adaptive-memory-multi-model-router 2.2.9 → 2.4.0

package/dist/cost/preCallCostEstimator.d.ts

@@ -1,114 +0,0 @@
-/**
- * A3M Router - Pre-Call Cost Estimator
- *
- * Estimates cost BEFORE making an API call based on input features.
- * Uses historical data patterns and token count estimation via character ratio.
- *
- * No external API calls - all estimation is local.
- *
- * Usage:
- *   const estimator = new PreCallCostEstimator();
- *   const estimate = estimator.estimate({
- *     query: "What is Python?",
- *     modelTier: 'mid',
- *     provider: 'groq'
- *   });
- *   console.log(estimate); // { estimatedTokens: 24, estimatedCost: 0.0014, estimatedLatency: 200 }
- */
-import { ProviderTier } from '../providers/providerConfig';
-export interface CostEstimateInput {
-    /** Query text to estimate tokens for */
-    query: string;
-    /** Model tier (free|cheap|mid|premium|enterprise) */
-    modelTier?: ProviderTier;
-    /** Provider ID for provider-specific cost lookup */
-    provider?: string;
-    /** Optional: explicit token count (if already known) */
-    explicitTokens?: number;
-    /** System prompt length (if using a system prompt) */
-    systemPromptLength?: number;
-}
-export interface CostEstimate {
-    /** Estimated total tokens (input + output) */
-    estimatedTokens: number;
-    /** Estimated input tokens */
-    estimatedInputTokens: number;
-    /** Estimated output tokens */
-    estimatedOutputTokens: number;
-    /** Estimated cost in USD */
-    estimatedCost: number;
-    /** Estimated latency in milliseconds */
-    estimatedLatency: number;
-    /** Confidence score 0-1 */
-    confidence: number;
-    /** Breakdown of estimation */
-    breakdown: {
-        inputCostPerM: number;
-        outputCostPerM: number;
-        charToTokenRatio: number;
-    };
-}
-export interface ProviderCostConfig {
-    input: number;
-    output: number;
-}
-export declare class PreCallCostEstimator {
-    private historicalData;
-    private slope;
-    private intercept;
-    private latencyHistory;
-    private ewmaAlpha;
-    constructor(historicalData?: Array<{
-        chars: number;
-        tokens: number;
-    }>);
-    /**
-     * Main estimation method - estimates tokens, cost, and latency.
-     */
-    estimate(input: CostEstimateInput): CostEstimate;
-    /**
-     * Estimate input tokens using character-to-token ratio.
-     * Uses linear regression if historical data is available.
-     */
-    estimateTokens(text: string, systemPromptLength?: number): number;
-    /**
-     * Estimate output tokens based on query complexity.
-     * More complex queries (code, analysis) tend to need more output.
-     */
-    estimateOutputTokens(query: string, inputTokens: number): number;
-    /**
-     * Calculate cost in USD.
-     */
-    calculateCost(totalTokens: number, config: ProviderCostConfig): number;
-    /**
-     * Estimate latency in milliseconds.
-     */
-    estimateLatency(tier: ProviderTier, inputTokens: number, totalTokens: number): number;
-    /**
-     * Record actual tokens for future regression improvements.
-     */
-    recordActualTokens(queryLength: number, actualTokens: number): void;
-    /**
-     * Record actual latency for EWMA updates.
-     */
-    recordActualLatency(latencyMs: number): void;
-    /**
-     * Fit linear regression to historical data.
-     * Uses ordinary least squares.
-     */
-    fitLinearRegression(data: Array<{
-        chars: number;
-        tokens: number;
-    }>): void;
-    private getCostConfig;
-    private containsKeyword;
-    private calculateEwma;
-    private calculateConfidence;
-    private estimateInputTokensFromTotal;
-    private estimateOutputTokensFromTotal;
-}
-export declare function createPreCallCostEstimator(historicalData?: Array<{
-    chars: number;
-    tokens: number;
-}>): PreCallCostEstimator;
-//# sourceMappingURL=preCallCostEstimator.d.ts.map

package/dist/cost/preCallCostEstimator.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"preCallCostEstimator.d.ts","sourceRoot":"","sources":["../../src/cost/preCallCostEstimator.