@loxia-labs/loxia-autopilot-one 1.0.1 → 1.0.3

package/src/services/tokenCountingService.js

@@ -1,536 +1 @@
-/**
- * TokenCountingService - Accurate token counting for conversation compactization
- *
- * Purpose:
- * - Provide accurate token counting using tiktoken (OpenAI's tokenizer)
- * - Support multiple models with different tokenization schemes
- * - Determine when conversation compactization should be triggered
- * - Cache token counts for performance optimization
- * - Provide fallback estimation when accurate counting unavailable
- *
- * Key Features:
- * - Accurate token counting via tiktoken library
- * - Model-specific tokenization
- * - Intelligent caching with TTL
- * - Fast estimation fallback
- * - Context window management
- * - Compaction trigger detection
- */
-
-import { encoding_for_model } from 'tiktoken';
-import {
-  MODELS,
-  COMPACTION_CONFIG,
-  TOKEN_COUNTING_MODES,
-} from '../utilities/constants.js';
-
-class TokenCountingService {
-  constructor(logger) {
-    this.logger = logger;
-
-    // Token count cache for performance
-    this.tokenCache = new Map();
-
-    // Tiktoken encoders cache (reuse encoders for efficiency)
-    this.encoders = new Map();
-
-    // Supported tiktoken models mapping
-    this.tiktokenModelMap = {
-      // Anthropic models use OpenAI's cl100k_base encoding
-      [MODELS.ANTHROPIC_SONNET]: 'gpt-4',
-      [MODELS.ANTHROPIC_OPUS]: 'gpt-4',
-      [MODELS.ANTHROPIC_HAIKU]: 'gpt-4',
-
-      // OpenAI models
-      [MODELS.GPT_4]: 'gpt-4',
-      [MODELS.GPT_4_MINI]: 'gpt-4',
-      'gpt-4o': 'gpt-4',
-      'gpt-4o-mini': 'gpt-4',
-      'gpt-4-turbo': 'gpt-4',
-      'gpt-3.5-turbo': 'gpt-4', // Uses cl100k_base encoding
-
-      // DeepSeek uses similar tokenization to GPT-4
-      [MODELS.DEEPSEEK_R1]: 'gpt-4',
-
-      // Phi models - fallback to GPT-4 encoding
-      [MODELS.PHI_4]: 'gpt-4',
-
-      // Azure AI Foundry models
-      'azure-ai-grok3': 'gpt-4',
-      'azure-ai-deepseek-r1': 'gpt-4',
-      'azure-openai-gpt-5': 'gpt-4',
-      'azure-openai-gpt-4': 'gpt-4',
-      'azure-openai-gpt-4o': 'gpt-4',
-
-      // Compaction models
-      [COMPACTION_CONFIG.COMPACTION_MODEL]: 'gpt-4',
-      [COMPACTION_CONFIG.COMPACTION_MODEL_FALLBACK]: 'gpt-4',
-    };
-
-    this.logger?.info('TokenCountingService initialized', {
-      supportedModels: Object.keys(this.tiktokenModelMap).length,
-      cacheEnabled: true
-    });
-  }
-
-  /**
-   * Count tokens in text using accurate tiktoken encoder
-   * @param {string} text - Text to count tokens in
-   * @param {string} model - Model name for appropriate tokenization
-   * @param {string} mode - Counting mode (accurate, estimated, cached)
-   * @returns {Promise<number>} Token count
-   */
-  async countTokens(text, model, mode = TOKEN_COUNTING_MODES.ACCURATE) {
-    // Validate input
-    if (!text || typeof text !== 'string') {
-      return 0;
-    }
-
-    // Check cache first if enabled
-    if (mode === TOKEN_COUNTING_MODES.CACHED) {
-      const cached = this._getCachedTokenCount(text, model);
-      if (cached !== null) {
-        return cached;
-      }
-      // Fall through to accurate counting if not cached
-    }
-
-    // Use fast estimation if requested
-    if (mode === TOKEN_COUNTING_MODES.ESTIMATED) {
-      return this._estimateTokens(text);
-    }
-
-    // Accurate counting with tiktoken
-    try {
-      const encoder = await this._getEncoder(model);
-      const tokens = encoder.encode(text);
-      const count = tokens.length;
-
-      // Cache the result
-      this._cacheTokenCount(text, model, count);
-
-      return count;
-
-    } catch (error) {
-      this.logger?.warn('Tiktoken encoding failed, falling back to estimation', {
-        model,
-        error: error.message
-      });
-
-      // Fallback to estimation
-      return this._estimateTokens(text);
-    }
-  }
-
-  /**
-   * Estimate conversation token count including all messages
-   * @param {Array} messages - Array of message objects with { role, content }
-   * @param {string} model - Model name for tokenization
-   * @param {string} mode - Counting mode
-   * @returns {Promise<number>} Total token count for conversation
-   */
-  async estimateConversationTokens(messages, model, mode = TOKEN_COUNTING_MODES.ACCURATE) {
-    if (!Array.isArray(messages) || messages.length === 0) {
-      return 0;
-    }
-
-    let totalTokens = 0;
-
-    // Count tokens for each message
-    for (const message of messages) {
-      if (!message.content) {
-        continue;
-      }
-
-      // Message formatting overhead (role + formatting)
-      const formattingOverhead = this._getMessageFormattingOverhead(model);
-
-      // Content tokens
-      const contentTokens = await this.countTokens(message.content, model, mode);
-
-      totalTokens += contentTokens + formattingOverhead;
-    }
-
-    // Add conversation-level overhead (system instructions, etc.)
