import { Injectable, Logger, Inject, forwardRef } from '@nestjs/common';
import { ConfigService } from '@nestjs/config';
import { ElasticsearchService } from '../elasticsearch/elasticsearch.service';
import { EmbeddingService } from '../knowledge-base/embedding.service';
import { ModelConfigService } from '../model-config/model-config.service';
import { RerankService } from './rerank.service';
import { I18nService } from '../i18n/i18n.service';
import { UserSettingService } from '../user-setting/user-setting.service';
import { ChatOpenAI } from '@langchain/openai';
import { ModelConfig } from '../types';

export interface RagSearchResult {
  content: string;
  fileName: string;
  score: number;
  chunkIndex: number;
  fileId?: string;
  originalScore?: number; // Rerank前のスコア（デバッグ用）
  metadata?: any;
}



@Injectable()
export class RagService {
  private readonly logger = new Logger(RagService.name);
  private readonly defaultDimensions: number;

  constructor(
    @Inject(forwardRef(() => ElasticsearchService))
    private elasticsearchService: ElasticsearchService,
    private embeddingService: EmbeddingService,
    private modelConfigService: ModelConfigService,
    private rerankService: RerankService,
    private configService: ConfigService,
    private i18nService: I18nService,
    private userSettingService: UserSettingService,
  ) {
    this.defaultDimensions = parseInt(
      this.configService.get<string>('DEFAULT_VECTOR_DIMENSIONS', '2560'),
    );
    this.logger.log(`RAG サービスのデフォルトベクトル次元数: ${this.defaultDimensions}`);
  }

  async searchKnowledge(
    query: string,
    userId: string,
    topK: number = 5,
    vectorSimilarityThreshold: number = 0.3, // ベクトル検索のしきい値
    embeddingModelId?: string,
    enableFullTextSearch: boolean = false,
    enableRerank: boolean = false,
    rerankModelId?: string,
    selectedGroups?: string[],
    effectiveFileIds?: string[],
    rerankSimilarityThreshold: number = 0.5, // Rerankのしきい値（デフォルト0.5）
    tenantId?: string, // New
    enableQueryExpansion?: boolean,
    enableHyDE?: boolean,
  ): Promise<RagSearchResult[]> {
    // 1. グローバル設定の取得
    const globalSettings = await this.userSettingService.getGlobalSettings();

    // パラメータが明示的に渡されていない場合はグローバル設定を使用
    const effectiveTopK = topK || globalSettings.topK || 5;
    const effectiveVectorThreshold = vectorSimilarityThreshold !== undefined ? vectorSimilarityThreshold : (globalSettings.similarityThreshold || 0.3);
    const effectiveRerankThreshold = rerankSimilarityThreshold !== undefined ? rerankSimilarityThreshold : (globalSettings.rerankSimilarityThreshold || 0.5);
    const effectiveEnableRerank = enableRerank !== undefined ? enableRerank : globalSettings.enableRerank;
    const effectiveEnableFullText = enableFullTextSearch !== undefined ? enableFullTextSearch : globalSettings.enableFullTextSearch;
    const effectiveEmbeddingId = embeddingModelId || globalSettings.selectedEmbeddingId;
    const effectiveRerankId = rerankModelId || globalSettings.selectedRerankId;
    const effectiveHybridWeight = globalSettings.hybridVectorWeight ?? 0.7;
    const effectiveEnableQueryExpansion = enableQueryExpansion !== undefined ? enableQueryExpansion : globalSettings.enableQueryExpansion;
    const effectiveEnableHyDE = enableHyDE !== undefined ? enableHyDE : globalSettings.enableHyDE;

    this.logger.log(
      `RAG search: query="${query}", topK=${effectiveTopK}, vectorThreshold=${effectiveVectorThreshold}, rerankThreshold=${effectiveRerankThreshold}, hybridWeight=${effectiveHybridWeight}, QueryExpansion=${effectiveEnableQueryExpansion}, HyDE=${effectiveEnableHyDE}`,
    );

    try {
      // 1. クエリの準備（拡張または HyDE）
      let queriesToSearch = [query];

      if (effectiveEnableHyDE) {
        const hydeDoc = await this.generateHyDE(query, userId);
        queriesToSearch = [hydeDoc]; // HyDE の場合は仮想ドキュメントをクエリとして使用
      } else if (effectiveEnableQueryExpansion) {
        const expanded = await this.expandQuery(query, userId);
        queriesToSearch = [...new Set([query, ...expanded])];
      }