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;GAgBG;AAEH,OAAO,EAAE,YAAY,EAAE,MAAM,6BAA6B,CAAC;AAM3D,MAAM,WAAW,iBAAiB;IAChC,wCAAwC;IACxC,KAAK,EAAE,MAAM,CAAC;IACd,qDAAqD;IACrD,SAAS,CAAC,EAAE,YAAY,CAAC;IACzB,oDAAoD;IACpD,QAAQ,CAAC,EAAE,MAAM,CAAC;IAClB,wDAAwD;IACxD,cAAc,CAAC,EAAE,MAAM,CAAC;IACxB,sDAAsD;IACtD,kBAAkB,CAAC,EAAE,MAAM,CAAC;CAC7B;AAED,MAAM,WAAW,YAAY;IAC3B,8CAA8C;IAC9C,eAAe,EAAE,MAAM,CAAC;IACxB,6BAA6B;IAC7B,oBAAoB,EAAE,MAAM,CAAC;IAC7B,8BAA8B;IAC9B,qBAAqB,EAAE,MAAM,CAAC;IAC9B,4BAA4B;IAC5B,aAAa,EAAE,MAAM,CAAC;IACtB,wCAAwC;IACxC,gBAAgB,EAAE,MAAM,CAAC;IACzB,2BAA2B;IAC3B,UAAU,EAAE,MAAM,CAAC;IACnB,8BAA8B;IAC9B,SAAS,EAAE;QACT,aAAa,EAAE,MAAM,CAAC;QACtB,cAAc,EAAE,MAAM,CAAC;QACvB,gBAAgB,EAAE,MAAM,CAAC;KAC1B,CAAC;CACH;AAED,MAAM,WAAW,kBAAkB;IACjC,KAAK,EAAE,MAAM,CAAC;IACd,MAAM,EAAE,MAAM,CAAC;CAChB;AAeD,qBAAa,oBAAoB;IAE/B,OAAO,CAAC,cAAc,CAAgD;IAEtE,OAAO,CAAC,KAAK,CAAQ;IACrB,OAAO,CAAC,SAAS,CAAM;IAEvB,OAAO,CAAC,cAAc,CAAgB;IACtC,OAAO,CAAC,SAAS,CAAO;gBAEZ,cAAc,CAAC,EAAE,KAAK,CAAC;QAAE,KAAK,EAAE,MAAM,CAAC;QAAC,MAAM,EAAE,MAAM,CAAA;KAAE,CAAC;IAMrE;;OAEG;IACH,QAAQ,CAAC,KAAK,EAAE,iBAAiB,GAAG,YAAY;IA+ChD;;;OAGG;IACH,cAAc,CAAC,IAAI,EAAE,MAAM,EAAE,kBAAkB,SAAI,GAAG,MAAM;IAY5D;;;OAGG;IACH,oBAAoB,CAAC,KAAK,EAAE,MAAM,EAAE,WAAW,EAAE,MAAM,GAAG,MAAM;IAwBhE;;OAEG;IACH,aAAa,CAAC,WAAW,EAAE,MAAM,EAAE,MAAM,EAAE,kBAAkB,GAAG,MAAM;IAMtE;;OAEG;IACH,eAAe,CAAC,IAAI,EAAE,YAAY,EAAE,WAAW,EAAE,MAAM,EAAE,WAAW,EAAE,MAAM,GAAG,MAAM;IAmBrF;;OAEG;IACH,kBAAkB,CAAC,WAAW,EAAE,MAAM,EAAE,YAAY,EAAE,MAAM,GAAG,IAAI;IAcnE;;OAEG;IACH,mBAAmB,CAAC,SAAS,EAAE,MAAM,GAAG,IAAI;IAO5C;;;OAGG;IACH,mBAAmB,CAAC,IAAI,EAAE,KAAK,CAAC;QAAE,KAAK,EAAE,MAAM,CAAC;QAAC,MAAM,EAAE,MAAM,CAAA;KAAE,CAAC,GAAG,IAAI;IA4BzE,OAAO,CAAC,aAAa;IA8BrB,OAAO,CAAC,eAAe;IAIvB,OAAO,CAAC,aAAa;IASrB,OAAO,CAAC,mBAAmB;IAQ3B,OAAO,CAAC,4BAA4B;IAIpC,OAAO,CAAC,6BAA6B;CAGtC;AAMD,wBAAgB,0BAA0B,CACxC,cAAc,CAAC,EAAE,KAAK,CAAC;IAAE,KAAK,EAAE,MAAM,CAAC;IAAC,MAAM,EAAE,MAAM,CAAA;CAAE,CAAC,GACxD,oBAAoB,CAEtB"}

package/dist/cost/preCallCostEstimator.js

@@ -1,256 +0,0 @@
-"use strict";
-/**
- * A3M Router - Pre-Call Cost Estimator
- *
- * Estimates cost BEFORE making an API call based on input features.
- * Uses historical data patterns and token count estimation via character ratio.
- *
- * No external API calls - all estimation is local.
- *
- * Usage:
- *   const estimator = new PreCallCostEstimator();
- *   const estimate = estimator.estimate({
- *     query: "What is Python?",
- *     modelTier: 'mid',
- *     provider: 'groq'
- *   });
- *   console.log(estimate); // { estimatedTokens: 24, estimatedCost: 0.0014, estimatedLatency: 200 }
- */
-Object.defineProperty(exports, "__esModule", { value: true });
-exports.PreCallCostEstimator = void 0;
-exports.createPreCallCostEstimator = createPreCallCostEstimator;
-// Default latency estimates per tier (ms)
-const TIER_LATENCY = {
-    free: { min: 500, max: 5000, avg: 2000 },
-    cheap: { min: 100, max: 800, avg: 300 },
-    mid: { min: 200, max: 1500, avg: 600 },
-    premium: { min: 300, max: 2000, avg: 800 },
-    enterprise: { min: 200, max: 1500, avg: 500 },
-};
-// ============================================================
-// PreCallCostEstimator
-// ============================================================
-class PreCallCostEstimator {
-    // Historical data for regression: [charCount, actualTokens][] tuples
-    historicalData = [];
-    // Linear regression coefficients
-    slope = 0.25; // chars per token ratio
-    intercept = 10;
-    // EWMA for latency estimation
-    latencyHistory = [];
-    ewmaAlpha = 0.3;
-    constructor(historicalData) {
-        if (historicalData && historicalData.length > 0) {
-            this.fitLinearRegression(historicalData);
-        }
-    }
-    /**
-     * Main estimation method - estimates tokens, cost, and latency.
-     */
-    estimate(input) {
-        const { query, modelTier = 'mid', provider, explicitTokens, systemPromptLength = 0, } = input;
-        // Token estimation
-        let inputTokens;
-        if (explicitTokens !== undefined) {
-            inputTokens = explicitTokens;
-        }
-        else {
-            inputTokens = this.estimateTokens(query, systemPromptLength);
-        }
-        // Output token estimation based on query complexity
-        const outputTokens = this.estimateOutputTokens(query, inputTokens);
-        const totalTokens = inputTokens + outputTokens;
-        // Cost estimation
-        const costConfig = this.getCostConfig(provider, modelTier);
-        const estimatedCost = this.calculateCost(totalTokens, costConfig);
-        // Latency estimation
-        const estimatedLatency = this.estimateLatency(modelTier, inputTokens, totalTokens);
-        // Confidence based on amount of historical data we have
-        const confidence = this.calculateConfidence();
-        return {
-            estimatedTokens: totalTokens,
-            estimatedInputTokens: inputTokens,
-            estimatedOutputTokens: outputTokens,
-            estimatedCost: Math.round(estimatedCost * 1000000) / 1000000, // 6 decimal places
-            estimatedLatency,
-            confidence,
-            breakdown: {
-                inputCostPerM: costConfig.input,
-                outputCostPerM: costConfig.output,
-                charToTokenRatio: this.slope,
-            },
-        };
-    }
-    /**
-     * Estimate input tokens using character-to-token ratio.