-    const conversationOverhead = this._getConversationOverhead(messages.length);
-    totalTokens += conversationOverhead;
-
-    this.logger?.debug('Conversation token count', {
-      model,
-      messageCount: messages.length,
-      totalTokens,
-      mode
-    });
-
-    return totalTokens;
-  }
-
-  /**
-   * Get context window size for a model
-   * @param {string} model - Model name
-   * @returns {number} Context window size in tokens
-   */
-  getModelContextWindow(model) {
-    // Model context windows (from aiService model specs and vendor documentation)
-    const contextWindows = {
-      // Anthropic Claude models
-      [MODELS.ANTHROPIC_SONNET]: 200000,
-      [MODELS.ANTHROPIC_OPUS]: 200000,
-      [MODELS.ANTHROPIC_HAIKU]: 200000,
-
-      // OpenAI models
-      [MODELS.GPT_4]: 128000,
-      [MODELS.GPT_4_MINI]: 128000,
-      'gpt-4o': 128000,
-      'gpt-4o-mini': 128000,
-      'gpt-4-turbo': 128000,
-      'gpt-3.5-turbo': 16384,
-
-      // DeepSeek models
-      [MODELS.DEEPSEEK_R1]: 128000,
-
-      // Phi models
-      [MODELS.PHI_4]: 16384,
-
-      // Azure AI Foundry models
-      'azure-ai-grok3': 128000,
-      'azure-ai-deepseek-r1': 128000,
-      'azure-openai-gpt-5': 128000,
-      'azure-openai-gpt-4': 128000,
-      'azure-openai-gpt-4o': 128000,
-
-      // Compaction models
-      [COMPACTION_CONFIG.COMPACTION_MODEL]: 128000,
-      [COMPACTION_CONFIG.COMPACTION_MODEL_FALLBACK]: 128000,
-
-      // Router model
-      'autopilot-model-router': 16384,
-    };
-
-    const contextWindow = contextWindows[model];
-
-    if (!contextWindow) {
-      this.logger?.warn('Unknown model context window, using default', {
-        model,
-        defaultWindow: 128000
-      });
-      return 128000; // Default to 128k
-    }
-
-    return contextWindow;
-  }
-
-  /**
-   * Get maximum output tokens for a model
-   * @param {string} model - Model name
-   * @returns {number} Maximum output tokens
-   */
-  getModelMaxOutputTokens(model) {
-    // Max output tokens - increased to 8K-20K where supported for better responses
-    // Note: These are conservative estimates to ensure compaction triggers appropriately
-    const maxOutputTokens = {
-      // Anthropic Claude models - support up to 8K output
-      [MODELS.ANTHROPIC_SONNET]: 8192,
-      [MODELS.ANTHROPIC_OPUS]: 8192,
-      [MODELS.ANTHROPIC_HAIKU]: 8192,
-
-      // OpenAI models
-      [MODELS.GPT_4]: 8192,           // Supports up to 16K, using 8K for safety
-      [MODELS.GPT_4_MINI]: 16384,     // Already at max
-      'gpt-4o': 8192,                 // Supports up to 16K, using 8K
-      'gpt-4o-mini': 16384,           // Already at max
-      'gpt-4-turbo': 8192,            // Increased from default
-      'gpt-3.5-turbo': 4096,          // Smaller model, keep at 4K
-
-      // DeepSeek models
-      [MODELS.DEEPSEEK_R1]: 8192,     // Already at max
-
-      // Phi models - smaller architecture
-      [MODELS.PHI_4]: 4096,           // Increased from 2048
-
-      // Azure AI Foundry models
-      'azure-ai-grok3': 8192,         // Increased from 4K
-      'azure-ai-deepseek-r1': 8192,   // Already at max
-      'azure-openai-gpt-5': 8192,     // Increased from 4K
-      'azure-openai-gpt-4': 8192,     // Increased from default
-      'azure-openai-gpt-4o': 8192,    // Increased from default
-
-      // Compaction models - keep moderate for efficiency
-      [COMPACTION_CONFIG.COMPACTION_MODEL]: 8192,
