Files
tasq/node_modules/@claude-flow/neural/src/sona-integration.ts
T
2026-04-09 19:01:53 +08:00

433 lines
12 KiB
TypeScript

/**
* SONA Integration for V3 Neural Module
*
* Wraps @ruvector/sona package for V3 usage with:
* - Trajectory tracking and verdict judgment
* - Pattern extraction and memory distillation
* - Sub-0.05ms learning performance target
* - Clean TypeScript API with proper types
*
* @module sona-integration
*/
import { SonaEngine, type JsSonaConfig, type JsLearnedPattern } from '@ruvector/sona';
import type {
Trajectory,
TrajectoryStep,
TrajectoryVerdict,
DistilledMemory,
SONAMode,
SONAModeConfig,
} from './types.js';
// =============================================================================
// Types
// =============================================================================
/**
* Context for SONA learning adaptation
*/
export interface Context {
/** Task domain */
domain: 'code' | 'creative' | 'reasoning' | 'chat' | 'math' | 'general';
/** Current query embedding */
queryEmbedding: Float32Array;
/** Additional context metadata */
metadata?: Record<string, unknown>;
}
/**
* Adapted behavior result from SONA
*/
export interface AdaptedBehavior {
/** Transformed query embedding after micro-LoRA */
transformedQuery: Float32Array;
/** Similar learned patterns */
patterns: JsLearnedPattern[];
/** Suggested route/model */
suggestedRoute?: string;
/** Confidence score */
confidence: number;
}
/**
* SONA engine statistics
*/
export interface SONAStats {
/** Total trajectories recorded */
totalTrajectories: number;
/** Patterns learned */
patternsLearned: number;
/** Average quality */
avgQuality: number;
/** Last learning time (ms) */
lastLearningMs: number;
/** Engine enabled state */
enabled: boolean;
}
// =============================================================================
// Mode Configuration Mapping
// =============================================================================
/**
* Convert V3 SONA mode to @ruvector/sona config
*/
function modeToConfig(mode: SONAMode, modeConfig: SONAModeConfig): JsSonaConfig {
const baseConfig: JsSonaConfig = {
hiddenDim: 768, // Standard transformer dimension
embeddingDim: 768,
microLoraRank: modeConfig.loraRank <= 2 ? modeConfig.loraRank : 1,
baseLoraRank: modeConfig.loraRank,
microLoraLr: modeConfig.learningRate,
baseLoraLr: modeConfig.learningRate * 0.1,
ewcLambda: modeConfig.ewcLambda,
patternClusters: modeConfig.patternClusters,
trajectoryCapacity: modeConfig.trajectoryCapacity,
qualityThreshold: modeConfig.qualityThreshold,
enableSimd: true,
};
// Mode-specific adjustments
switch (mode) {
case 'real-time':
return {
...baseConfig,
microLoraRank: 1,
backgroundIntervalMs: 60000, // 1 minute
};
case 'edge':
return {
...baseConfig,
hiddenDim: 384, // Smaller for edge devices
embeddingDim: 384,
microLoraRank: 1,
patternClusters: 25,
backgroundIntervalMs: 300000, // 5 minutes
};
case 'research':
return {
...baseConfig,
baseLoraRank: 16,
backgroundIntervalMs: 3600000, // 1 hour
};
case 'batch':
return {
...baseConfig,
backgroundIntervalMs: 7200000, // 2 hours
};
case 'balanced':
default:
return {
...baseConfig,
backgroundIntervalMs: 1800000, // 30 minutes
};
}
}
// =============================================================================
// SONA Learning Engine
// =============================================================================
/**
* SONA Learning Engine - wraps @ruvector/sona for V3 usage
*
* Performance targets:
* - learn(): <0.05ms
* - adapt(): <0.1ms
* - Full learning cycle: <10ms
*/
export class SONALearningEngine {
private engine: SonaEngine;
private trajectoryMap: Map<string, number> = new Map();
private adaptationTimeMs: number = 0;
private learningTimeMs: number = 0;
private mode: SONAMode;
private modeConfig: SONAModeConfig;
constructor(mode: SONAMode, modeConfig: SONAModeConfig) {
this.mode = mode;
this.modeConfig = modeConfig;
const config = modeToConfig(mode, modeConfig);
this.engine = SonaEngine.withConfig(config);
}
/**
* Learn from a trajectory
*
* Performance target: <0.05ms
*
* @param trajectory - Trajectory to learn from
*/
async learn(trajectory: Trajectory): Promise<void> {
const startTime = performance.now();
try {
// Begin trajectory recording
const queryEmbedding = this.trajectoryToQueryEmbedding(trajectory);
const trajectoryId = this.engine.beginTrajectory(
Array.from(queryEmbedding)
);
// Add trajectory steps
for (const step of trajectory.steps) {
const activations = this.stateToActivations(step.stateBefore);
const attentionWeights = this.stateToAttentionWeights(step.stateAfter);
this.engine.addTrajectoryStep(
trajectoryId,
Array.from(activations),
Array.from(attentionWeights),
step.reward
);
}
// Set context if available
if (trajectory.domain) {
this.engine.addTrajectoryContext(trajectoryId, trajectory.domain);
}
// Complete trajectory with quality score
const quality = this.calculateQuality(trajectory);
this.engine.endTrajectory(trajectoryId, quality);
// Flush instant updates
this.engine.flush();
this.learningTimeMs = performance.now() - startTime;
} catch (error) {
throw new Error(`SONA learning failed: ${error}`);
}
}
/**
* Adapt behavior based on context
*
* @param context - Current context for adaptation
* @returns Adapted behavior with transformed embeddings
*/
async adapt(context: Context): Promise<AdaptedBehavior> {
