Compare several reasoning chains and synthesize an answer

multi_chain_comparison() builds a module that runs an inner module M times, then makes one more call that reads all the attempts and writes a final answer with its own reasoning (DSPy's MultiChainComparison).

Usage

multi_chain_comparison(
  signature,
  inner_module = NULL,
  M = 3L,
  temperature = 0.7,
  comparison_template = NULL,
  config = list(),
  chat = NULL,
  ...
)

Arguments

signature

A signature from signature(), or a signature string.

inner_module

The module that produces each attempt. The default is chain_of_thought(signature).

M

Number of attempts.

temperature

Temperature applied to the attempts, to make them differ; NULL sends none. Reasoning models may reject a temperature: gpt-6-luna accepts it only with reasoning_effort = "none".

comparison_template

A glue template for the comparison prompt. It can use {M}, {attempts_text} (each attempt's output fields under an "=== Attempt i ===" heading) and input fields such as {question}. The default shows the attempts but not the original inputs.

config, chat

As in module().

...

Must be empty.

Value

A module (an R6 object of class MultiChainComparisonModule).

Details

Each run() makes M + 1 model calls. A failed attempt gives a warning and is left out; the module fails only if every attempt fails. The final output has a reasoning field followed by the signature's output fields. The returned module's get_attempts() method lists the attempts of the last run.

With the response cache on, identical attempts are served from the cache, so pass .cache = FALSE to run() to get M independent attempts.

See also

Other program constructors: chain_of_thought(), code_act(), flex(), module(), module_fn(), program_of_thought(), rag_module(), react(), rlm(), rlm_module()

Examples

mcc <- multi_chain_comparison("question -> answer", M = 3L)
mcc
#> 
#> ── MultiChainComparisonModule ──
#> 
#> M: 3 reasoning chains
#> Temperature: 0.7
#> Inner module: <PredictModule>

# Let the comparison step see the question too
mcc_with_question <- multi_chain_comparison(
  "question -> answer",
  M = 3L,
  comparison_template = paste(
    "Question: {question}",
    "Here are {M} attempts:",
    "{attempts_text}",
    "Write the best final answer.",
    sep = "\n\n"
  )
)

if (FALSE) { # \dontrun{
run(
  mcc_with_question,
  question = "A bat and a ball cost $1.10. The bat costs $1 more. What does the ball cost?",
  .llm = ellmer::chat_openai(
    model = "gpt-6-luna",
    params = ellmer::params(reasoning_effort = "none")
  ),
  .cache = FALSE
)
} # }