-     * Uses linear regression if historical data is available.
-     */
-    estimateTokens(text, systemPromptLength = 0) {
-        const totalChars = text.length + systemPromptLength;
-        if (this.historicalData.length >= 5) {
-            // Use linear regression
-            return Math.max(1, Math.round(this.slope * totalChars + this.intercept));
-        }
-        // Fallback: general English average ~4 chars per token
-        return Math.max(1, Math.round(totalChars / 4));
-    }
-    /**
-     * Estimate output tokens based on query complexity.
-     * More complex queries (code, analysis) tend to need more output.
-     */
-    estimateOutputTokens(query, inputTokens) {
-        const lower = query.toLowerCase();
-        // Base estimate: ~30% of input tokens as output
-        let multiplier = 0.3;
-        // Complexity adjustments
-        if (this.containsKeyword(lower, ['code', 'implement', 'function', 'class', 'algorithm'])) {
-            multiplier = 0.5; // Code needs more output
-        }
-        else if (this.containsKeyword(lower, ['explain', 'describe', 'what is', 'how does'])) {
-            multiplier = 0.35; // Explanations need moderate output
-        }
-        else if (this.containsKeyword(lower, ['list', 'count', 'find all'])) {
-            multiplier = 0.4; // List queries need more output
-        }
-        else if (this.containsKeyword(lower, ['yes', 'no', 'is', 'are', 'does'])) {
-            multiplier = 0.1; // Simple questions need minimal output
-        }
-        // Cap to reasonable bounds
-        return Math.min(Math.max(10, Math.round(inputTokens * multiplier)), 4000 // Max 4k output tokens
-        );
-    }
-    /**
-     * Calculate cost in USD.
-     */
-    calculateCost(totalTokens, config) {
-        const inputM = this.estimateInputTokensFromTotal(totalTokens) / 1_000_000;
-        const outputM = this.estimateOutputTokensFromTotal(totalTokens) / 1_000_000;
-        return inputM * config.input + outputM * config.output;
-    }
-    /**
-     * Estimate latency in milliseconds.
-     */
-    estimateLatency(tier, inputTokens, totalTokens) {
-        const tierLatency = TIER_LATENCY[tier];
-        // Base latency from tier
-        let latency = tierLatency.avg;
-        // Scale by token count (rough linear approximation)
-        const tokenScale = totalTokens / 100;
-        latency *= Math.max(0.5, Math.min(3, tokenScale));
-        // Adjust for historical EWMA if available
-        if (this.latencyHistory.length > 0) {
-            const ewmaLatency = this.calculateEwma();
-            latency = latency * 0.7 + ewmaLatency * 0.3;
-        }
-        return Math.round(latency);
-    }
-    /**
-     * Record actual tokens for future regression improvements.
-     */
-    recordActualTokens(queryLength, actualTokens) {
-        this.historicalData.push({ chars: queryLength, tokens: actualTokens });
-        // Keep only last 100 data points
-        if (this.historicalData.length > 100) {
-            this.historicalData.shift();
-        }
-        // Refit regression periodically
-        if (this.historicalData.length % 10 === 0) {
-            this.fitLinearRegression(this.historicalData);
-        }
-    }
-    /**
-     * Record actual latency for EWMA updates.
-     */
-    recordActualLatency(latencyMs) {
-        this.latencyHistory.push(latencyMs);
-        if (this.latencyHistory.length > 50) {
-            this.latencyHistory.shift();
-        }
-    }
-    /**
-     * Fit linear regression to historical data.
-     * Uses ordinary least squares.
-     */
-    fitLinearRegression(data) {
-        if (data.length < 2)
-            return;
-        const n = data.length;
-        let sumX = 0, sumY = 0, sumXY = 0, sumX2 = 0;
-        for (const { chars, tokens } of data) {
-            sumX += chars;
-            sumY += tokens;
