rlm-marketplace

RLM - Recursive Language Model

A modern Python 3.11+ implementation of the Recursive Language Model paradigm from MIT CSAIL research (arXiv:2512.24601).

What is RLM?

Unlike traditional RAG (Retrieval-Augmented Generation), RLM treats document context as an external variable in a Python REPL environment. The LLM doesn't see the full document - instead, it writes Python code to:

1. Inspect the document (len(CONTEXT), CONTEXT[:1000])
2. Search it (re.findall(r'pattern', CONTEXT))
3. Chunk it and process recursively (llm_query(f"Summarize: {chunk}"))
4. Synthesize results and return a final answer (FINAL(answer))

This approach enables processing of documents far exceeding typical context windows while maintaining adaptive, task-specific exploration strategies.

Architecture

┌─────────────────────────────────────────────────────────┐
│                      User Query                          │
└─────────────────────┬───────────────────────────────────┘
                      ▼
┌─────────────────────────────────────────────────────────┐
│                    RLM Orchestrator                      │
│  • Manages iteration loop                                │
│  • Parses code blocks from LLM response                  │
│  • Detects FINAL() answers                               │
└─────────────────────┬───────────────────────────────────┘
                      ▼
┌─────────────────────────────────────────────────────────┐
│                    Root LLM                              │
│  • Receives query + system prompt (NOT the context!)     │
│  • Generates Python code to explore CONTEXT              │
│  • Calls llm_query() for sub-processing                  │
└─────────────────────┬───────────────────────────────────┘
                      ▼
┌─────────────────────────────────────────────────────────┐
│                    REPL Environment                      │
│  • CONTEXT variable (holds the massive text)             │
│  • llm_query(snippet, task) for sub-RLM calls              │
│  • FINAL(answer) for returning results                    │
│  • Isolation delegated to container runtime               │
└─────────────────────────────────────────────────────────┘

Installation

pip install -r requirements.txt

Requirements:

• Python 3.11+ (tested with 3.13+)
• anthropic>=0.39.0 (for Anthropic API backend)
• openai>=1.0.0 (optional, for local models via Ollama/vLLM)

Quick Start

Using Anthropic API

from rlm import RLM
from rlm.backends import AnthropicBackend

# Initialize with API key (or set ANTHROPIC_API_KEY env var)
backend = AnthropicBackend()
rlm = RLM(
    backend,
    model="claude-sonnet-4-20250514",
    sub_rlm_model="claude-haiku-3-20250813",  # Cheaper model for sub-RLM calls
    verbose=True,
)

# Process a large document
with open("large_document.txt") as f:
    context = f.read()

result = rlm.completion(
    context=context,
    query="What are the main themes discussed in this document?"
)

print(result.answer)
print(rlm.cost_summary())

Using OpenAI

OPENAI_API_KEY=... uvx --with openai rlm --backend openai --model gpt-4o --context-file doc.txt --query "Summarize"

Using OpenRouter

OPENROUTER_API_KEY=... uvx --with openai rlm --backend openrouter --model anthropic/claude-sonnet-4 --context-file doc.txt --query "Summarize"

Using Hugging Face

HF_TOKEN=... uvx --with openai rlm --backend huggingface --model Qwen/Qwen2.5-Coder-32B-Instruct --context-file doc.txt --query "Summarize"

Using Local Models (Ollama)

from rlm import RLM
from rlm.backends import OpenAICompatibleBackend

backend = OpenAICompatibleBackend(
    base_url="http://localhost:11434/v1",
    api_key="ollama",
)
rlm = RLM(backend, model="llama3.2", verbose=True)

result = rlm.completion(context=doc, query="Summarize this document")

Demo

# With Anthropic API (set ANTHROPIC_API_KEY)
python demo.py --verbose

# With Ollama
python demo.py --backend ollama --model llama3.2 --verbose

# Custom context file
python demo.py --context-file /path/to/document.txt --query "Your question here"

Configuration File

Use --config rlm.yaml to configure per-role backends, models, and system prompts:

defaults:
  backend: anthropic
  model: claude-sonnet-4-20250514

roles:
  root:
    model: claude-sonnet-4-20250514
  sub_rlm:
    backend: ollama
    model: llama3.2
    base_url: http://localhost:11434/v1
  verifier:
    model: claude-haiku-4-5-20251001

settings:
  max_iterations: 10
  verbose: true
  verify: true

uvx rlm --config rlm.yaml --context-file doc.txt --query "Summarize"

CLI flags always override config values. Merge priority: CLI flags > roles.{role} > defaults > hardcoded defaults.