Ravel is meant to feel closer to an analysis copilot than a generic chat box. The point is not just to answer a question about R syntax. The point is to reason with the current script, the selected code, the loaded objects, the models already in memory, the surrounding project, and the reporting workflow.
Imagine you just fit a model and the output looks suspicious. Coefficients are huge, standard errors are unstable, or the interaction terms are hard to explain.
fit <- lm(mpg ~ wt * am, data = mtcars)
ravel::ravel_summarize_model(fit)
#> $name
#> [1] "fit"
#>
#> $class
#> [1] "lm"
#>
#> $formula
#> [1] "mpg ~ wt * am"
#>
#> $fit
#> $fit$n
#> [1] 32
#>
#> $fit$aic
#> [1] 157.476
#>
#> $fit$bic
#> [1] 164.8046
#>
#> $fit$r_squared
#> [1] 0.8330375
#>
#> $fit$adj_r_squared
#> [1] 0.8151486
#>
#> $fit$sigma
#> [1] 2.591247
#>
#>
#> $coefficients
#> # A tibble: 4 × 5
#> term estimate std_error statistic p_value
#> <chr> <dbl> <dbl> <dbl> <dbl>
#> 1 (Intercept) 31.4 3.02 10.4 4.00e-11
#> 2 wt -3.79 0.786 -4.82 4.55e- 5
#> 3 am 14.9 4.26 3.49 1.62e- 3
#> 4 wt:am -5.30 1.44 -3.67 1.02e- 3Ravel is designed for prompts like:
Explain this model in plain English, tell me how the interaction changes the interpretation of the main effects, and list the next diagnostics I should run before I trust it.
cat(ravel::ravel_interpret_model(fit))
#> Model type: lm.
#> Formula: mpg ~ wt * am.
#> The model explains about 83.3% of the observed variance (R-squared = 0.833).
#> Term `am` has a positive association (estimate = 14.878) with statistical evidence at the 0.05 level.
#> Term `wt:am` has a negative association (estimate = -5.298) with statistical evidence at the 0.05 level.
#> Term `wt` has a negative association (estimate = -3.786) with statistical evidence at the 0.05 level.
#> Interaction terms are present (wt:am), so main effects should be interpreted conditionally.ravel::ravel_suggest_diagnostics(fit)
#> [1] "Inspect residuals versus fitted values for non-linearity and heteroskedasticity."
#> [2] "Review a Q-Q plot of residuals for heavy tails or strong departures from normality."
#> [3] "Check leverage and Cook's distance for influential observations."
#> [4] "Consider multicollinearity diagnostics if predictors are strongly related."Once you already have a fitted object, the next bottleneck is often writing. Ravel can help scaffold methods, results, and diagnostics sections without leaving the IDE.
#> ## Results
#>
#> ```{r results}
#> # Fit or summarize the final model here
#> ```
#>
#> Model type: lm.
#> Formula: mpg ~ wt * am.
#> The model explains about 83.3% of the observed variance (R-squared = 0.833).
#> Term `am` has a positive association (estimate = 14.878) with statistical evidence at the 0.05 level.
#> Term `wt:am` has a negative association (estimate = -5.298) with statistical evidence at the 0.05 level.
#> Term `wt` has a negative association (estimate = -3.786) with statistical evidence at the 0.05 level.
#> Interaction terms are present (wt:am), so main effects should be interpreted conditionally.For a more diagnostics-heavy turn:
Draft a Quarto diagnostics section for this model, summarize the main caveats, and include a chunk for the checks I should run next.
#> ## Diagnostics
#>
#> ```{r diagnostics}
#> # Produce diagnostic plots or checks here
#> ```
#>
#> Model type: lm.
#> Formula: mpg ~ wt * am.
#> The model explains about 83.3% of the observed variance (R-squared = 0.833).
#> Term `am` has a positive association (estimate = 14.878) with statistical evidence at the 0.05 level.
#> Term `wt:am` has a negative association (estimate = -5.298) with statistical evidence at the 0.05 level.
#> Term `wt` has a negative association (estimate = -3.786) with statistical evidence at the 0.05 level.
#> Interaction terms are present (wt:am), so main effects should be interpreted conditionally.
#>
#> Suggested checks:
#> - Inspect residuals versus fitted values for non-linearity and heteroskedasticity.
#> - Review a Q-Q plot of residuals for heavy tails or strong departures from normality.
#> - Check leverage and Cook's distance for influential observations.
