Classify spec variables via an LLM chat backend — classify_variables

Runs the batch translate-validate-retry pipeline. With `samples > 1` the pipeline is repeated and per-variable signatures (step-layer chains) are voted on to counter LLM output variance.

Usage

classify_variables_llm(
  vars,
  chat,
  max_attempts = 2,
  batch_size = 4,
  samples = 1,
  consensus = c("majority", "first"),
  prompt_variant = c("full", "neutral", "minimal")
)

Arguments

vars: spec data frame from [read_spec_df()].
chat: ellmer chat object (or any object with a `chat(prompt)` method).
max_attempts: validation-repair attempts per batch.
batch_size: variables per LLM batch.
samples: number of independent pipeline runs; `1` (default) keeps the classic single-run behavior; `> 1` enables consensus mode.
consensus: `majority` = per-variable majority vote across samples; ties or all-failed samples mark the variable `needs_human`. `first` = skip voting and take the earliest successful sample.
prompt_variant: system-prompt variant: `full` (default) carries the complete convention block; `neutral` and `minimal` strip it down for agreement measurement, so consensus reflects independent translations rather than prompt compliance.

Value

List of `aa_variable_ir` objects. For `samples > 1` the list carries a `consensus` attribute: a data.frame with columns `variable`, `signature_votes` (e.g. `"impute_dtc->merge_var:2"`), `chosen`, `unanimous`.

Examples

if (FALSE) { # \dontrun{
# Requires the 'ellmer' package and a provider API key:
chat <- ellmer::chat_anthropic()
vars <- spec_variables(mock_spec_adsl(), "ADSL")
ir <- classify_variables_llm(vars, chat)
} # }