      // 埋め込みモデルIDが提供されているか確認
      if (!effectiveEmbeddingId) {
        throw new Error('埋め込みモデルIDが提供されていません');
      }

      // 2. 複数のクエリに対して並列検索
      const searchTasks = queriesToSearch.map(async (searchQuery) => {
        // クエリベクトルの取得
        const queryEmbedding = await this.embeddingService.getEmbeddings(
          [searchQuery],
          userId,
          effectiveEmbeddingId,
        );
        const queryVector = queryEmbedding[0];

        // 設定に基づいた検索戦略の選択
        let results;
        if (effectiveEnableFullText) {
          results = await this.elasticsearchService.hybridSearch(
            queryVector,
            searchQuery,
            userId,
            effectiveTopK * 4,
            effectiveHybridWeight,
            undefined,
            effectiveFileIds,
            tenantId,
          );
        } else {
          let vectorSearchResults = await this.elasticsearchService.searchSimilar(
            queryVector,
            userId,
            effectiveTopK * 4,
            tenantId,
          );
          if (effectiveFileIds && effectiveFileIds.length > 0) {
            results = vectorSearchResults.filter(r => effectiveFileIds.includes(r.fileId));
          } else {
            results = vectorSearchResults;
          }
        }
        return results;
      });

      const allResultsRaw = await Promise.all(searchTasks);
      let searchResults = this.deduplicateResults(allResultsRaw.flat());

      // 初回の類似度フィルタリング
      const initialCount = searchResults.length;

      // ログ出力
      searchResults.forEach((r, idx) => {
        this.logger.log(`Hit ${idx}: score=${r.score.toFixed(4)}, fileName=${r.fileName}`);
      });

      // 閾値フィルタリングを適用
      searchResults = searchResults.filter(r => r.score >= effectiveVectorThreshold);
      this.logger.log(`Initial hits: ${initialCount} -> filtered by vectorThreshold: ${searchResults.length}`);

      // 3. リランク (Rerank)
      let finalResults = searchResults;

      if (effectiveEnableRerank && effectiveRerankId && searchResults.length > 0) {
        try {
          const docs = searchResults.map(r => r.content);
          const rerankedIndices = await this.rerankService.rerank(
            query,
            docs,
            userId,
            effectiveRerankId,
            effectiveTopK * 2 // 少し多めに残す
          );

          finalResults = rerankedIndices.map(r => {
            const originalItem = searchResults[r.index];
            return {
              ...originalItem,
              score: r.score, // Rerank スコア
              originalScore: originalItem.score // 元のスコア
            };
          });

          // Rerank後のフィルタリング
          const beforeRerankFilter = finalResults.length;
          finalResults = finalResults.filter(r => r.score >= effectiveRerankThreshold);
          this.logger.log(`After rerank: ${beforeRerankFilter} -> filtered by rerankThreshold: ${finalResults.length}`);

        } catch (error) {
          this.logger.warn(`Rerank failed, falling back to filtered vector search: ${error.message}`);
          // 失敗した場合はベクトル検索の結果をそのまま使う
        }
      }

      // 最終的な件数制限
      finalResults = finalResults.slice(0, effectiveTopK);

      // 4. RAG 結果形式に変換
      const ragResults: RagSearchResult[] = finalResults.map((result) => ({
        content: result.content,
        fileName: result.fileName,
        score: result.score,
        originalScore: result.originalScore !== undefined ? result.originalScore : result.score,
        chunkIndex: result.chunkIndex,
        fileId: result.fileId,
        metadata: result.metadata,
      }));

      return ragResults;
    } catch (error) {
      this.logger.error('RAG search failed:', error);
      return [];
    }
  }

  buildRagPrompt(
    query: string,
    searchResults: RagSearchResult[],
    language: string = 'ja',
  ): string {
    const lang = language || 'ja';

    // コンテキストの構築
    let context = '';
    if (searchResults.length === 0) {
      context = this.i18nService.getMessage('ragNoDocumentFound', lang);
    } else {
      // ファイルごとにグループ化
      const fileGroups = new Map<string, RagSearchResult[]>();
      searchResults.forEach((result) => {
        if (!fileGroups.has(result.fileName)) {
          fileGroups.set(result.fileName, []);
        }
        fileGroups.get(result.fileName)!.push(result);
      });

