Gemini RLM (Minimal)

Purpose: Provide a lightweight, CLI-based implementation of the Recursive Language Model architecture using Google’s Gemini models. This skill allows for processing extremely large documents by orchestrating chunking, sub-LLM processing, and synthesis entirely via a Python script and the Gemini API.

Architecture

Based on arXiv:2512.24601 – Recursive Language Models.

Prerequisites

• Environment Variable: GEMINI_API_KEY must be set in your shell environment.

  export GEMINI_API_KEY="your_api_key_here"
  

Usage

The primary entry point is the gem_rlm.py script.

Syntax

${SKILLS_ROOT}/gemini-rlm-min/gem_rlm.py --context <path_to_large_file> --query <"your query"> [options]

Options

• --chunk-size: Size of chunks in characters (default: 50000)
• --overlap: Overlap between chunks in characters (default: 0)

Examples

Analyze a large log file:

export GEMINI_API_KEY="AIza..."
${SKILLS_ROOT}/gemini-rlm-min/gem_rlm.py --context ./large_logs.txt --query "Identify all security exceptions and their timestamps"

Summarize a book:

${SKILLS_ROOT}/gemini-rlm-min/gem_rlm.py --context ./mobydick.txt --query "Summarize the relationship between Ahab and Starbuck" --chunk-size 100000

How It Works

1. Initialization: The script initializes a persistent Python REPL (rlm_repl.py) and loads the large context file into memory.
2. Chunking: The context is split into manageable chunks (e.g., 50k chars) using the REPL.
3. Sub-LLM Processing: The script iterates through each chunk, sending it to gemini-2.0-flash-exp with a prompt to extract relevant information.
4. Synthesis: The extracted findings from all chunks are aggregated and sent to the Root LLM (also Gemini 2.0 Flash) to generate the final answer.

File Structure

gemini-rlm-min/
├── SKILL.md              # This definition file
├── gem_rlm.py            # Main CLI Orchestrator
├── scripts/
│   └── rlm_repl.py       # Persistent REPL environment
└── state/                # Runtime state storage (chunks, pickle files)

Integration with IRP

This skill serves as a high-speed, low-overhead alternative to the full rlm-context-manager when:

• Quick analysis is needed via CLI.
• The context needs to be processed entirely by Gemini models.
• Minimal dependencies are preferred (no complex agent setup required).