-            sumXY += chars * tokens;
-            sumX2 += chars * chars;
-        }
-        const denominator = n * sumX2 - sumX * sumX;
-        if (denominator === 0)
-            return;
-        this.slope = (n * sumXY - sumX * sumY) / denominator;
-        this.intercept = (sumY - this.slope * sumX) / n;
-        // Sanity check
-        if (this.slope <= 0 || this.slope > 1) {
-            this.slope = 0.25; // Reset to default if outlier
-            this.intercept = 10;
-        }
-    }
-    // ---- Private helpers ----
-    getCostConfig(provider, tier) {
-        // Provider-specific costs (from providerConfig.ts patterns)
-        const providerCosts = {
-            groq: { input: 0.59, output: 0.79 },
-            cerebras: { input: 0.60, output: 0.60 },
-            deepseek: { input: 0.14, output: 0.28 },
-            deepinfra: { input: 0.05, output: 0.05 },
-            together: { input: 0.18, output: 0.18 },
-            fireworks: { input: 0.20, output: 0.20 },
-            mistral: { input: 0.20, output: 0.60 },
-            openai: { input: 2.50, output: 10.00 },
-            anthropic: { input: 3.00, output: 15.00 },
-        };
-        if (provider && providerCosts[provider]) {
-            return providerCosts[provider];
-        }
-        // Tier fallback
-        const tierCosts = {
-            free: { input: 0, output: 0 },
-            cheap: { input: 0.20, output: 0.40 },
-            mid: { input: 1.00, output: 3.00 },
-            premium: { input: 3.00, output: 12.00 },
-            enterprise: { input: 5.00, output: 20.00 },
-        };
-        return tier ? tierCosts[tier] : tierCosts.mid;
-    }
-    containsKeyword(text, keywords) {
-        return keywords.some(kw => text.includes(kw));
-    }
-    calculateEwma() {
-        if (this.latencyHistory.length === 0)
-            return 0;
-        let ewma = this.latencyHistory[0];
-        for (let i = 1; i < this.latencyHistory.length; i++) {
-            ewma = this.ewmaAlpha * this.latencyHistory[i] + (1 - this.ewmaAlpha) * ewma;
-        }
-        return ewma;
-    }
-    calculateConfidence() {
-        // More historical data = higher confidence
-        const dataFactor = Math.min(this.historicalData.length / 50, 1.0);
-        // More latency history = higher confidence
-        const latencyFactor = Math.min(this.latencyHistory.length / 20, 1.0);
-        return Math.round((dataFactor * 0.6 + latencyFactor * 0.4) * 100) / 100;
-    }
-    estimateInputTokensFromTotal(total) {
-        return Math.round(total * 0.7); // Assume 70% input
-    }
-    estimateOutputTokensFromTotal(total) {
-        return Math.round(total * 0.3); // Assume 30% output
-    }
-}
-exports.PreCallCostEstimator = PreCallCostEstimator;
-// ============================================================
-// Factory
-// ============================================================
-function createPreCallCostEstimator(historicalData) {
-    return new PreCallCostEstimator(historicalData);
-}
-//# sourceMappingURL=preCallCostEstimator.js.map

package/dist/cost/preCallCostEstimator.js.map

@@ -1 +0,0 @@
-{"version":3,"file":"preCallCostEstimator.js","sourceRoot":"","sources":["../../src/cost/preCallCostEstimator.ts"],"names":[],"mappings":";AAAA;;;;;;;;;;;;;;;;GAgBG;;;AAoUH,gEAIC;AAzRD,0CAA0C;AAC1C,MAAM,YAAY,GAAoE;IACpF,IAAI,EAAI,EAAE,GAAG,EAAE,GAAG,EAAG,GAAG,EAAE,IAAI,EAAG,GAAG,EAAE,IAAI,EAAE;IAC5C,KAAK,EAAG,EAAE,GAAG,EAAE,GAAG,EAAG,GAAG,EAAE,GAAG,EAAI,GAAG,EAAE,GAAG,EAAE;IAC3C,GAAG,EAAK,EAAE,GAAG,EAAE,GAAG,EAAG,GAAG,EAAE,IAAI,EAAG,GAAG,EAAE,GAAG,EAAE;IAC3C,OAAO,EAAC,EAAE,GAAG,EAAE,GAAG,EAAG,GAAG,EAAE,IAAI,EAAG,GAAG,EAAE,GAAG,EAAE;IAC3C,UAAU,EAAE,EAAE,GAAG,EAAE,GAAG,EAAE,GAAG,EAAE,IAAI,EAAE,GAAG,EAAE,GAAG,EAAE;CAC9C,CAAC;AAEF,+DAA+D;AAC/D,uBAAuB;AACvB,+DAA+D;AAE/D,MAAa,oBAAoB;IAC/B,qEAAqE;IAC7D,cAAc,GAA6C,EAAE,CAAC;IACtE,iCAAiC;IACzB,KAAK,GAAG,IAAI,CAAC,CAAC,wBAAwB;IACtC,SAAS,GAAG,EAAE,CAAC;IACvB,8BAA8B;IACtB,cAAc,GAAa,EAAE,CAAC;IAC9B,SAAS,GAAG,GAAG,CAAC;IAExB,YAAY,cAAyD;QACnE,IAAI,cAAc,IAAI,cAAc,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;YAChD,IAAI,CAAC,mBAAmB,CAAC,cAAc,CAAC,CAAC;QAC3C,CAAC;IACH,CAAC;IAED;;OAEG;IACH,QAAQ,CAAC,KAAwB;QAC/B,MAAM,EACJ,KAAK,EACL,SAAS,GAAG,KAAK,EACjB,QAAQ,EACR,cAAc,EACd,kBAAkB,GAAG,CAAC,GACvB,GAAG,KAAK,CAAC;QAEV,mBAAmB;QACnB,IAAI,WAAmB,CAAC;QACxB,IAAI,cAAc,KAAK,SAAS,EAAE,CAAC;YACjC,WAAW,GAAG,cAAc,CAAC;QAC/B,CAAC;aAAM,CAAC;YACN,WAAW,GAAG,IAAI,CAAC,cAAc,CAAC,KAAK,EAAE,kBAAkB,CAAC,CAAC;QAC/D,CAAC;QAED,oDAAoD;QACpD,MAAM,YAAY,GAAG,IAAI,CAAC,oBAAoB,CAAC,KAAK,EAAE,WAAW,CAAC,CAAC;QAEnE,MAAM,WAAW,GAAG,WAAW,GAAG,YAAY,CAAC;QAE/C,kBAAkB;QAClB,MAAM,UAAU,GAAG,IAAI,CAAC,aAAa,CAAC,QAAQ,EAAE,SAAS,CAAC,CAAC;QAC3D,MAAM,aAAa,GAAG,IAAI,CAAC,aAAa,CAAC,WAAW,EAAE,UAAU,CAAC,CAAC;QAElE,qBAAqB;QACrB,MAAM,gBAAgB,GAAG,IAAI,CAAC,eAAe,CAAC,SAAS,EAAE,WAAW,EAAE,WAAW,CAAC,CAAC;QAEnF,wDAAwD;QACxD,MAAM,UAAU,GAAG,IAAI,CAAC,mBAAmB,EAAE,CAAC;QAE9C,OAAO;YACL,eAAe,EAAE,WAAW;YAC5B,oBAAoB,EAAE,WAAW;YACjC,qBAAqB,EAAE,YAAY;YACnC,aAAa,EAAE,IAAI,CAAC,KAAK,CAAC,aAAa,GAAG,OAAO,CAAC,GAAG,OAAO,EAAE,mBAAmB;YACjF,gBAAgB;YAChB,UAAU;YACV,SAAS,EAAE;gBACT,aAAa,EAAE,UAAU,CAAC,KAAK;gBAC/B,cAAc,EAAE,UAAU,CAAC,MAAM;gBACjC,gBAAgB,EAAE,IAAI,CAAC,KAAK;aAC7B;SACF,CAAC;IACJ,CAAC;IAED;;;OAGG;IACH,cAAc,CAAC,IAAY,EAAE,kBAAkB,GAAG,CAAC;QACjD,MAAM,UAAU,GAAG,IAAI,CAAC,MAAM,GAAG,kBAAkB,CAAC;QAEpD,IAAI,IAAI,CAAC,cAAc,CAAC,MAAM,IAAI,CAAC,EAAE,CAAC;YACpC,wBAAwB;YACxB,OAAO,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,KAAK,GAAG,UAAU,GAAG,IAAI,CAAC,SAAS,CAAC,CAAC,CAAC;QAC3E,CAAC;QAED,uDAAuD;QACvD,OAAO,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,IAAI,CAAC,KAAK,CAAC,UAAU,GAAG,CAAC,CAAC,CAAC,CAAC;IACjD,CAAC;IAED;;;OAGG;IACH,oBAAoB,CAAC,KAAa,EAAE,WAAmB;QACrD,MAAM,KAAK,GAAG,KAAK,CAAC,WAAW,EAAE,CAAC;QAElC,gDAAgD;QAChD,IAAI,UAAU,GAAG,GAAG,CAAC;QAErB,yBAAyB;QACzB,IAAI,IAAI,CAAC,eAAe,CAAC,KAAK,EAAE,CAAC,MAAM,EAAE,WAAW,EAAE,UAAU,EAAE,OAAO,EAAE,WAAW,CAAC,CAAC,EAAE,CAAC;YACzF,UAAU,GAAG,GAAG,CAAC,CAAC,yBAAyB;QAC7C,CAAC;aAAM,IAAI,IAAI,CAAC,eAAe,CAAC,KAAK,EAAE,CAAC,SAAS,EAAE,UAAU,EAAE,SAAS,EAAE,UAAU,CAAC,CAAC,EAAE,CAAC;YACvF,UAAU,GAAG,IAAI,CAAC,CAAC,oCAAoC;QACzD,CAAC;aAAM,IAAI,IAAI,CAAC,eAAe,CAAC,KAAK,EAAE,CAAC,MAAM,EAAE,OAAO,EAAE,UAAU,CAAC,CAAC,EAAE,CAAC;YACtE,UAAU,GAAG,GAAG,CAAC,CAAC,gCAAgC;QACpD,CAAC;aAAM,IAAI,IAAI,CAAC,eAAe,CAAC,KAAK,EAAE,CAAC,KAAK,EAAE,IAAI,EAAE,IAAI,EAAE,KAAK,EAAE,MAAM,CAAC,CAAC,EAAE,CAAC;YAC3E,UAAU,GAAG,GAAG,CAAC,CAAC,uCAAuC;QAC3D,CAAC;QAED,2BAA2B;QAC3B,OAAO,IAAI,CAAC,GAAG,CACb,IAAI,CAAC,GAAG,CAAC,EAAE,EAAE,IAAI,CAAC,KAAK,CAAC,WAAW,GAAG,UAAU,CAAC,CAAC,EAClD,IAAI,CAAC,uBAAuB;SAC7B,CAAC;IACJ,CAAC;IAED;;OAEG;IACH,aAAa,CAAC,WAAmB,EAAE,MAA0B;QAC3D,MAAM,MAAM,GAAG,IAAI,CAAC,4BAA4B,CAAC,WAAW,CAAC,GAAG,SAAS,CAAC;QAC1E,MAAM,OAAO,GAAG,IAAI,CAAC,6BAA6B,CAAC,WAAW,CAAC,GAAG,SAAS,CAAC;QAC5E,OAAO,MAAM,GAAG,MAAM,CAAC,KAAK,GAAG,OAAO,GAAG,MAAM,CAAC,MAAM,CAAC;IACzD,CAAC;IAED;;OAEG;IACH,eAAe,CAAC,IAAkB,EAAE,WAAmB,EAAE,WAAmB;QAC1E,MAAM,WAAW,GAAG,YAAY,CAAC,IAAI,CAAC,CAAC;QAEvC,yBAAyB;QACzB,IAAI,OAAO,GAAG,WAAW,CAAC,GAAG,CAAC;QAE9B,oDAAoD;QACpD,MAAM,UAAU,GAAG,WAAW,GAAG,GAAG,CAAC;QACrC,OAAO,IAAI,IAAI,CAAC,GAAG,CAAC,GAAG,EAAE,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,UAAU,CAAC,CAAC,CAAC;QAElD,0CAA0C;QAC1C,IAAI,IAAI,CAAC,cAAc,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;YACnC,MAAM,WAAW,GAAG,IAAI,CAAC,aAAa,EAAE,CAAC;YACzC,OAAO,GAAG,OAAO,GAAG,GAAG,GAAG,WAAW,GAAG,GAAG,CAAC;QAC9C,CAAC;QAED,OAAO,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC;IAC7B,CAAC;IAED;;OAEG;IACH,kBAAkB,CAAC,WAAmB,EAAE,YAAoB;QAC1D,IAAI,CAAC,cAAc,CAAC,IAAI,CAAC,EAAE,KAAK,EAAE,WAAW,EAAE,MAAM,EAAE,YAAY,EAAE,CAAC,CAAC;QAEvE,iCAAiC;QACjC,IAAI,IAAI,CAAC,cAAc,CAAC,MAAM,GAAG,GAAG,EAAE,CAAC;YACrC,IAAI,CAAC,cAAc,CAAC,KAAK,EAAE,CAAC;QAC9B,CAAC;QAED,gCAAgC;QAChC,IAAI,IAAI,CAAC,cAAc,CAAC,MAAM,GAAG,EAAE,KAAK,CAAC,EAAE,CAAC;YAC1C,IAAI,CAAC,mBAAmB,CAAC,IAAI,CAAC,cAAc,CAAC,CAAC;QAChD,CAAC;IACH,CAAC;IAED;;OAEG;IACH,mBAAmB,CAAC,SAAiB;QACnC,IAAI,CAAC,cAAc,CAAC,IAAI,CAAC,SAAS,CAAC,CAAC;QACpC,IAAI,IAAI,CAAC,cAAc,CAAC,MAAM,GAAG,EAAE,EAAE,CAAC;YACpC,IAAI,CAAC,cAAc,CAAC,KAAK,EAAE,CAAC;QAC9B,CAAC;IACH,CAAC;IAED;;;OAGG;IACH,mBAAmB,CAAC,IAA8C;QAChE,IAAI,IAAI,CAAC,MAAM,GAAG,CAAC;YAAE,OAAO;QAE5B,MAAM,CAAC,GAAG,IAAI,CAAC,MAAM,CAAC;QACtB,IAAI,IAAI,GAAG,CAAC,EAAE,IAAI,GAAG,CAAC,EAAE,KAAK,GAAG,CAAC,EAAE,KAAK,GAAG,CAAC,CAAC;QAE7C,KAAK,MAAM,EAAE,KAAK,EAAE,MAAM,EAAE,IAAI,IAAI,EAAE,CAAC;YACrC,IAAI,IAAI,KAAK,CAAC;YACd,IAAI,IAAI,MAAM,CAAC;YACf,KAAK,IAAI,KAAK,GAAG,MAAM,CAAC;YACxB,KAAK,IAAI,KAAK,GAAG,KAAK,CAAC;QACzB,CAAC;QAED,MAAM,WAAW,GAAG,CAAC,GAAG,KAAK,GAAG,IAAI,GAAG,IAAI,CAAC;QAC5C,IAAI,WAAW,KAAK,CAAC;YAAE,OAAO;QAE9B,IAAI,CAAC,KAAK,GAAG,CAAC,CAAC,GAAG,KAAK,GAAG,IAAI,GAAG,IAAI,CAAC,GAAG,WAAW,CAAC;QACrD,IAAI,CAAC,SAAS,GAAG,CAAC,IAAI,GAAG,IAAI,CAAC,KAAK,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC;QAEhD,eAAe;QACf,IAAI,IAAI,CAAC,KAAK,IAAI,CAAC,IAAI,IAAI,CAAC,KAAK,GAAG,CAAC,EAAE,CAAC;YACtC,IAAI,CAAC,KAAK,GAAG,IAAI,CAAC,CAAC,8BAA8B;YACjD,IAAI,CAAC,SAAS,GAAG,EAAE,CAAC;QACtB,CAAC;IACH,CAAC;IAED,4BAA4B;IAEpB,aAAa,CAAC,QAAiB,EAAE,IAAmB;QAC1D,4DAA4D;QAC5D,MAAM,aAAa,GAAgD;YACjE,IAAI,EAAO,EAAE,KAAK,EAAE,IAAI,EAAG,MAAM,EAAE,IAAI,EAAE;YACzC,QAAQ,EAAG,EAAE,KAAK,EAAE,IAAI,EAAG,MAAM,EAAE,IAAI,EAAE;YACzC,QAAQ,EAAG,EAAE,KAAK,EAAE,IAAI,EAAG,MAAM,EAAE,IAAI,EAAE;YACzC,SAAS,EAAE,EAAE,KAAK,EAAE,IAAI,EAAG,MAAM,EAAE,IAAI,EAAE;YACzC,QAAQ,EAAG,EAAE,KAAK,EAAE,IAAI,EAAG,MAAM,EAAE,IAAI,EAAE;YACzC,SAAS,EAAE,EAAE,KAAK,EAAE,IAAI,EAAG,MAAM,EAAE,IAAI,EAAE;YACzC,OAAO,EAAI,EAAE,KAAK,EAAE,IAAI,EAAG,MAAM,EAAE,IAAI,EAAE;YACzC,MAAM,EAAK,EAAE,KAAK,EAAE,IAAI,EAAG,MAAM,EAAE,KAAK,EAAE;YAC1C,SAAS,EAAE,EAAE,KAAK,EAAE,IAAI,EAAG,MAAM,EAAE,KAAK,EAAE;SAC3C,CAAC;QAEF,IAAI,QAAQ,IAAI,aAAa,CAAC,QAAQ,CAAC,EAAE,CAAC;YACxC,OAAO,aAAa,CAAC,QAAQ,CAAE,CAAC;QAClC,CAAC;QAED,gBAAgB;QAChB,MAAM,SAAS,GAA6C;YAC1D,IAAI,EAAQ,EAAE,KAAK,EAAE,CAAC,EAAM,MAAM,EAAE,CAAC,EAAE;YACvC,KAAK,EAAO,EAAE,KAAK,EAAE,IAAI,EAAG,MAAM,EAAE,IAAI,EAAE;YAC1C,GAAG,EAAS,EAAE,KAAK,EAAE,IAAI,EAAG,MAAM,EAAE,IAAI,EAAE;YAC1C,OAAO,EAAK,EAAE,KAAK,EAAE,IAAI,EAAG,MAAM,EAAE,KAAK,EAAE;YAC3C,UAAU,EAAE,EAAE,KAAK,EAAE,IAAI,EAAG,MAAM,EAAE,KAAK,EAAE;SAC5C,CAAC;QAEF,OAAO,IAAI,CAAC,CAAC,CAAC,SAAS,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,SAAS,CAAC,GAAG,CAAC;IAChD,CAAC;IAEO,eAAe,CAAC,IAAY,EAAE,QAAkB;QACtD,OAAO,QAAQ,CAAC,IAAI,CAAC,EAAE,CAAC,EAAE,CAAC,IAAI,CAAC,QAAQ,CAAC,EAAE,CAAC,CAAC,CAAC;IAChD,CAAC;IAEO,aAAa;QACnB,IAAI,IAAI,CAAC,cAAc,CAAC,MAAM,KAAK,CAAC;YAAE,OAAO,CAAC,CAAC;QAC/C,IAAI,IAAI,GAAG,IAAI,CAAC,cAAc,CAAC,CAAC,CAAC,CAAC;QAClC,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,IAAI,CAAC,cAAc,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;YACpD,IAAI,GAAG,IAAI,CAAC,SAAS,GAAG,IAAI,CAAC,cAAc,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,GAAG,IAAI,CAAC,SAAS,CAAC,GAAG,IAAI,CAAC;QAC/E,CAAC;QACD,OAAO,IAAI,CAAC;IACd,CAAC;IAEO,mBAAmB;QACzB,2CAA2C;QAC3C,MAAM,UAAU,GAAG,IAAI,CAAC,GAAG,CAAC,IAAI,CAAC,cAAc,CAAC,MAAM,GAAG,EAAE,EAAE,GAAG,CAAC,CAAC;QAClE,2CAA2C;QAC3C,MAAM,aAAa,GAAG,IAAI,CAAC,GAAG,CAAC,IAAI,CAAC,cAAc,CAAC,MAAM,GAAG,EAAE,EAAE,GAAG,CAAC,CAAC;QACrE,OAAO,IAAI,CAAC,KAAK,CAAC,CAAC,UAAU,GAAG,GAAG,GAAG,aAAa,GAAG,GAAG,CAAC,GAAG,GAAG,CAAC,GAAG,GAAG,CAAC;IAC1E,CAAC;IAEO,4BAA4B,CAAC,KAAa;QAChD,OAAO,IAAI,CAAC,KAAK,CAAC,KAAK,GAAG,GAAG,CAAC,CAAC,CAAC,mBAAmB;IACrD,CAAC;IAEO,6BAA6B,CAAC,KAAa;QACjD,OAAO,IAAI,CAAC,KAAK,CAAC,KAAK,GAAG,GAAG,CAAC,CAAC,CAAC,oBAAoB;IACtD,CAAC;CACF;AAlQD,oDAkQC;AAED,+DAA+D;AAC/D,UAAU;AACV,+DAA+D;AAE/D,SAAgB,0BAA0B,CACxC,cAAyD;IAEzD,OAAO,IAAI,oBAAoB,CAAC,cAAc,CAAC,CAAC;AAClD,CAAC"}

package/dist/inference/speculativeDecoding.d.ts

@@ -1,133 +0,0 @@
-/**
- * A3M Router - Speculative Decoding Interface
- *
- * Interface for integrating speculative decoding.
