Crate librlm 

librlm

Implementation of the Recursive Language Models (RLM) algorithm as described in “Recursive Language Models” (Zhang, Kraska, Khattab — MIT CSAIL, Jan 2026).

RLM enables LLMs to handle arbitrarily long prompts by treating them as part of an external environment. The LLM interacts with the prompt through a persistent Lua REPL, writing code to peek at, decompose, and recursively invoke sub-LLMs over manageable chunks.

Quick Start

use librlm::Rlm;

let rlm = Rlm::builder()
    .root_model("gpt-5")
    .root_api_key("sk-...")
    .sub_model("gpt-5-mini")
    .max_iterations(30)
    .build()?;

let result = rlm.completion("very long prompt...", Some("What is X?")).await?;
println!("{}", result.response);

Structs

CodeBlock

A code block extracted from LLM output.

CompletionResponse

Response from an LLM completion call.

Message

A single message in a conversation.

OpenAiBackend

OpenAI-compatible API backend.

ReplResult

Result of executing code in the REPL.

Rlm

The main RLM (Recursive Language Model) engine.

RlmBuilder

Builder for constructing an Rlm instance.

RlmCompletion

The final result of an RLM completion.

RlmConfig

Configuration for the RLM algorithm.

UsageInfo

Token usage info from an LLM response.

Enums

FinalAnswer

How the final answer was signaled.

RlmError

All errors that can occur in the RLM library.

Role

Role in a conversation message.

Traits

LlmBackend

Trait for LLM backends. Implement this for custom providers.