gander takes care of describing your computational environment to the model. For example, imagine you're working in a .R file that looks like this: ```r library(ggplot2) data("stackoverflow", package = "modeldata") stackoverflow ``` [stackoverflow](https://modeldata.tidymodels.org/reference/stackoverflow.html) is a data set containing results from the annual Stack Overflow Developer Survey. You select `stackoverflow`, trigger the addin, and type `plot salary vs. years coded`. To effectively respond to this input, a model would need to know a few additional things: 1) Code context: a) What framework does this user want to use to plot this data? b) In what style should the model respond? Notably, this is a `.R` file, so any exposition would need to live inside of comments. 2) What does the data look like? Specifically, what are the column names related to the salary and number of years coded? From the surrounding contents in the file, we can answer **1a)**; the user likely wants to use ggplot2 to do so. We can answer **1b)** by inspecting the file extension—since this is an `.R` file, the model should respond only with code, omitting exposition and backticks. We can answer **2)** using the user's R environment; if they've run that `data("stackoverflow", package = "modeldata")` line, then the `stackoverflow` object will live in their environment and we can print it out to learn more about it. The gander addin will preemptively answer all of these questions for the model by including context from each of these sources. To demonstrate, we can run this `stackoverflow` example ourselves and then call `gander_peek()` to peek at everything gander actually sent to the model. Under the hood, gander creates an [ellmer chat](https://ellmer.tidyverse.org/reference/Chat.html) and provides it a bunch of context---`gander_peek()` returns that chat object from your most recent interaction with gander as-is. So, when I do so: ```{r} #| eval: false gander::gander_peek() #> #> #> ── system ───────────────────────────────────────────────────────────────────── #> You are a helpful but terse R data scientist. Respond only with valid R code: #> no exposition, no backticks. Always provide a minimal solution and refrain #> from unnecessary additions. Use tidyverse style and, when relevant, tidyverse #> packages. For example, when asked to plot something, use ggplot2, or when #> asked to transform data, using dplyr and/or tidyr unless explicitly instructed #> otherwise. #> #> ── user ─────────────────────────────────────────────────────────────────────── #> Up to this point, the contents of my r file reads: #> #> ``` #> library(ggplot2) #> #> data("stackoverflow", package = "modeldata") #> #> ``` #> #> Now, plot salary vs. years coded in the following: #> #> stackoverflow #> #> Here's some information about the objects in my R environment: #> #> ``` #> stackoverflow #> #> tibble [5,594 × 21] (S3: tbl_df/tbl/data.frame) #> #> $ Country : Factor w/ 5 levels "Canada","Germany",..: 4 5 5 2 3 5 5 2 5 2 ... #> #> $ Salary : num [1:5594] 100000 130000 175000 64516 6636 ... #> #> $ YearsCodedJob: int [1:5594] 20 20 16 4 1 1 13 4 7 17 ... #> #> $ OpenSource : num [1:5594] 0 1 0 0 0 0 0 1 1 1 ... #> #> $ Hobby : num [1:5594] 1 1 1 0 1 1 1 0 1 1 ... #> #> $ CompanySizeNumber : num [1:5594] 5000 1000 10000 1000 5000 20 20 5000 20 20 ... #> #> $ Remote : Factor w/ 2 levels "Remote","Not remote": 1 1 2 2 2 2 2 2 2 2 ... #> #> $ CareerSatisfaction : int [1:5594] 8 9 7 9 5 8 7 7 8 9 ... #> #> $ Data_scientist : num [1:5594] 0 0 0 0 0 0 0 0 0 0 ... #> #> $ Database_administrator : num [1:5594] 0 0 0 0 0 0 0 0 0 0 ... #> #> $ Desktop_applications_developer : num [1:5594] 0 0 0 0 0 0 0 0 1 1 ... #> #> $ Developer_with_stats_math_background: num [1:5594] 0 0 0 0 0 0 0 0 0 1 ... #> #> $ DevOps : num [1:5594] 0 1 0 0 0 0 0 0 0 0 ... #> #> $ Embedded_developer : num [1:5594] 1 1 0 0 0 0 0 0 0 0 ... #> #> $ Graphic_designer : num [1:5594] 0 0 0 0 0 0 0 0 0 0 ... #> #> $ Graphics_programming : num [1:5594] 0 0 0 0 0 0 0 0 0 0... #> #> $ Machine_learning_specialist : num [1:5594] 0 0 0 0 0 0 0 0 0 0 ... #> #> $ Mobile_developer : num [1:5594] 0 0 0 0 0 0 0 0 1 0 ... #> #> $ Quality_assurance_engineer : num [1:5594] 0 1 0 0 0 0 0 0 0 0 ... #> #> $ Systems_administrator : num [1:5594] 0 0 0 0 0 0 0 0 0 0 ... #> #> $ Web_developer : num [1:5594] 0 1 1 1 1 1 1 1 0 0 ... #> ``` #> #> ── assistant ────────────────────────────────────────────────────────────────── #> ggplot(stackoverflow, aes(x = YearsCodedJob, y = Salary)) + #> geom_point() + #> labs(x = "Years Coded", y = "Salary") ``` Even though the user selected only a variable name and typed 5 words, that request was automatically situated in enough context for the model to reply only in R code that uses ggplot2 and references column names that the user hasn't supplied. First, gander knows the framework to use since the preceding lines in the file have been inlined into the user prompt. (By default, gander recommends that models lean on the tidyverse in their responses anyway. You can adjust this with `?gander_options`.) Then, with respect to style, gander has detected that we're working in a `.R` file and thus writes "Respond only with valid R code" in the system prompt. Finally, gander inspects the code selection and describes variables from the global environment if they're mentioned either in the selection or in the text inputted in the addin. In this case, that's only the `stackoverflow` data. Altogether, this allows for the model to tailor its response to the user's development environment without the user needing to supply any context themselves.