#> - Consider multicollinearity diagnostics if predictors are strongly related.Ravel stages actions before execution, which matters for analysis work where a small change can silently alter results.
action <- ravel::ravel_preview_code("
summary(mtcars$mpg)
quantile(mtcars$mpg)
")
action$label
#> [1] "Run generated R code"You can then approve and run it explicitly:
approved <- ravel::ravel_approve_action(action)
ravel::ravel_run_code(approved)
#> $success
#> [1] TRUE
#>
#> $output
#> [1] " Min. 1st Qu. Median Mean 3rd Qu. Max. "
#> [2] " 10.40 15.43 19.20 20.09 22.80 33.90 "
#> [3] " 0% 25% 50% 75% 100% "
#> [4] "10.400 15.425 19.200 22.800 33.900 "
#>
#> $warnings
#> character(0)
#>
#> $messages
#> character(0)
#>
#> $value
#> 0% 25% 50% 75% 100%
#> 10.400 15.425 19.200 22.800 33.900
#>
#> $error
#> NULLWhen Ravel inspects the current session, it can summarize data frames, models, formulas, and other objects directly.
analysis_objects <- list(
data = mtcars,
formula = mpg ~ wt * am,
model = fit
)
lapply(analysis_objects, ravel::ravel_summarize_object)
#> $data
#> $data$name
#> [1] "X[[i]]"
#>
#> $data$kind
#> [1] "data.frame"
#>
#> $data$class
#> [1] "data.frame"
#>
#> $data$rows
#> [1] 32
#>
#> $data$cols
#> [1] 11
#>
#> $data$columns
#> [1] "mpg" "cyl" "disp" "hp" "drat" "wt" "qsec" "vs" "am" "gear"
#> [11] "carb"
#>
#> $data$column_types
#> [1] "numeric" "numeric" "numeric" "numeric" "numeric" "numeric" "numeric"
#> [8] "numeric" "numeric" "numeric" "numeric"
#>
#> $data$preview
#> [1] " mpg cyl disp hp drat wt qsec vs am gear carb\nMazda RX4 21.0 6 160 110 3.90 2.620 16.46 0 1 4 4\nMazda RX4 Wag 21.0 6 160 110 3.90 2.875 17.02 0 1 4 4\nDatsun 710 22.8 4 108 93 3.85 2.320 18.61 1 1 4 1"
#>
#>
#> $formula
#> $formula$name
#> [1] "X[[i]]"
#>
#> $formula$kind
#> [1] "formula"
#>
#> $formula$class
#> [1] "formula"
#>
#> $formula$formula
#> [1] "mpg ~ wt * am"
#>
#>
#> $model
#> $model$name
#> [1] "X[[i]]"
#>
#> $model$class
#> [1] "lm"
#>
#> $model$formula
#> [1] "mpg ~ wt * am"
#>
#> $model$fit
#> $model$fit$n
#> [1] 32
#>
#> $model$fit$aic
#> [1] 157.476
#>
#> $model$fit$bic
#> [1] 164.8046
#>
#> $model$fit$r_squared
#> [1] 0.8330375
#>
#> $model$fit$adj_r_squared
#> [1] 0.8151486
#>
#> $model$fit$sigma
#> [1] 2.591247
#>
#>
#> $model$coefficients
#> # A tibble: 4 × 5
#> term estimate std_error statistic p_value
#> <chr> <dbl> <dbl> <dbl> <dbl>
#> 1 (Intercept) 31.4 3.02 10.4 4.00e-11
#> 2 wt -3.79 0.786 -4.82 4.55e- 5
#> 3 am 14.9 4.26 3.49 1.62e- 3
#> 4 wt:am -5.30 1.44 -3.67 1.02e- 3
#>
#> $model$kind
#> [1] "model"This enables prompts like:
Summarize the objects I have loaded, tell me which one looks like the main modeling artifact, and propose the next analysis step.
Ravel is also designed for prompts that depend on the active editor and project-level context:
Use the selected code and the current project context to tell me whether this refactor changes results or just changes style.
Summarize the latest analysis diff and tell me what I should validate before I merge it.
Convert this selected tidyverse pipeline to base R, but keep it readable and consistent with the rest of the file.
These are exactly the kinds of tasks where a project-aware assistant feels different from a plain browser chat.
The intended flow is:
Once the addin is open, Ravel can answer questions like:
The long-term goal is not to mimic a chat website inside RStudio. It is to make RStudio feel like it has an analysis-aware copilot that understands:
That is the difference Ravel is trying to push.