Handling edge cases & guardrails

Missing or empty text

df <- tibble(text = c("ok", NA, "   "))

df |>
  mutate(
    label = classify_llm(
      text,
      categories = c("Yes", "No")
    )
  )

Out-of-set or ambiguous responses

df <- tibble(text = c("fine", "bad", "so-so", "excellent"))

df |>
  mutate(
    raw = classify_llm(text, categories = c("Positive","Negative")),
    label = dplyr::recode(raw,
      "so-so" = "Neutral",
      .default = raw
    )
  )

Warning

Tip: You can always post-process LLM output with recode() or a lookup table to enforce strict categories.

Long inputs and rate limits

Use batch_size and delay to control throughput:

df |>
  mutate(
    category = classify_llm(
      text,
      categories = c("Info","Complaint","Request"),
      batch_size = 5,
      delay = 2
    )
  )