rtemis.llm R package

Unified interface for creating LLM and Agent objects, generating responses, and performing batch inference.
Built on a type-checked and validated ‘S7’ backend.
Features reasoning, structured output, memory management, and tool use.
Supports Ollama, OpenAI-compatible, and Anthropic-compatible endpoints, and Apple Foundation Models on-device through the rtemis-afm bridge.

Features

Installation

CRAN

{r} install.packages("rtemis.llm")

or

{r} pak::pak("rtemis.llm")

R-universe

{r} install.packages("rtemis.llm", repos = "https://rtemis-org.r-universe.dev")

or

pak::repo_add(myuniverse = "https://rtemis-org.r-universe.dev")
pak::pak("rtemis.llm")

GitHub

pak::pak("rtemis-org/llm")

Documentation

For detailed documentation, see the rtemis.llm documentation.

Quick Usage

library(rtemis.llm)

List available Ollama models

ollama_list_models()

LLM

Create an LLM object

llm <- create_Ollama(
  model_name = "gemma4:26b",
  system_prompt = "You are a meticulous research assistant.",
  temperature = 0.3
)

generate(llm, "What is the role of the telomere?")

Agent

Create an Agent object

agent <- create_agent(
  llmconfig = config_Ollama(
    model_name = "gemma4:26b",
    temperature = 0.3
  ),
  system_prompt = "You are a meticulous research assistant.",
  name = "Kaimana"
)

generate(agent, "Explain quantum superposition in seven bullet points.")

Apple Foundation Models

On an Apple silicon Mac with macOS 27 and Apple Intelligence turned on, the on-device model is served by the rtemis-afm bridge. Install and start it once in a terminal (curl -fsSL https://live.rtemis.org/afm.sh | sh, or brew install rtemis-org/tap/rtemis-afm then rtemis-afm); no API key is needed.

llm <- create_Apple(system_prompt = "You are a meticulous research assistant.")
generate(llm, "What is the role of the telomere?")

agent <- create_agent(config_Apple(), tools = list(tool_datetime))
generate(agent, "What is the date today?")

config_Apple() checks the bridge’s health first and says what to do if it is not running or the model is unavailable; apple_health() reports the served model and its context window (8,192 tokens on macOS 27.0).

Structured output validation

Validation runs locally when an output schema is supplied. Invalid output is retained by default, with an informational message through rtemis.core::warn() (not an R warning). This applies to single responses and batches, including small local models that may not reliably follow schemas.

sch <- schema("Count", field("n", type = "integer"))
out <- llmapply(
  c("How many days are in a week?", "How many months are in a year?"),
  "gemma4:e4b",
  output_schema = sch
)
report <- validation_results(out)
report@status   # valid, invalid, unavailable, or not_validated
report@issues   # input index, JSON path, keyword, and diagnostic message

Set on_validation_failure = "collect" to record diagnostics silently, or "abort" to raise an error on a mismatch. Batch validation occurs per response; the default logs a single summary. Validation aborts follow the batch’s on_error policy, with rejected text retained in the validation report.

You can also generate with validate_output = FALSE and validate later, or check any saved JSON directly:

report <- validate_output(c('{"n":10}', '{"n":"10"}'), sch)
report@status  # "valid" "invalid"

Validation checks the requested schema without coercing values, stripping Markdown, or repairing JSON. Current schemas allow extra properties, optional fields permit omission but not null, and array/object fields constrain only the outer type.