Skip to contents

Rebuild a callable request from a logged record

Source: R/log_read.R
llm_request_from_log.Rd

Turns one parsed log record (the rec element of a llm_log_read() entry) back into the config and messages an experiments row needs, so an archived call can be re-issued live or matched against a config-driven call. The provider-specific request body is canonicalized to provider-neutral role/content turns and its generation parameters are recovered onto a reconstructed llm_config().

Usage

llm_request_from_log(record, on_unsupported = c("error", "warn", "quiet"))

Arguments

record

A parsed log record: a list with at least provider, model, and request (as found in llm_log_read(x)$records[[i]]$rec).

on_unsupported

What to do when the request body uses a shape this reconstructor does not cover (so the rebuilt messages would be empty or lossy) – chiefly the OpenAI Responses API (input/instructions) and non-text content blocks (images, files). "error" (default) aborts; "warn" warns and returns the best-effort result; "quiet" returns it silently. Failing loudly is the default because a silently empty replay is worse than none.

Value

A list with config (an llm_config()), messages (a named character vector of role/content turns), and complete (a logical: TRUE when the body was fully reconstructed, FALSE when content was dropped). NULL when the record carries no request body.

See also

llm_log_read(), llm_request_hash()

Examples

rec <- list(provider = "openai", model = "gpt-4o-mini",
            request = list(
              messages = list(list(role = "user", content = "Hi")),
              temperature = 0))
req <- llm_request_from_log(rec)
req$messages