List recent model calls

inspect_history() returns the most recent model calls of the session as a tibble, across all modules, like DSPy's dspy.inspect_history(). It can also write them to a plain-text transcript.

Usage

inspect_history(
  n = 10,
  include_prompts = TRUE,
  include_responses = TRUE,
  file = NULL
)

Arguments

n

Number of recent calls to return.

include_prompts

If TRUE (the default), include the full prompt text.

include_responses

If TRUE (the default), include the full response text.

file

A file path or writable connection. If given, the selected calls are also written there as a plain-text transcript.

Value

A tibble with one row per call and columns timestamp, source (the module class that made the call), model, tokens_in, tokens_out, cost (in US dollars, when known), duration_s, program_artifact_id, trace_context, and prompt and response when requested. With no recorded calls, an empty tibble and a message.

Details

The history is kept in memory for the session and holds the most recent 100 calls by default (options(dsprrr.prompt_history_max = )). Prediction modules (module(), chain_of_thought()) and rag_module(), react(), rlm_module() and flex() modules add their calls to it; calls made inside wrappers and pipelines, and by run_async() or run_stream(), are not recorded. clear_prompt_history() empties it.

See also

Other inspection: accessors, clear_prompt_history(), clear_traces(), export_traces(), get_last_prompt(), session_cost(), summarize_traces()

Examples

# Empty until a module has called a model
inspect_history(n = 5)
#> No LLM calls recorded yet
#> # A tibble: 0 × 0

if (FALSE) { # \dontrun{
qa <- module(signature("question -> answer"))
run(qa, question = "What is 2 + 2?", .llm = ellmer::chat_openai(model = "gpt-6-luna"))

history <- inspect_history(n = 20)
sum(history$cost, na.rm = TRUE)

# A plain-text transcript for sharing
inspect_history(n = 20, file = tempfile(fileext = ".txt"))
} # }