rlm.pi

Recursive Language Models for Pi.

Use a cheap model to read very large contexts with high accuracy.

_{English · 中文 · Русский}

Install

	Install	Upgrade	Remove
Pi	pi install npm:@hicaru/pi-rlm	add --force	pi uninstall npm:@hicaru/pi-rlm
oh-my-pi	omp plugin install @hicaru/pi-rlm	add --force	omp plugin uninstall @hicaru/pi-rlm

1. Install the plugin.
2. Type /reload, or restart Pi.
3. Make sure that /rlm, /rlm-config and /rlm-stop show under [Extensions].
4. Push Ctrl+Shift+R or type /rlm. Your prompts now go to the RLM engine.

How it works

 you ──► prompt ──► RLM engine ──► main model (Python in repl(), one turn at a time)
                        │              ├─► search / grep / outline   (no tokens)
                        │              ├─► llm_query ──► cheap workers
                        │              └─► rlm_query ──► child RLMs (depth limit)
                        └──► state Σ + validation + limits (tokens, time, errors)

• The context stays in a Python sandbox. The model sees only an index.
• The main model writes Python code. The code finds the data that the model needs.
• Cheap worker models read and summarize. Your main model does not read the full text.
• A child RLM solves a difficult sub-task. It gets the full sandbox at no token cost.
• The engine keeps the run state Σ and validates each change to it.
• /compact makes a digest of the session. It does not use an LLM.

Benchmarks

The benchmark is OOLONG (oolong-synth): 24 tasks, context up to 65,536 tokens, temperature 0.

Model	OOLONG	Tokens/task	Cost/task
zai/glm-5.3	95.8% (23/24)	~28k	~$0.00 *
b-ai/qwen3.8-27b	91.7% (22/24)	~37k	~$0.00 *
openrouter/google/gemma-4-31b-it:free	pending	—	—

The model list is in bench/panel.json. Each model has its own results folder: bench/runs/<model>/. To run all models:

export ZAI_API_KEY=... BAI_API_KEY=... OPENROUTER_API_KEY=...
bun run bench --oolong-max-cl 65536 --oolong-limit 24

For more data, see bench/README.md.

Commands

Command	Shortcut	Function
/rlm	Ctrl+Shift+R	Set RLM mode on or off
/rlm-stop		Stop the current run
/rlm-config		Set sampling, budgets and limits

Sandbox API

The model can use only these retrieval and delegation functions:

search("needle", k=10)        # BM25 hits in the context
grep_context(pattern, k=50)   # regex hits with line numbers
outline(path)                 # definitions in a file
add_context("report.pdf")     # add a PDF, DOCX, XLSX, EPUB, CSV or HTML file
llm_query(prompt)             # send text to a cheap worker
map_files(paths, prompt)      # ask one question about many files
rlm_query(task, paths=...)    # start a child RLM
answer["content"] = "..."     # write the final answer
answer["ready"] = True        # end the run

Security

• The engine does not write to the disk. The answer is the only output of a run.
• API keys do not go into the sandbox. The host sends all sub-LLM calls.
• Python runs in a guarded subprocess with a list of reserved names.

License

MIT © hicaru contributors