- * Currently a stub/interface for future Medusa/Lookahead integration.
- *
- * Speculative decoding uses a smaller "draft" model to predict
- * multiple tokens ahead, which are then verified in parallel by
- * the main model. This can provide 2-3x speedup in generation.
- *
- * Usage:
- *   const specDec = new SpeculativeDecoding();
- *   if (specDec.shouldUse(true)) {
- *     const draftModel = specDec.getDraftModel('medusa');
- *     // ... use draft model to generate and verify
- *   }
- */
-export type DraftModelType = 'medusa' | 'lookahead' | 'eagle' | 'spark';
-export interface SpeculativeConfig {
-    /** Enable speculative decoding */
-    enabled: boolean;
-    /** Draft model type */
-    draftModelType?: DraftModelType;
-    /** Number of speculative tokens to generate */
-    speculationWindow?: number;
-    /** Temperature for draft model */
-    temperature?: number;
-    /** Provider to use for draft model */
-    provider?: string;
-    /** Model to use for draft model */
-    model?: string;
-}
-export interface DraftModel {
-    /** Model identifier */
-    id: string;
-    /** Model type */
-    type: DraftModelType;
-    /** Provider for this model */
-    provider: string;
-    /** Model size description */
-    size: string;
-    /** Supported speculation windows */
-    supportedWindows: number[];
-}
-export interface SpeculativeResult {
-    /** Whether speculative decoding was used */
-    used: boolean;
-    /** Number of tokens in draft */
-    draftTokens: number;
-    /** Number of tokens accepted */
-    acceptedTokens: number;
-    /** Acceptance rate */
-    acceptanceRate: number;
-    /** Time saved (ms, estimated) */
-    estimatedTimeSaved: number;
-}
-export interface VerificationResult {
-    /** All tokens verified successfully */
-    allAccepted: boolean;
-    /** Indices of accepted tokens */
-    acceptedIndices: number[];
-    /** Indices of rejected tokens */
-    rejectedIndices: number[];
-    /** Actual tokens to use (may differ from draft) */
-    actualTokens: string[];
-    /** Number of tokens to rewind */
-    rewindCount: number;
-}
-export declare class SpeculativeDecoding {
-    private config;
-    private draftModels;
-    constructor(config?: Partial<SpeculativeConfig>);
-    /**
-     * Determine if speculative decoding should be used.
-     * Checks config, provider support, and model availability.
-     */
-    shouldUse(forceEnable?: boolean): boolean;
-    /**
-     * Get the draft model configuration.
-     */
-    getDraftModel(type?: DraftModelType): DraftModel | null;
-    /**
-     * Generate draft tokens using the draft model.
-     * This is an async method to support API-based draft models.
-     */
-    generateDraft(prompt: string, maxTokens: number): Promise<{
-        tokens: string[];
-        scores: number[];
-    }>;
-    /**
-     * Verify draft tokens against the main model.
-     * Returns which tokens were accepted and corrections if needed.
-     */
-    verifyDraft(prompt: string, draftTokens: string[]): Promise<VerificationResult>;
-    /**
-     * Calculate speedup from speculative decoding results.
-     */
-    calculateSpeedup(result: SpeculativeResult): number;
-    /**
-     * Update configuration.
-     */
-    updateConfig(updates: Partial<SpeculativeConfig>): void;
-    /**
-     * Get current configuration.
-     */
-    getConfig(): SpeculativeConfig;
-    /**
-     * Check if a provider supports speculative decoding.
-     */
-    supportsSpeculativeDecoding(provider: string): boolean;
-    /**
-     * Get recommended speculation window based on model size.
-     */
-    getRecommendedWindow(modelSize: 'small' | 'medium' | 'large'): number;
-    private registerDraftModels;
-}
-export declare class SpeculativeDecodingWrapper {
-    private specDec;
-    private mainModelCall;
-    constructor(mainModelCall: (prompt: string, options?: any) => Promise<string>, config?: Partial<SpeculativeConfig>);
-    /**
-     * Generate with speculative decoding.