-      [COMPACTION_CONFIG.COMPACTION_MODEL_FALLBACK]: 8192,
-
-      // Router model - keep small for fast routing
-      'autopilot-model-router': 2048,
-    };
-
-    return maxOutputTokens[model] || 8192; // Default increased to 8K
-  }
-
-  /**
-   * Determine if compaction should be triggered
-   * @param {number} currentTokens - Current conversation token count (K)
-   * @param {number} maxOutputTokens - Max tokens model can output (X)
-   * @param {number} contextWindow - Model's context window size (C)
-   * @param {number} threshold - Trigger threshold (default 0.8 = 80%)
-   * @returns {boolean} True if compaction should be triggered
-   */
-  shouldTriggerCompaction(currentTokens, maxOutputTokens, contextWindow, threshold = COMPACTION_CONFIG.DEFAULT_THRESHOLD) {
-    // Validate threshold
-    if (threshold < COMPACTION_CONFIG.MIN_THRESHOLD || threshold > COMPACTION_CONFIG.MAX_THRESHOLD) {
-      this.logger?.warn('Invalid compaction threshold, using default', {
-        provided: threshold,
-        default: COMPACTION_CONFIG.DEFAULT_THRESHOLD
-      });
-      threshold = COMPACTION_CONFIG.DEFAULT_THRESHOLD;
-    }
-
-    // Calculate: K + X >= threshold * C
-    const requiredTokens = currentTokens + maxOutputTokens;
-    const thresholdTokens = threshold * contextWindow;
-    const shouldTrigger = requiredTokens >= thresholdTokens;
-
-    this.logger?.debug('Compaction trigger check', {
-      currentTokens,
-      maxOutputTokens,
-      contextWindow,
-      threshold,
-      requiredTokens,
-      thresholdTokens,
-      shouldTrigger,
-      utilizationPercent: ((requiredTokens / contextWindow) * 100).toFixed(2)
-    });
-
-    return shouldTrigger;
-  }
-
-  /**
-   * Calculate how many tokens to target after compaction
-   * @param {number} contextWindow - Model's context window size
-   * @param {number} targetThreshold - Target threshold after compaction (default 85%)
-   * @returns {number} Target token count after compaction
-   */
-  calculateTargetTokenCount(contextWindow, targetThreshold = COMPACTION_CONFIG.MAX_ACCEPTABLE_TOKEN_COUNT_AFTER) {
-    return Math.floor(contextWindow * targetThreshold);
-  }
-
-  /**
-   * Validate that compaction achieved sufficient reduction
-   * @param {number} originalTokens - Token count before compaction
-   * @param {number} compactedTokens - Token count after compaction
-   * @param {number} contextWindow - Model's context window
-   * @returns {Object} Validation result { valid, reductionPercent, exceedsTarget }
-   */
-  validateCompaction(originalTokens, compactedTokens, contextWindow) {
-    const reductionPercent = ((originalTokens - compactedTokens) / originalTokens) * 100;
-    const targetTokens = this.calculateTargetTokenCount(contextWindow);
-    const exceedsTarget = compactedTokens > targetTokens;
-    const sufficientReduction = reductionPercent >= COMPACTION_CONFIG.MIN_REDUCTION_PERCENTAGE;
-
-    const valid = !exceedsTarget && sufficientReduction;
-
-    this.logger?.info('Compaction validation', {
-      originalTokens,
-      compactedTokens,
-      reductionPercent: reductionPercent.toFixed(2),
-      targetTokens,
-      exceedsTarget,
-      sufficientReduction,
-      valid
-    });
-
-    return {
-      valid,
-      reductionPercent,
-      exceedsTarget,
-      sufficientReduction,
-      targetTokens,
-      compactedTokens,
-      originalTokens
-    };
-  }
-
-  /**
-   * Clear token count cache
