Skip to contents

Multi-extractor consensus over LLM-extracted causal claims

Source: R/causal_llm_vote.R
causal_llm_vote.Rd

Runs N LLM backends on the same abstract and returns a single consensus data frame that resolves the inevitable disagreements by one of three voting rules. See the file header for the design rationale.

Usage

causal_llm_vote(
  abstract,
  backends,
  voting = c("majority", "weighted", "intersection"),
  min_support = NULL,
  threshold = NULL,
  weights = NULL,
  timeout_sec = 120
)

Arguments

abstract

Character scalar with the passage to annotate.

backends

List of backend configurations; see above.

voting

"majority" (default), "weighted", or "intersection".

min_support

Integer — required number of backends for "majority" voting. Defaults to ceiling(length(backends) / 2).

threshold

Numeric — weighted-confidence threshold for "weighted" voting. Defaults to 0.5 * sum(weights).

weights

Optional numeric vector, one element per backend (same order as backends). Defaults to uniform weights.

timeout_sec

Default per-request timeout; may be overridden per-backend via the timeout_sec field of each config entry.

Value

A data frame with one row per consensus edge and columns cause, effect, n_backends, mean_confidence, weighted_confidence, backends (" | "-separated list of supporting backend ids), evidence (concatenated quotations).

Voting rules

"majority" (default)

Keep edges asserted by at least min_support backends. When min_support = NULL, defaults to ceiling(length(backends) / 2) so 2-out-of-3 or 3-out-of-5 agreement survives.

"weighted"

Keep edges whose weighted confidence \(\sum_i w_i c_i\) exceeds threshold. Weights are supplied per backend via weights.

"intersection"

Keep only edges asserted by every backend. Most conservative; maximises precision at the expense of recall.

Expected backends shape

A list of lists. Each inner list configures one backend and must contain at least a backend field (one of "ollama", "openai", "anthropic") and may optionally carry model, host (for Ollama), temperature, timeout_sec, api_key, id. Example:


backends = list(
  list(backend = "ollama",    model = "gemma4:latest",
       host   = "http://localhost:11434"),
  list(backend = "openai",    model = "gpt-4o-mini"),
  list(backend = "anthropic", model = "claude-sonnet-4-6")
)

Failed backends (timeouts, missing API keys) emit a warning and contribute zero claims; the vote continues with the remaining backends.

See also

causal_llm_ingest_abstract_voted() for the KG-insertion wrapper; causal_llm_extract() for the single-backend call.

Examples

if (FALSE) { # \dontrun{
  backends <- list(
    list(backend = "ollama",    model = "gemma4:latest"),
    list(backend = "openai",    model = "gpt-4o-mini"),
    list(backend = "anthropic", model = "claude-sonnet-4-6")
  )
  cons <- causal_llm_vote(
    abstract = "In Cerrado Oxisols, precipitation drives SOC...",
    backends = backends,
    voting   = "majority"
  )
  cons
} # }