Extending krt with plugins

library(krt)
#> krt 0.1.0: author, validate, and export Key Resources Tables.
#>   Start with new_krt(); see https://choxos.github.io/krt/

krt is built on five registries, and each is a public extension point. An institution can add profiles, validators, resolvers, LLM providers, and autocomplete sources without forking the package.

krt_plugin_api()
#>             kind                register
#> 1        profile        register_profile
#> 2      validator      register_validator
#> 3       resolver       register_resolver
#> 4   llm_provider   register_llm_provider
#> 5 suggest_source register_suggest_source
#>                                                                       contract
#> 1          a directory with schema.yml + mappings.yml, or a krt_profile object
#> 2                                  function(x, ctx) returning a list of issues
#> 3         function(id, resolve = TRUE, ...) returning a normalized result list
#> 4                      function(prompt, llm) returning the model's text output
#> 5 function(query, n) returning a data frame (label, id, authority, score, uri)

Check an object against its contract before registering it:

validate_plugin_contract("validator", function(x, ctx) list())

A custom validator

A validator is function(x, ctx) returning a list of issues. Here is one that flags datasets lacking a license.

license_rule <- function(x, ctx) {
  out <- list()
  for (r in x$resources) {
    if (identical(r$resource_type, "Dataset") && is.null(r$license)) {
      out <- c(out, list(list(message = "Dataset has no license.",
                              resource_id = r$resource_id, field = "license")))
    }
  }
  out
}
register_validator("inst-dataset-license", license_rule,
                   layer = "semantic", severity = "warning")

k <- add_resource(new_krt("Demo"), "Dataset", "D",
                  doi = "10.5281/zenodo.1", new_or_reuse = "new")
"inst-dataset-license" %in% as.data.frame(validate_krt(k))$rule_id
#> [1] TRUE

A custom institutional profile

A profile is a directory with schema.yml and mappings.yml. Register it by path:

register_profile(name = "my-institute", path = "path/to/profile/dir")
export_krt(k, file.path(tempdir(), "out.csv"), profile = "my-institute")

A custom resolver or suggest source

register_resolver("myscheme", function(id, resolve = TRUE, ...) {
  list(input = id, normalized = id, resolved = FALSE, source = "mine")
})

register_suggest_source("mysource", function(query, n) {
  data.frame(label = query, id = "X:1", authority = "mine",
             score = 1, uri = NA_character_, stringsAsFactors = FALSE)
})

Registered plugins run inside the same error isolation as the built-ins: a rule or resolver that throws is caught, so a faulty plugin cannot abort a validation or resolution run. Plugins are still trusted code. A plugin registered with replace = TRUE deliberately overrides a built-in, and a validator whose predicate or body errors is downgraded to a warning rather than failing the run, so a plugin can change validation and output behavior.