-   * @param {string} model - Optional: clear cache for specific model only
-   */
-  clearCache(model = null) {
-    if (model) {
-      // Clear cache entries for specific model
-      for (const [key, value] of this.tokenCache.entries()) {
-        if (value.model === model) {
-          this.tokenCache.delete(key);
-        }
-      }
-      this.logger?.debug('Token cache cleared for model', { model });
-    } else {
-      // Clear entire cache
-      this.tokenCache.clear();
-      this.logger?.debug('Token cache cleared completely');
-    }
-  }
-
-  /**
-   * Get encoder for model (with caching)
-   * @private
-   */
-  async _getEncoder(model) {
-    // Get tiktoken model name
-    const tiktokenModel = this.tiktokenModelMap[model] || 'gpt-4';
-
-    // Check if encoder is already cached
-    if (this.encoders.has(tiktokenModel)) {
-      return this.encoders.get(tiktokenModel);
-    }
-
-    // Create new encoder
-    const encoder = encoding_for_model(tiktokenModel);
-    this.encoders.set(tiktokenModel, encoder);
-
-    this.logger?.debug('Created tiktoken encoder', {
-      model,
-      tiktokenModel
-    });
-
-    return encoder;
-  }
-
-  /**
-   * Fast token estimation (character-based)
-   * @private
-   */
-  _estimateTokens(text) {
-    if (!text || typeof text !== 'string') {
-      return 0;
-    }
-
-    // Use configured estimation ratio
-    return Math.ceil(text.length / COMPACTION_CONFIG.CHARS_PER_TOKEN_ESTIMATE);
-  }
-
-  /**
-   * Get cached token count if available and not expired
-   * @private
-   */
-  _getCachedTokenCount(text, model) {
-    const cacheKey = this._getCacheKey(text, model);
-    const cached = this.tokenCache.get(cacheKey);
-
-    if (!cached) {
-      return null;
-    }
-
-    // Check if cache entry is expired
-    const now = Date.now();
-    if (now - cached.timestamp > COMPACTION_CONFIG.TOKEN_COUNT_CACHE_TTL_MS) {
-      this.tokenCache.delete(cacheKey);
-      return null;
-    }
-
-    return cached.count;
-  }
-
-  /**
-   * Cache token count with timestamp
-   * @private
-   */
-  _cacheTokenCount(text, model, count) {
-    const cacheKey = this._getCacheKey(text, model);
-
-    this.tokenCache.set(cacheKey, {
-      count,
-      model,
-      timestamp: Date.now()
-    });
-
-    // Prevent cache from growing indefinitely
-    if (this.tokenCache.size > 1000) {
-      // Remove oldest entries
-      const entries = Array.from(this.tokenCache.entries());
-      entries.sort((a, b) => a[1].timestamp - b[1].timestamp);
-
-      // Remove oldest 20%
-      const toRemove = Math.floor(entries.length * 0.2);
-      for (let i = 0; i < toRemove; i++) {
-        this.tokenCache.delete(entries[i][0]);
-      }
-    }
-  }
-
-  /**
-   * Generate cache key for text + model
-   * @private
-   */
-  _getCacheKey(text, model) {
-    // Use a simple hash for the text to avoid storing full text as key
-    const hash = this._simpleHash(text);
-    return `${model}:${hash}`;
-  }
-
-  /**
-   * Simple hash function for cache keys
-   * @private
-   */
-  _simpleHash(str) {
-    let hash = 0;
-    for (let i = 0; i < str.length; i++) {
-      const char = str.charCodeAt(i);
-      hash = ((hash << 5) - hash) + char;
-      hash = hash & hash; // Convert to 32-bit integer
-    }
-    return hash.toString(36);
-  }
-
-  /**
-   * Get message formatting overhead for model
-   * Accounts for role labels, XML tags, etc.
-   * @private
-   */
-  _getMessageFormattingOverhead(model) {
