pourquoi

Why settle for one noisy answer when you can vote & verify?

Add robust step-by-step reasoning to any base language model. Generate diverse reasoning traces, vote on the consensus, and verify with the same LLM or via a separate low-cost LLM.

pourquoi uses Verified Self-Consistency (VSC), a minimal ensemble technique: generate diverse traces with a base model for reasoning, normalize answers for consensus voting, then verify with a final LLM verifier.

library(pourquoi)

prompt <- "how do i say why in French?"

pourquoi::why(
  prompt = prompt,
  reasoner = "ollama/gemma3:4b",
  verifier = "ollama/qwen3:0.6b",
  n_traces = 3
)
#> [1] "CONFIRMED ANSWER: pourquoi"

prompt <- "how many r's are in the word Strawberry?"

pourquoi::why(
  prompt = prompt,
  reasoner = "ollama/ministral-3:8b",
  verifier = "ollama/ministral-3:3b",
  n_traces = 5
)
#> [1] "3"