avilum/minrlm

Implements a REPL-based execution model where the LLM generates Python code to query data directly, keeping raw context out of the prompt entirely; uses entropy profiling via zlib compression to identify relevant sections and task-specific routing to optimize code patterns for structured data, search, math, and code retrieval tasks. Wraps execution in a DockerREPL sandbox (seccomp, stdlib-only) and optionally delegates smaller sub-tasks to a secondary LLM on filtered evidence, achieving 30pp accuracy gains over vanilla on frontier models while maintaining flat token cost regardless of document size.

Used by 1 other package. Available on PyPI.