const startTime = performance.now();
try {
// Apply micro-LoRA transformation
const transformedQuery = this.engine.applyMicroLora(
Array.from(context.queryEmbedding)
);
// Find similar patterns
const patterns = this.engine.findPatterns(
Array.from(context.queryEmbedding),
5
);
// Determine suggested route from patterns
const suggestedRoute = this.inferRoute(patterns, context);
const confidence = patterns.length > 0 ? patterns[0].avgQuality : 0.5;
this.adaptationTimeMs = performance.now() - startTime;
return {
transformedQuery: new Float32Array(transformedQuery),
patterns,
suggestedRoute,
confidence,
};
} catch (error) {
throw new Error(`SONA adaptation failed: ${error}`);
}
}
/**
* Get last adaptation time
*
* @returns Adaptation time in milliseconds
*/
getAdaptationTime(): number {
return this.adaptationTimeMs;
}
/**
* Get last learning time
*
* @returns Learning time in milliseconds
*/
getLearningTime(): number {
return this.learningTimeMs;
}
/**
* Reset learning state
*/
resetLearning(): void {
// Create a new engine with the same config
const config = modeToConfig(this.mode, this.modeConfig);
this.engine = SonaEngine.withConfig(config);
this.trajectoryMap.clear();
this.adaptationTimeMs = 0;
this.learningTimeMs = 0;
}
/**
* Force immediate learning cycle
*
* @returns Status message
*/
forceLearning(): string {
return this.engine.forceLearn();
}
/**
* Tick background learning (call periodically)
*
* @returns Status message if learning occurred
*/
tick(): string | null {
return this.engine.tick();
}
/**
* Get engine statistics
*
* @returns SONA engine statistics
*/
getStats(): SONAStats {
const statsJson = this.engine.getStats();
const stats = JSON.parse(statsJson);
return {
totalTrajectories: stats.total_trajectories || 0,
patternsLearned: stats.patterns_learned || 0,
avgQuality: stats.avg_quality || 0,
lastLearningMs: this.learningTimeMs,
enabled: this.engine.isEnabled(),
};
}
/**
* Enable or disable the engine
*
* @param enabled - Whether to enable the engine
*/
setEnabled(enabled: boolean): void {
this.engine.setEnabled(enabled);
}
/**
* Check if engine is enabled
*
* @returns Whether the engine is enabled
*/
isEnabled(): boolean {
return this.engine.isEnabled();
}
/**
* Find learned patterns similar to query
*
* @param queryEmbedding - Query embedding
* @param k - Number of patterns to return
* @returns Learned patterns
*/
findPatterns(queryEmbedding: Float32Array, k: number = 5): JsLearnedPattern[] {
return this.engine.findPatterns(Array.from(queryEmbedding), k);
}
// =============================================================================
// Private Helpers
// =============================================================================
/**
* Convert trajectory to query embedding
*/
private trajectoryToQueryEmbedding(trajectory: Trajectory): Float32Array {
// Use the first step's state as query
if (trajectory.steps.length > 0) {
return trajectory.steps[0].stateBefore;
}
// Fallback to zero embedding
return new Float32Array(768);
}
/**
* Convert state embedding to activations
*/
private stateToActivations(state: Float32Array): Float32Array {
// For now, use state directly as activations
// In a real implementation, this would extract layer activations
return state;
}
/**
* Convert state embedding to attention weights
*/
private stateToAttentionWeights(state: Float32Array): Float32Array {
// For now, use normalized state as attention weights
// In a real implementation, this would extract attention patterns
const sum = state.reduce((acc, val) => acc + Math.abs(val), 0);
if (sum === 0) return state;
const weights = new Float32Array(state.length);
for (let i = 0; i < state.length; i++) {
weights[i] = Math.abs(state[i]) / sum;
}
return weights;
}
/**
* Calculate quality score for trajectory
*/
private calculateQuality(trajectory: Trajectory): number {
if (trajectory.qualityScore !== undefined) {
return trajectory.qualityScore;
}
// Calculate from steps
if (trajectory.steps.length === 0) return 0.5;
const avgReward = trajectory.steps.reduce((sum, step) => sum + step.reward, 0) /
trajectory.steps.length;
// Normalize to [0, 1]
return Math.max(0, Math.min(1, (avgReward + 1) / 2));
}
/**
* Infer suggested route from patterns and context
*/
private inferRoute(patterns: JsLearnedPattern[], context: Context): string | undefined {
if (patterns.length === 0) return undefined;
// Use the highest quality pattern's type as route
const bestPattern = patterns.reduce((best, pattern) =>
pattern.avgQuality > best.avgQuality ? pattern : best
);
return bestPattern.patternType || `${context.domain}-default`;
}
}
// =============================================================================
// Factory Functions
// =============================================================================
/**
* Create a SONA learning engine
*
* @param mode - SONA learning mode
* @param modeConfig - Mode configuration
* @returns SONA learning engine instance
*/
export function createSONALearningEngine(
mode: SONAMode,
modeConfig: SONAModeConfig
): SONALearningEngine {
return new SONALearningEngine(mode, modeConfig);
}
// =============================================================================
// Exports
// =============================================================================
export type { JsLearnedPattern, JsSonaConfig };