Extract structured data with the Batch API — foundry_extract

Prepare JSONL requests for structured extraction, upload them, and create a batch. With wait = TRUE, waits for completion and returns parsed results joined back to the input rows.

Usage

foundry_extract_batch(
  data,
  text_col,
  schema,
  model,
  wait = FALSE,
  path = tempfile(fileext = ".jsonl"),
  schema_name = "ExtractedData",
  strict = TRUE,
  instructions = NULL,
  completion_window = "24h",
  api_key = NULL,
  token = NULL,
  endpoint_url = NULL,
  api_version = NULL
)

Arguments

data: Data frame containing input rows.
text_col: Character. Name of the column containing input text.
schema: List. JSON Schema object for structured extraction.
model: Character. Model deployment name to include in each request.
wait: Logical. Whether to block until the batch reaches a terminal state and parse results.
path: Character. Optional JSONL path. Defaults to a temporary file.
schema_name: Character. Name for schema when supplied.
strict: Logical. Whether structured output should be strict.
instructions: Character. Optional instructions for Responses API requests.
completion_window: Character. Batch completion window, usually "24h".
api_key: Character. Optional API key override.
token: Character. Optional bearer token override.
endpoint_url: Character. Optional Foundry endpoint override.
api_version: Character. Optional API version query value.

Value

A batch tibble when wait = FALSE, or parsed result rows when wait = TRUE.

Examples

if (FALSE) { # \dontrun{
jobs <- data.frame(text = c("Great service.", "Slow support."))
schema <- foundry_schema(sentiment = schema_string())
foundry_extract_batch(jobs, text_col = "text", schema = schema, model = "gpt-5-nano")
} # }