      // コンテキスト文字列を構築
      const contextParts: string[] = [];
      fileGroups.forEach((chunks, fileName) => {
        contextParts.push(this.i18nService.formatMessage('ragSource', { fileName }, lang));
        chunks.forEach((chunk, index) => {
          contextParts.push(
            this.i18nService.formatMessage('ragSegment', {
              index: index + 1,
              score: chunk.score.toFixed(3)
            }, lang),
          );
          contextParts.push(chunk.content);
          contextParts.push('');
        });
      });
      context = contextParts.join('\n');
    }

    const langText =
      lang === 'zh' ? '中文' : lang === 'en' ? 'English' : '日本語';

    const systemPrompt = this.i18nService.getMessage('ragSystemPrompt', lang);
    const rules = this.i18nService.formatMessage('ragRules', { lang: langText }, lang);
    const docContentHeader = this.i18nService.getMessage('ragDocumentContent', lang);
    const userQuestionHeader = this.i18nService.getMessage('ragUserQuestion', lang);
    const answerHeader = this.i18nService.getMessage('ragAnswer', lang);

    return `${systemPrompt}

${rules}

${docContentHeader}
${context}

${userQuestionHeader}
${query}

${answerHeader}`;
  }

  extractSources(searchResults: RagSearchResult[]): string[] {
    const uniqueFiles = new Set<string>();
    searchResults.forEach((result) => {
      uniqueFiles.add(result.fileName);
    });
    return Array.from(uniqueFiles);
  }

  /**
   * 検索結果の重複排除
   */
  private deduplicateResults(results: any[]): any[] {
    const unique = new Map<string, any>();
    results.forEach(r => {
      const key = `${r.fileId}_${r.chunkIndex}`;
      if (!unique.has(key) || unique.get(key)!.score < r.score) {
        unique.set(key, r);
      }
    });
    return Array.from(unique.values()).sort((a, b) => b.score - a.score);
  }

  /**
   * クエリを拡張してバリエーションを生成
   */
  async expandQuery(query: string, userId: string, tenantId?: string): Promise<string[]> {
    try {
      const llm = await this.getInternalLlm(userId, tenantId || 'default');
      if (!llm) return [query];

      const userSettings = await this.userSettingService.findOrCreate(userId);
      const lang = userSettings.language || 'ja';
      const prompt = this.i18nService.formatMessage('queryExpansionPrompt', { query }, lang);

      const response = await llm.invoke(prompt);
      const content = String(response.content);

      const expandedQueries = content
        .split('\n')
        .map(q => q.trim())
        .filter(q => q.length > 0)
        .slice(0, 3); // 最大3つに制限

      this.logger.log(`Query expanded: "${query}" -> [${expandedQueries.join(', ')}]`);
      return expandedQueries.length > 0 ? expandedQueries : [query];
    } catch (error) {
      this.logger.error('Query expansion failed:', error);
      return [query];
    }
  }

  /**
   * 仮想的なドキュメント（HyDE）を生成
   */
  async generateHyDE(query: string, userId: string, tenantId?: string): Promise<string> {
    try {
      const llm = await this.getInternalLlm(userId, tenantId || 'default');
      if (!llm) return query;

      const userSettings = await this.userSettingService.findOrCreate(userId);
      const lang = userSettings.language || 'ja';
      const prompt = this.i18nService.formatMessage('hydePrompt', { query }, lang);

      const response = await llm.invoke(prompt);
      const hydeDoc = String(response.content).trim();

      this.logger.log(`HyDE generated for: "${query}" (length: ${hydeDoc.length})`);
      return hydeDoc || query;
    } catch (error) {
      this.logger.error('HyDE generation failed:', error);
      return query;
    }
  }

  /**
   * 内部タスク用の LLM インスタンスを取得
   */
  private async getInternalLlm(userId: string, tenantId: string): Promise<ChatOpenAI | null> {
    try {
      const models = await this.modelConfigService.findAll(userId, tenantId || 'default');
      const defaultLlm = models.find(m => m.type === 'llm' && m.isDefault && m.isEnabled !== false);

      if (!defaultLlm) {
        this.logger.warn('No default LLM configured for internal tasks');
        return null;
      }

      return new ChatOpenAI({
        apiKey: defaultLlm.apiKey || 'ollama',
        temperature: 0.3,
        modelName: defaultLlm.modelId,
        configuration: {
          baseURL: defaultLlm.baseUrl || 'http://localhost:11434/v1',
        },
      });
    } catch (error) {
      this.logger.error('Failed to get internal LLM:', error);
      return null;
    }
  }
}