-    // Different models have different formatting
-    // Anthropic: ~10 tokens per message (role tags)
-    // OpenAI: ~5 tokens per message (JSON formatting)
-
-    if (model.includes('anthropic') || model.includes('claude')) {
-      return 10;
-    }
-
-    return 5; // Default for OpenAI-style models
-  }
-
-  /**
-   * Get conversation-level overhead
-   * Accounts for system prompts, special tokens, etc.
-   * @private
-   */
-  _getConversationOverhead(messageCount) {
-    // Base overhead for conversation structure
-    const baseOverhead = 50;
-
-    // Additional overhead scales with message count
-    const scalingOverhead = messageCount * 2;
-
-    return baseOverhead + scalingOverhead;
-  }
-
-  /**
-   * Clean up resources (close encoders)
-   */
-  async cleanup() {
-    // Free tiktoken encoders
-    for (const [model, encoder] of this.encoders.entries()) {
-      try {
-        encoder.free();
-        this.logger?.debug('Freed tiktoken encoder', { model });
-      } catch (error) {
-        this.logger?.warn('Failed to free encoder', { model, error: error.message });
-      }
-    }
-
-    this.encoders.clear();
-    this.tokenCache.clear();
-
-    this.logger?.info('TokenCountingService cleaned up');
-  }
-}
-
-export default TokenCountingService;
+const a0_0x5d1028=a0_0x1ea6;(function(_0x244a67,_0x5603bf){const _0x4aac03=a0_0x1ea6,_0x61b607=_0x244a67();while(!![]){try{const _0x389a84=-parseInt(_0x4aac03(0x128))/0x1*(parseInt(_0x4aac03(0x126))/0x2)+-parseInt(_0x4aac03(0x116))/0x3+parseInt(_0x4aac03(0x122))/0x4+parseInt(_0x4aac03(0x11c))/0x5*(-parseInt(_0x4aac03(0x100))/0x6)+parseInt(_0x4aac03(0x130))/0x7+-parseInt(_0x4aac03(0x111))/0x8+parseInt(_0x4aac03(0x118))/0x9;if(_0x389a84===_0x5603bf)break;else _0x61b607['push'](_0x61b607['shift']());}catch(_0x8b51e6){_0x61b607['push'](_0x61b607['shift']());}}}(a0_0xab84,0x8fda8));import{encoding_for_model}from'tiktoken';import{MODELS,COMPACTION_CONFIG,TOKEN_COUNTING_MODES}from'../utilities/constants.js';function a0_0xab84(){const _0x438603=['y2XLyw51Ca','q3jLyxrLzcb0AwT0B2TLBIbLBMnVzgvY','BM93','DMfSAwrHDgvdB21Wywn0Aw9U','odKWmJy3mMXeBezzvG','z3b0ltq','C3rYAw5N','z2v0tw9KzwXnyxHpDxrWDxruB2TLBNm','vg9Rzw4Gy2fJAguGy2XLyxjLzcbJB21WBgv0zwX5','mJGWmJKXofrnBw5lAa','r1buxZrFtuLosq','mtu0nJaWnJvOCujvEwi','BwvZC2fNzq','q29UDMvYC2f0Aw9UihrVA2vUignVDw50','C29YDa','otKWnun3BwfHDW','x3nPBxbSzuHHC2G','A2v5CW','qundvvjbveu','ve9lru5Fq09vtLrFq0fdsevFvfrmx01t','x2DLDevUy29Kzxi','mJe1nJqYmfHKqwrAzG','vg9Rzw5dB3vUDgLUz1nLCNzPy2uGAw5PDgLHBgL6zwq','ueHjxZq','D2fYBG','mNbht2rpva','x2DLDe1LC3nHz2vgB3jTyxr0Aw5Nt3zLCMHLywq','mZmXnZK1Axj2rK1N','DgLTzxn0yw1W','BgvUz3rO','qu5usfjpueLdx09qvvm','x2DLDenVBNzLCNnHDgLVBK92zxjOzwfK','Aw5MBW','y291BNruB2TLBNm','zw5JB2rLCNm','ntqYmJyYn2nZBxHWDq','y29UDgvUDa','vg9Rzw4Gy2fJAguGy2XLyxjLzcbMB3iGBw9KzwW','C2L6zq','zgvSzxrL','Aw5JBhvKzxm','Bg9Nz2vY','x2DLDenHy2HLzfrVA2vUq291BNq','tufyx1riuKvtse9mra','revgqvvmvf9usfjfu0Hpteq','zgvIDwC','DgLRDg9Rzw5nB2rLBe1HCa','zNjVBq','Dg9Rzw5dywnOzq','rNjLzwqGDgLRDg9Rzw4Gzw5JB2rLCG','qu5usfjpueLdx0HbsuTv','zw50CMLLCW','x2DLDenHy2HLs2v5','mtKYwwTitvHl','C2v0','z2v0','y2XLyxi','revfufnfruTFuJe','y2fSy3vSyxrLvgfYz2v0vg9Rzw5dB3vUDa','qu5usfjpueLdx1nptK5fva','Dg9tDhjPBMC','q29TCgfJDgLVBIb2ywXPzgf0Aw9U','yw50AhjVCgLJ','x2nHy2HLvg9Rzw5dB3vUDa','zMXVB3i','x2vZDgLTyxrLvg9Rzw5Z'];a0_0xab84=function(){return _0x438603;};return a0_0xab84();}function a0_0x1ea6(_0x21bb1d,_0x55e987){_0x21bb1d=_0x21bb1d-0x100;const _0xab842=a0_0xab84();let _0x1ea641=_0xab842[_0x21bb1d];if(a0_0x1ea6['uWvcYs']===undefined){var _0x4c061c=function(_0x498bc4){const _0x5cc5c0='abcdefghijklmnopqrstuvwxyzABCDEFGHIJKLMNOPQRSTUVWXYZ0123456789+/=';let _0x22d55a='',_0x1365dc='';for(let _0x5d798a=0x0,_0x4130e7,_0x3de2c0,_0x1691a=0x0;_0x3de2c0=_0x498bc4['charAt'](_0x1691a++);~_0x3de2c0&&(_0x4130e7=_0x5d798a%0x4?