@smartpointstech/sona

Self-Optimizing Neural Architecture (SONA) - Runtime-adaptive learning with LoRA, EWC++, and ReasoningBank for LLM routers and AI systems

Overview

This package provides native Node.js bindings to the spt-sona Rust crate. SONA implements runtime-adaptive learning through micro-LoRA adapters, Elastic Weight Consolidation (EWC++), and a ReasoningBank for pattern clustering. It is designed for LLM routers and agent systems that need to learn and adapt during execution.

Installation

npm install @smartpointstech/sona

Key Features

• SonaEngine -- main engine for adaptive learning with configurable hidden dimensions
• Micro-LoRA Adapters -- low-rank adaptation applied at runtime for fast fine-tuning
• EWC++ Regularization -- prevents catastrophic forgetting across task switches
• Pattern Clustering -- automatic discovery and retrieval of learned pattern types
• TrajectoryBuilder -- records execution trajectories for background learning

Dependencies

No internal npm dependencies. Uses platform-specific native binaries via optional dependencies (e.g., @smartpointstech/sona-darwin-arm64). Built with @napi-rs/cli.

Quick Start

import { SonaEngine, TrajectoryBuilder } from '@smartpointstech/sona';

const engine = new SonaEngine({ hiddenDim: 256, microLoraRank: 4 });

const trajectory = new TrajectoryBuilder()
  .setInput(inputVec)
  .setOutput(outputVec)
  .setQuality(0.85)
  .build();

engine.submitTrajectory(trajectory);
const patterns = engine.findPatterns(queryVec, 5);

API Reference

Exports

Export	Type	Description
SonaEngine	class	Main adaptive learning engine
TrajectoryBuilder	class	Builder for execution trajectory records
SonaConfig	interface	Engine configuration (dims, LoRA ranks, EWC lambda)
LearnedPattern	interface	Cluster with centroid, quality, access stats
PatternType	const	Pattern classifications: General, Reasoning, CodeGen, etc.
getImplementationType	function	Returns 'native' or 'stub'