Trace an LLM Call — trace_llm_call • securetrace

Wraps an ellmer chat call with automatic span and token recording. Requires the ellmer package.

Usage

trace_llm_call(chat, prompt, ..., stream = FALSE)

Arguments

chat: An ellmer chat object.
prompt: The prompt string to send.
...: Additional arguments passed to the chat method.
stream: Logical. If TRUE, calls chat$stream(prompt, ...) instead of chat$chat(prompt, ...). Default FALSE.

Value

The chat response.

Details

When the chat object supports ellmer's get_model() and get_tokens() methods, the model name and token usage are automatically extracted and recorded on the span. Model names are resolved through resolve_model() for proper cost calculation with cloud provider model IDs.

Auto-extraction is best-effort: if the chat object does not support these methods (e.g., a non-ellmer Chat object), the span will still be created with latency recorded but without model or token data.

Examples

if (FALSE) { # \dontrun{
# Requires ellmer package
chat <- ellmer::chat_openai(model = "gpt-4o")
with_trace("llm-demo", {
  response <- trace_llm_call(chat, "What is 2 + 2?")
})
} # }