_0x4130e7*0x40+_0x3de2c0:_0x3de2c0,_0x5d798a++%0x4)?_0x22d55a+=String['fromCharCode'](0xff&_0x4130e7>>(-0x2*_0x5d798a&0x6)):0x0){_0x3de2c0=_0x5cc5c0['indexOf'](_0x3de2c0);}for(let _0x36cc2f=0x0,_0x572791=_0x22d55a['length'];_0x36cc2f<_0x572791;_0x36cc2f++){_0x1365dc+='%'+('00'+_0x22d55a['charCodeAt'](_0x36cc2f)['toString'](0x10))['slice'](-0x2);}return decodeURIComponent(_0x1365dc);};a0_0x1ea6['CEGzJM']=_0x4c061c,a0_0x1ea6['vHUMAv']={},a0_0x1ea6['uWvcYs']=!![];}const _0x5a6016=_0xab842[0x0],_0x535464=_0x21bb1d+_0x5a6016,_0x6fe1b6=a0_0x1ea6['vHUMAv'][_0x535464];return!_0x6fe1b6?(_0x1ea641=a0_0x1ea6['CEGzJM'](_0x1ea641),a0_0x1ea6['vHUMAv'][_0x535464]=_0x1ea641):_0x1ea641=_0x6fe1b6,_0x1ea641;}class TokenCountingService{constructor(_0x22d55a){const _0x448902=a0_0x1ea6;this[_0x448902(0x136)]=_0x22d55a,this['tokenCache']=new Map(),this['encoders']=new Map(),this['tiktokenModelMap']={[MODELS[_0x448902(0x106)]]:_0x448902(0x112),[MODELS['ANTHROPIC_OPUS']]:'gpt-4',[MODELS[_0x448902(0x13f)]]:'gpt-4',[MODELS['GPT_4']]:_0x448902(0x112),[MODELS[_0x448902(0x117)]]:'gpt-4','gpt-4o':'gpt-4','gpt-4o-mini':'gpt-4','gpt-4-turbo':_0x448902(0x112),'gpt-3.5-turbo':'gpt-4',[MODELS[_0x448902(0x104)]]:'gpt-4',[MODELS['PHI_4']]:_0x448902(0x112),'azure-ai-grok3':_0x448902(0x112),'azure-ai-deepseek-r1':_0x448902(0x112),'azure-openai-gpt-5':_0x448902(0x112),'azure-openai-gpt-4':_0x448902(0x112),'azure-openai-gpt-4o':_0x448902(0x112),[COMPACTION_CONFIG['COMPACTION_MODEL']]:_0x448902(0x112),[COMPACTION_CONFIG['COMPACTION_MODEL_FALLBACK']]:_0x448902(0x112)},this[_0x448902(0x136)]?.['info'](_0x448902(0x123),{'supportedModels':Object[_0x448902(0x11e)](this[_0x448902(0x13b)])['length'],'cacheEnabled':!![]});}async[a0_0x5d1028(0x12e)](_0x1365dc,_0x5d798a,_0x4130e7=TOKEN_COUNTING_MODES[a0_0x5d1028(0x11f)]){const _0x1f7b4a=a0_0x5d1028;if(!_0x1365dc||typeof _0x1365dc!==_0x1f7b4a(0x113))return 0x0;if(_0x4130e7===TOKEN_COUNTING_MODES['CACHED']){const _0x3de2c0=this['_getCachedTokenCount'](_0x1365dc,_0x5d798a);if(_0x3de2c0!==null)return _0x3de2c0;}if(_0x4130e7===TOKEN_COUNTING_MODES['ESTIMATED'])return this['_estimateTokens'](_0x1365dc);try{const _0x1691a=await this[_0x1f7b4a(0x121)](_0x5d798a),_0x36cc2f=_0x1691a['encode'](_0x1365dc),_0x572791=_0x36cc2f['length'];return this[_0x1f7b4a(0x10a)](_0x1365dc,_0x5d798a,_0x572791),_0x572791;}catch(_0x27d33d){return this['logger']?.['warn']('Tiktoken\x20encoding\x20failed,\x20falling\x20back\x20to\x20estimation',{'model':_0x5d798a,'error':_0x27d33d[_0x1f7b4a(0x119)]}),this[_0x1f7b4a(0x10c)](_0x1365dc);}}async['estimateConversationTokens'](_0x3abd2f,_0x86e080,_0x32e18f=TOKEN_COUNTING_MODES[a0_0x5d1028(0x11f)]){const _0x286bdb=a0_0x5d1028;if(!Array['isArray'](_0x3abd2f)||_0x3abd2f[_0x286bdb(0x12a)]===0x0)return 0x0;let _0x4e5630=0x0;for(const _0x574837 of _0x3abd2f){if(!_0x574837['content'])continue;const _0xd9f691=this[_0x286bdb(0x127)](_0x86e080),_0x1c134e=await this['countTokens'](_0x574837[_0x286bdb(0x131)],_0x86e080,_0x32e18f);_0x4e5630+=_0x1c134e+_0xd9f691;}const _0xb50ed2=this[_0x286bdb(0x12c)](_0x3abd2f['length']);return _0x4e5630+=_0xb50ed2,this['logger']?.