Create a Recursive Language Model (RLM) Module

Create an RLM whose implementation can adaptively explore R objects at inference time. Use RLM when the inspection path is not known in advance; use ordinary R when that path becomes stable, or Flex when labeled examples should discover a reusable implementation.

Usage

rlm_module(
  signature,
  runner = NULL,
  interpreter_factory = NULL,
  max_iterations = 20L,
  max_llm_calls = 50L,
  max_output_chars = 10000L,
  sub_lm = NULL,
  verbose = FALSE,
  tools = list(),
  config = list(),
  chat = NULL,
  generate_action = NULL,
  extract = NULL,
  ...
)

Arguments

signature: A Signature object or string notation defining inputs/outputs with explicit ellmer string, number, integer, boolean, enum, array, or object output types. Opaque TypeJsonSchema outputs are unsupported.
runner: Optional caller-owned code runner implementing execute() and policy(). Its policy must declare persistent = TRUE. It is retained, never automatically shut down, and must not be shared concurrently. For the trusted callr backend, use r_code_runner(persistent = TRUE).
interpreter_factory: Optional zero-argument function returning a fresh runner with execute(), policy(), optional start(), and idempotent terminal shutdown(). Its policy must advertise persistent = TRUE for RLM. Supply exactly one of runner and interpreter_factory.
max_iterations: Maximum REPL iterations before fallback (default 20)
max_llm_calls: Maximum recursive LLM calls allowed (default 50)
max_output_chars: Maximum model-visible characters per execution output. Longer output is shown as a head-and-tail excerpt. Default 10000.
sub_lm: Optional ellmer Chat for recursive queries. NULL inherits the invocation's outer Chat. Set max_llm_calls = 0 to disable recursion.
verbose: Logical. Print execution progress (default FALSE)
tools: Named list of user-defined host functions or ellmer ToolDef objects. Guest code emits an invocation-bound request, dsprrr validates it and invokes the original function in the host, and the guest is replayed with the response. Closures are never deparsed or serialized into generated code. These tools execute in the host process, outside the guest runner sandbox, with the host's permissions. ToolDef schemas guide generation; the callable must still enforce semantic constraints beyond the bridge's lossless JSON-compatible value checks. A protocol safety ceiling permits at most 1,000 host-tool calls in one generated R step.
config: Optional prediction configuration.
chat: Optional ellmer Chat object.
generate_action: Optional advanced action predictor.
extract: Optional advanced extraction predictor.
...: Must be empty.

Value

An RLMModule object

Examples

if (FALSE) { # \dontrun{
analyst <- rlm_module(
  "document, question -> answer",
  interpreter_factory = function() mcp_repl_runner(timeout = 30)
)
result <- run(
  analyst,
  document = "Owner: team-a\nObligation: rotate keys quarterly",
  question = "Which obligations have no owner?",
  .llm = ellmer::chat_openai()
)
} # }