Exched/DeepSeek_Coder

1/**2 * Calculate optimal max_tokens based on provider capabilities and input size3 * 4 * @param selectedProvider - The selected provider object from getBestProvider5 * @param inputTokens - Estimated input tokens (prompt + system message + context)6 * @param isStreaming - Whether this is a streaming request (affects buffer)7 * @returns Optimal max_tokens value8 */9export function calculateMaxTokens(10  selectedProvider: any,11  inputTokens: number = 0,12  isStreaming: boolean = false13): number {14  if (!selectedProvider?.context_length) {15    // Fallback for unknown providers - use conservative default16    return 4096;17  }18 19  const contextLength = selectedProvider.context_length;20  21  // Reserve buffer for safety and potential tokenization differences22  const safetyBuffer = isStreaming ? 1000 : 500;23  24  // Calculate available tokens for output25  const availableTokens = contextLength - inputTokens - safetyBuffer;26  27  // Define reasonable max output limits based on use case28  const useCase = {29    // For HTML generation, we typically need substantial output30    htmlGeneration: Math.min(32_000, availableTokens),31    // For code editing, moderate output is usually sufficient  32    codeEditing: Math.min(16_000, availableTokens),33    // Conservative fallback34    default: Math.min(8_000, availableTokens)35  };36  37  // Choose based on available tokens and provider capabilities38  let targetTokens: number;39  40  if (availableTokens >= 32_000) {41    targetTokens = useCase.htmlGeneration;42  } else if (availableTokens >= 16_000) {43    targetTokens = useCase.codeEditing;44  } else {45    targetTokens = useCase.default;46  }47  48  // Ensure we don't go below minimum viable output49  const minimumViableOutput = 2048;50  if (targetTokens < minimumViableOutput) {51    // If we can't provide minimum viable output, try with minimal buffer52    const minimalBuffer = 200;53    targetTokens = Math.max(54      minimumViableOutput,55      contextLength - inputTokens - minimalBuffer56    );57  }58  59  // Final safety check - never exceed context length60  return Math.min(targetTokens, contextLength - inputTokens - 100);61}62 63/**64 * Estimate input tokens for a request (rough estimation)65 * 66 * @param systemPrompt - System prompt content67 * @param userPrompt - User prompt content  68 * @param additionalContext - Additional context (templates, pages, etc.)69 * @returns Estimated token count70 */71export function estimateInputTokens(72  systemPrompt: string = "",73  userPrompt: string = "",74  additionalContext: string = ""75): number {76  // Rough estimation: ~4 characters per token for English text77  // This is conservative - actual tokenization may vary78  const totalChars = systemPrompt.length + userPrompt.length + additionalContext.length;79  return Math.ceil(totalChars / 3.5); // Slightly more conservative than 4 chars/token80}81 82/**83 * Get max_tokens configuration for specific providers with special handling84 */85export function getProviderSpecificConfig(selectedProvider: any, baseMaxTokens: number) {86  const providerName = selectedProvider?.provider;87  88  switch (providerName) {89    case "sambanova":90      // SambaNova has specific limitations - don't set max_tokens91      return {};92    default:93      return { max_tokens: baseMaxTokens };94  }95}96