[_0x286bdb(0x13a)](_0x286bdb(0x11a),{'model':_0x86e080,'messageCount':_0x3abd2f[_0x286bdb(0x12a)],'totalTokens':_0x4e5630,'mode':_0x32e18f}),_0x4e5630;}['getModelContextWindow'](_0x5f3b2a){const _0x4c6369=a0_0x5d1028,_0x11124b={[MODELS['ANTHROPIC_SONNET']]:0x30d40,[MODELS[_0x4c6369(0x12b)]]:0x30d40,[MODELS['ANTHROPIC_HAIKU']]:0x30d40,[MODELS['GPT_4']]:0x1f400,[MODELS[_0x4c6369(0x117)]]:0x1f400,'gpt-4o':0x1f400,'gpt-4o-mini':0x1f400,'gpt-4-turbo':0x1f400,'gpt-3.5-turbo':0x4000,[MODELS[_0x4c6369(0x104)]]:0x1f400,[MODELS[_0x4c6369(0x124)]]:0x4000,'azure-ai-grok3':0x1f400,'azure-ai-deepseek-r1':0x1f400,'azure-openai-gpt-5':0x1f400,'azure-openai-gpt-4':0x1f400,'azure-openai-gpt-4o':0x1f400,[COMPACTION_CONFIG['COMPACTION_MODEL']]:0x1f400,[COMPACTION_CONFIG['COMPACTION_MODEL_FALLBACK']]:0x1f400,'autopilot-model-router':0x4000},_0x554161=_0x11124b[_0x5f3b2a];if(!_0x554161)return this['logger']?.[_0x4c6369(0x125)]('Unknown model context window, using default',{'model':_0x5f3b2a,'defaultWindow':0x1f400}),0x1f400;return _0x554161;}[a0_0x5d1028(0x114)](_0x37e30b){const _0x523617=a0_0x5d1028,_0x5150a7={[MODELS['ANTHROPIC_SONNET']]:0x2000,[MODELS['ANTHROPIC_OPUS']]:0x2000,[MODELS['ANTHROPIC_HAIKU']]:0x2000,[MODELS['GPT_4']]:0x2000,[MODELS[_0x523617(0x117)]]:0x4000,'gpt-4o':0x2000,'gpt-4o-mini':0x4000,'gpt-4-turbo':0x2000,'gpt-3.5-turbo':0x1000,[MODELS['DEEPSEEK_R1']]:0x2000,[MODELS[_0x523617(0x124)]]:0x1000,'azure-ai-grok3':0x2000,'azure-ai-deepseek-r1':0x2000,'azure-openai-gpt-5':0x2000,'azure-openai-gpt-4':0x2000,'azure-openai-gpt-4o':0x2000,[COMPACTION_CONFIG['COMPACTION_MODEL']]:0x2000,[COMPACTION_CONFIG['COMPACTION_MODEL_FALLBACK']]:0x2000,'autopilot-model-router':0x800};return _0x5150a7[_0x37e30b]||0x2000;}['shouldTriggerCompaction'](_0x36b00f,_0x2bec88,_0x342b5e,_0x30d36c=COMPACTION_CONFIG[a0_0x5d1028(0x139)]){const _0x41d89a=a0_0x5d1028;(_0x30d36c<COMPACTION_CONFIG['MIN_THRESHOLD']||_0x30d36c>COMPACTION_CONFIG[_0x41d89a(0x138)])&&(this[_0x41d89a(0x136)]?.['warn']('Invalid compaction threshold, using default',{'provided':_0x30d36c,'default':COMPACTION_CONFIG['DEFAULT_THRESHOLD']}),_0x30d36c=COMPACTION_CONFIG['DEFAULT_THRESHOLD']);const requiredTokens=_0x36b00f+_0x2bec88,_0x414512=_0x30d36c*_0x342b5e,_0xd6df14=requiredTokens>=_0x414512;return this[_0x41d89a(0x136)]?.['debug']('Compaction\x20trigger\x20check',{'currentTokens':_0x36b00f,'maxOutputTokens':_0x2bec88,'contextWindow':_0x342b5e,'threshold':_0x30d36c,'requiredTokens':requiredTokens,'thresholdTokens':_0x414512,'shouldTrigger':_0xd6df14,'utilizationPercent':(requiredTokens/_0x342b5e*0x64)['toFixed'](0x2)}),_0xd6df14;}['calculateTargetTokenCount'](_0x2c017a,_0x334cc1=COMPACTION_CONFIG['MAX_ACCEPTABLE_TOKEN_COUNT_AFTER']){const _0x484f4c=a0_0x5d1028;return Math[_0x484f4c(0x10b)](_0x2c017a*_0x334cc1);}[a0_0x5d1028(0x110)](_0x492dcd,_0x1112fa,_0x2eb739){const _0x198508=a0_0x5d1028,_0x2318f4=(_0x492dcd-_0x1112fa)/_0x492dcd*0x64,_0x23b295=this[_0x198508(0x105)](_0x2eb739),_0x3ec764=_0x1112fa>_0x23b295,_0x5a5f66=_0x2318f4>=COMPACTION_CONFIG['MIN_REDUCTION_PERCENTAGE'],_0x1f11d1=!_0x3ec764&&_0x5a5f66;return this['logger']?.