LLMR

LLMR gives R a single tidy-friendly interface to hosted and local language models. One llm_config() object selects the provider; every other function works the same regardless of which model is behind it.

Providers: OpenAI, Anthropic, Gemini (incl. Vertex AI), Groq, Together AI, DeepSeek, xAI, Voyage AI, Ollama, Xiaomi MiMo, Alibaba (Qwen), Zhipu, Moonshot.

Install

install.packages("LLMR")                        # CRAN
# remotes::install_github("asanaei/LLMR")       # dev

Setup

Store API keys as environment variables (OPENAI_API_KEY, ANTHROPIC_API_KEY, GEMINI_API_KEY, etc.) in ~/.Renviron.

library(LLMR)

cfg <- llm_config("openai", "gpt-4.1-nano", temperature = 0.2, max_tokens = 256)

One call

r <- call_llm(cfg, c(system = "Be concise.", user = "Capital of Mongolia?"))
r                    # prints text + status line
as.character(r)      # just the text
tokens(r)            # token counts

Data-frame pipelines

llm_mutate() adds LLM-generated columns. The shorthand puts the output name and glue prompt in one argument.

library(dplyr)

df <- tibble::tibble(city = c("Cairo", "Lima", "Oslo"))

df |>
  llm_mutate(country = "What country is {city} in? One word.", .config = cfg)

For vectors without a data frame, use llm_fn().

Structured output (JSON)

Get typed columns from a JSON schema:

schema <- list(
  type = "object",
  properties = list(
    answer = list(type = "string"),
    confidence = list(type = "number")
  ),
  required = list("answer", "confidence"),
  additionalProperties = FALSE
)

df |>
  llm_mutate(
    result = "What country is {city} in?",
    .config = cfg,
    .structured = TRUE,
    .schema = schema
  )

Soft structured output (XML-like tags)

When strict JSON is unnecessary, request fields as tags:

df |>
  llm_mutate(
    geo = "Where is {city}? Give country and continent.",
    .config = cfg,
    .tags = c("country", "continent")
  )

Embeddings

emb_cfg <- llm_config("voyage", "voyage-3.5-lite", embedding = TRUE)

texts <- c("Quiet rivers mirror bright skies.", "Thunder shakes the mountain path.")
emb   <- get_batched_embeddings(texts, emb_cfg, batch_size = 8)
dim(emb)

Chat sessions

chat <- chat_session(cfg, system = "You teach statistics tersely.")
chat$send("Explain p-values in 12 words.")
chat$send("Now give a three-word analogy.")

Parallel experiments

setup_llm_parallel(workers = 4)

experiments <- build_factorial_experiments(
  configs = list(cfg),
  user_prompts = c("Summarize: The Apollo program.", "Summarize: The Manhattan Project."),
  system_prompts = "One sentence."
)

res <- call_llm_par(experiments, progress = TRUE)
reset_llm_parallel()

Which function do I need?

Task	Function
Single prompt	call_llm() / call_llm_robust()
Vector of prompts	llm_fn()
Data-frame pipeline	llm_mutate()
JSON with schema	.structured = TRUE / llm_mutate_structured()
Tag-based extraction	.tags / llm_mutate_tags()
Parse existing column	llm_parse_structured_col() / llm_parse_tags_col()
Factorial designs	build_factorial_experiments() + call_llm_par()

The LLMR ecosystem

LLMR is the provider layer of a family of packages for LLM-assisted research. LLMRagent builds agents and multi-agent designs on top of it. LLMRcontent carries the measurement workflow in one package: codebook-first annotation with sealed gold-set validation, robustness audits of LLM-mediated estimates, and verifiable replication archives built from LLMR’s audit logs. LLMRpanel provides calibrated silicon samples for survey and experiment design, and FocusGroup supports moderated multi-agent discussions. An overview of the family lives at the ecosystem page.

Contributing

Bug reports and feature requests: GitHub Issues. Pull requests welcome. See CONTRIBUTING.md.