Skip to contents

Extract causal claims from text via an LLM backend

Source: R/causal_llm.R
causal_llm_extract.Rd

Calls a Large Language Model and returns a tidy table of causal claims — one row per extracted (cause, effect) pair — with the quoted evidence and the confidence reported by the model.

Usage

causal_llm_extract(
  text,
  backend = c("ollama", "openai", "anthropic"),
  model = NULL,
  host = "http://localhost:11434",
  temperature = 0,
  timeout_sec = 120,
  api_key = NULL
)

Arguments

text

Character scalar — the passage to annotate.

backend

One of "ollama", "openai", "anthropic".

model

Optional model name; defaults vary by backend.

host

Ollama server URL (ignored by the hosted backends).

temperature

Sampling temperature. Defaults to 0 for reproducibility.

timeout_sec

Request timeout in seconds.

api_key

Optional API key overriding the environment variable.

Value

A data frame with columns cause, effect, evidence, confidence. Empty (zero rows) when the model returns no claims or the response cannot be parsed.

Details

Three backends are supported via a uniform prompt:

  • "ollama" — local, zero-cost Ollama server (default host = "http://localhost:11434", default model "gemma4:latest"). For higher extraction quality switch to model = "gemma4:26b".

  • "openai" — OpenAI Chat Completions API with JSON mode; needs the OPENAI_API_KEY environment variable (default model "gpt-4o-mini").

  • "anthropic" — Claude Messages API; needs ANTHROPIC_API_KEY (default model "claude-sonnet-4-5").

The return value is backend-independent, so downstream causal_llm_ingest_abstract() and causal_kg_add_edge() work the same regardless of which model produced the claims.

Examples

if (FALSE) { # \dontrun{
  abstract <- "In Cerrado soils, higher mean annual precipitation is
    associated with significantly increased topsoil organic carbon."
  causal_llm_extract(abstract, backend = "ollama",
                     model = "gemma4:latest")
} # }