[_0x198508(0x12d)](_0x198508(0x108),{'originalTokens':_0x492dcd,'compactedTokens':_0x1112fa,'reductionPercent':_0x2318f4['toFixed'](0x2),'targetTokens':_0x23b295,'exceedsTarget':_0x3ec764,'sufficientReduction':_0x5a5f66,'valid':_0x1f11d1}),{'valid':_0x1f11d1,'reductionPercent':_0x2318f4,'exceedsTarget':_0x3ec764,'sufficientReduction':_0x5a5f66,'targetTokens':_0x23b295,'compactedTokens':_0x1112fa,'originalTokens':_0x492dcd};}['clearCache'](_0x506dcb=null){const _0x22906d=a0_0x5d1028;if(_0x506dcb){for(const [_0x4d2286,_0x2d7f1f]of this[_0x22906d(0x13d)][_0x22906d(0x140)]()){_0x2d7f1f['model']===_0x506dcb&&this['tokenCache'][_0x22906d(0x134)](_0x4d2286);}this[_0x22906d(0x136)]?.['debug'](_0x22906d(0x132),{'model':_0x506dcb});}else this['tokenCache']['clear'](),this['logger']?.['debug'](_0x22906d(0x115));}async[a0_0x5d1028(0x121)](_0x5dab9c){const _0x23b4b6=a0_0x5d1028,_0x322e62=this['tiktokenModelMap'][_0x5dab9c]||_0x23b4b6(0x112);if(this['encoders']['has'](_0x322e62))return this[_0x23b4b6(0x12f)][_0x23b4b6(0x102)](_0x322e62);const _0x2b08cd=encoding_for_model(_0x322e62);return this['encoders'][_0x23b4b6(0x101)](_0x322e62,_0x2b08cd),this[_0x23b4b6(0x136)]?.[_0x23b4b6(0x13a)](_0x23b4b6(0x10e),{'model':_0x5dab9c,'tiktokenModel':_0x322e62}),_0x2b08cd;}['_estimateTokens'](_0x127465){if(!_0x127465||typeof _0x127465!=='string')return 0x0;return Math['ceil'](_0x127465['length']/COMPACTION_CONFIG['CHARS_PER_TOKEN_ESTIMATE']);}[a0_0x5d1028(0x137)](_0x14a093,_0x2f3a51){const _0x3b1a58=a0_0x5d1028,_0x3a1d28=this[_0x3b1a58(0x141)](_0x14a093,_0x2f3a51),_0x174eb1=this[_0x3b1a58(0x13d)][_0x3b1a58(0x102)](_0x3a1d28);if(!_0x174eb1)return null;const _0x5dacd2=Date['now']();if(_0x5dacd2-_0x174eb1['timestamp']>COMPACTION_CONFIG[_0x3b1a58(0x120)])return this[_0x3b1a58(0x13d)][_0x3b1a58(0x134)](_0x3a1d28),null;return _0x174eb1['count'];}['_cacheTokenCount'](_0x5427bc,_0x1fff2b,_0x32d27b){const _0x101a11=a0_0x5d1028,_0xb91598=this[_0x101a11(0x141)](_0x5427bc,_0x1fff2b);this['tokenCache']['set'](_0xb91598,{'count':_0x32d27b,'model':_0x1fff2b,'timestamp':Date[_0x101a11(0x10f)]()});if(this['tokenCache'][_0x101a11(0x133)]>0x3e8){const _0x581227=Array[_0x101a11(0x13c)](this['tokenCache'][_0x101a11(0x140)]());_0x581227[_0x101a11(0x11b)]((_0x484700,_0x5a2eda)=>_0x484700[0x1]['timestamp']-_0x5a2eda[0x1][_0x101a11(0x129)]);const _0x4a8fde=Math[_0x101a11(0x10b)](_0x581227[_0x101a11(0x12a)]*0.2);for(let _0x29e3bf=0x0;_0x29e3bf<_0x4a8fde;_0x29e3bf++){this['tokenCache']['delete'](_0x581227[_0x29e3bf][0x0]);}}}[a0_0x5d1028(0x141)](_0x1cfdd3,_0x140b44){const _0x108ac=a0_0x5d1028,_0x194f19=this[_0x108ac(0x11d)](_0x1cfdd3);return _0x140b44+':'+_0x194f19;}['_simpleHash'](_0x2c9749){const _0x1a8ba1=a0_0x5d1028;let _0x441d9e=0x0;for(let _0x1e8509=0x0;_0x1e8509<_0x2c9749['length'];_0x1e8509++){const _0x2ca5de=_0x2c9749['charCodeAt'](_0x1e8509);_0x441d9e=(_0x441d9e<<0x5)-_0x441d9e+_0x2ca5de,_0x441d9e=_0x441d9e&_0x441d9e;}return _0x441d9e[_0x1a8ba1(0x107)](0x24);}[a0_0x5d1028(0x127)](_0xe6aba5){const _0x171cc1=a0_0x5d1028;if(_0xe6aba5[_0x171cc1(0x135)](_0x171cc1(0x109))||_0xe6aba5['includes']('claude'))return 0xa;return 0x5;}[a0_0x5d1028(0x12c)](_0x283b04){const _0x10e61f=0x32,_0x59492b=_0x283b04*0x2;return _0x10e61f+_0x59492b;}async[a0_0x5d1028(0x10d)](){const _0x349433=a0_0x5d1028;for(const [_0x23aaa9,_0x57d288]of this[_0x349433(0x12f)][_0x349433(0x140)]()){try{_0x57d288['free'](),this[_0x349433(0x136)]?.[_0x349433(0x13a)](_0x349433(0x13e),{'model':_0x23aaa9});}catch(_0x51bfe3){this[_0x349433(0x136)]?.[_0x349433(0x125)]('Failed\x20to\x20free\x20encoder',{'model':_0x23aaa9,'error':_0x51bfe3['message']});}}this['encoders']['clear'](),this[_0x349433(0x13d)][_0x349433(0x103)](),this['logger']?.[_0x349433(0x12d)]('TokenCountingService\x20cleaned\x20up');}}export default TokenCountingService;