-     */
-    generate(prompt: string, maxTokens: number, options?: {
-        temperature?: number;
-        useSpecDec?: boolean;
-    }): Promise<{
-        text: string;
-        result: SpeculativeResult;
-    }>;
-}
-export declare function createSpeculativeDecoding(config?: Partial<SpeculativeConfig>): SpeculativeDecoding;
-//# sourceMappingURL=speculativeDecoding.d.ts.map

package/dist/inference/speculativeDecoding.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"speculativeDecoding.d.ts","sourceRoot":"","sources":["../../src/inference/speculativeDecoding.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;GAgBG;AAQH,MAAM,MAAM,cAAc,GAAG,QAAQ,GAAG,WAAW,GAAG,OAAO,GAAG,OAAO,CAAC;AAExE,MAAM,WAAW,iBAAiB;IAChC,kCAAkC;IAClC,OAAO,EAAE,OAAO,CAAC;IACjB,uBAAuB;IACvB,cAAc,CAAC,EAAE,cAAc,CAAC;IAChC,+CAA+C;IAC/C,iBAAiB,CAAC,EAAE,MAAM,CAAC;IAC3B,kCAAkC;IAClC,WAAW,CAAC,EAAE,MAAM,CAAC;IACrB,sCAAsC;IACtC,QAAQ,CAAC,EAAE,MAAM,CAAC;IAClB,mCAAmC;IACnC,KAAK,CAAC,EAAE,MAAM,CAAC;CAChB;AAED,MAAM,WAAW,UAAU;IACzB,uBAAuB;IACvB,EAAE,EAAE,MAAM,CAAC;IACX,iBAAiB;IACjB,IAAI,EAAE,cAAc,CAAC;IACrB,8BAA8B;IAC9B,QAAQ,EAAE,MAAM,CAAC;IACjB,6BAA6B;IAC7B,IAAI,EAAE,MAAM,CAAC;IACb,oCAAoC;IACpC,gBAAgB,EAAE,MAAM,EAAE,CAAC;CAC5B;AAED,MAAM,WAAW,iBAAiB;IAChC,4CAA4C;IAC5C,IAAI,EAAE,OAAO,CAAC;IACd,gCAAgC;IAChC,WAAW,EAAE,MAAM,CAAC;IACpB,gCAAgC;IAChC,cAAc,EAAE,MAAM,CAAC;IACvB,sBAAsB;IACtB,cAAc,EAAE,MAAM,CAAC;IACvB,iCAAiC;IACjC,kBAAkB,EAAE,MAAM,CAAC;CAC5B;AAED,MAAM,WAAW,kBAAkB;IACjC,uCAAuC;IACvC,WAAW,EAAE,OAAO,CAAC;IACrB,iCAAiC;IACjC,eAAe,EAAE,MAAM,EAAE,CAAC;IAC1B,iCAAiC;IACjC,eAAe,EAAE,MAAM,EAAE,CAAC;IAC1B,mDAAmD;IACnD,YAAY,EAAE,MAAM,EAAE,CAAC;IACvB,iCAAiC;IACjC,WAAW,EAAE,MAAM,CAAC;CACrB;AAMD,qBAAa,mBAAmB;IAC9B,OAAO,CAAC,MAAM,CAAoB;IAClC,OAAO,CAAC,WAAW,CAAsC;gBAE7C,MAAM,CAAC,EAAE,OAAO,CAAC,iBAAiB,CAAC;IAc/C;;;OAGG;IACH,SAAS,CAAC,WAAW,CAAC,EAAE,OAAO,GAAG,OAAO;IASzC;;OAEG;IACH,aAAa,CAAC,IAAI,CAAC,EAAE,cAAc,GAAG,UAAU,GAAG,IAAI;IAkBvD;;;OAGG;IACG,aAAa,CACjB,MAAM,EAAE,MAAM,EACd,SAAS,EAAE,MAAM,GAChB,OAAO,CAAC;QAAE,MAAM,EAAE,MAAM,EAAE,CAAC;QAAC,MAAM,EAAE,MAAM,EAAE,CAAA;KAAE,CAAC;IAclD;;;OAGG;IACG,WAAW,CACf,MAAM,EAAE,MAAM,EACd,WAAW,EAAE,MAAM,EAAE,GACpB,OAAO,CAAC,kBAAkB,CAAC;IAsB9B;;OAEG;IACH,gBAAgB,CAAC,MAAM,EAAE,iBAAiB,GAAG,MAAM;IAgBnD;;OAEG;IACH,YAAY,CAAC,OAAO,EAAE,OAAO,CAAC,iBAAiB,CAAC,GAAG,IAAI;IAIvD;;OAEG;IACH,SAAS,IAAI,iBAAiB;IAI9B;;OAEG;IACH,2BAA2B,CAAC,QAAQ,EAAE,MAAM,GAAG,OAAO;IAUtD;;OAEG;IACH,oBAAoB,CAAC,SAAS,EAAE,OAAO,GAAG,QAAQ,GAAG,OAAO,GAAG,MAAM;IAWrE,OAAO,CAAC,mBAAmB;CAqC5B;AAMD,qBAAa,0BAA0B;IACrC,OAAO,CAAC,OAAO,CAAsB;IACrC,OAAO,CAAC,aAAa,CAAqD;gBAGxE,aAAa,EAAE,CAAC,MAAM,EAAE,MAAM,EAAE,OAAO,CAAC,EAAE,GAAG,KAAK,OAAO,CAAC,MAAM,CAAC,EACjE,MAAM,CAAC,EAAE,OAAO,CAAC,iBAAiB,CAAC;IAMrC;;OAEG;IACG,QAAQ,CACZ,MAAM,EAAE,MAAM,EACd,SAAS,EAAE,MAAM,EACjB,OAAO,CAAC,EAAE;QAAE,WAAW,CAAC,EAAE,MAAM,CAAC;QAAC,UAAU,CAAC,EAAE,OAAO,CAAA;KAAE,GACvD,OAAO,CAAC;QAAE,IAAI,EAAE,MAAM,CAAC;QAAC,MAAM,EAAE,iBAAiB,CAAA;KAAE,CAAC;CAsDxD;AAMD,wBAAgB,yBAAyB,CACvC,MAAM,CAAC,EAAE,OAAO,CAAC,iBAAiB,CAAC,GAClC,mBAAmB,CAErB"}