BaaS-MCP

import {BM25_CONFIGS, BM25Config, MIN_SCORE_RATIO, SearchMode,} from "../constants/search-mode.js"; import {BaaSDocument} from "./baas-document.js"; import {DocumentChunk} from "./types.js"; type Score = { id: number; chunkId: number; score: number; totalTF: number }; export class BaaSBM25Calculator { private readonly allChunks: DocumentChunk[]; private readonly totalCount: number; private readonly averageDocLength: number; private readonly N: number; constructor(documents: BaaSDocument[]) { this.allChunks = documents.flatMap((doc) => doc.getChunks()); this.totalCount = this.allChunks.reduce( (count, doc) => count + doc.wordCount, 0 ); this.averageDocLength = this.totalCount / this.allChunks.length; this.N = this.allChunks.length; } calculate(keywords: string[], searchMode: SearchMode): BM25Result[] { const regexp = this.createRegexp(keywords); const config = BM25_CONFIGS[searchMode]; const { termFrequencies, docFrequencies } = this.calculateFrequencies(regexp); const scores = this.calculateScore(termFrequencies, docFrequencies, config); scores.sort((a, b) => b.score !== a.score ? b.score - a.score : b.totalTF - a.totalTF ); const filteredScores = this.applyRelativeFilter(scores, searchMode); return filteredScores.map(({ id, score, chunkId }) => ({ id, chunkId, score, })); } /** * 최고 점수 대비 상대적 임계값을 적용하여 필터링 */ private applyRelativeFilter( scores: Score[], searchMode: SearchMode ): Score[] { const minScoreRatio = MIN_SCORE_RATIO[searchMode]; if (scores.length === 0) return scores; const maxScore = scores[0].score; const relativeThreshold = maxScore * this.getThresholdRatio(minScoreRatio); const filtered = scores.filter((s) => s.score >= relativeThreshold); // 최소 1개는 보장 (최고점수 문서는 항상 포함) return filtered.length > 0 ? filtered : [scores[0]]; } /** * minScore 값을 상대적 비율로 매핑 */ private getThresholdRatio(minScore: number): number { // minScore 0.1~1.0 -> 비율 0.05~0.7로 매핑 return Math.max(0.05, Math.min(0.7, minScore * 0.7)); } private createRegexp(keywords: string[]) { const set = new Set( keywords.map((keyword) => this.escapeRegExp(keyword.trim())) ); const query = Array.from(set).join("|"); return new RegExp(query, "gi"); } private calculateFrequencies(regexp: RegExp) { const termFrequencies: Record<number, Record<string, number>> = {}; const docFrequencies: Record<string, number> = {}; for (const doc of this.allChunks) { const text = doc.text; const matches = Array.from(text.matchAll(regexp)); const termCounts: Record<string, number> = {}; for (const match of matches) { const term = match[0].toLowerCase(); termCounts[term] = (termCounts[term] || 0) + 1; } if (Object.keys(termCounts).length > 0) { termFrequencies[doc.chunkId] = termCounts; for (const term of Object.keys(termCounts)) { docFrequencies[term] = (docFrequencies[term] || 0) + 1; } } } return { termFrequencies, docFrequencies }; } private calculateScore( termFrequencies: Record<number, Record<string, number>>, docFrequencies: Record<string, number>, config: BM25Config ): Score[] { return this.allChunks .filter((chunk) => termFrequencies[chunk.chunkId]) .map((chunk) => { const tf = termFrequencies[chunk.chunkId]; const len = chunk.wordCount; const score = Object.keys(tf) .map((term) => { const df = docFrequencies[term]; const idf = Math.max(0.001, Math.log((this.N - df + 0.5) / (df + 0.5))); const numerator = tf[term] * (config.k1 + 1); const denominator = tf[term] + config.k1 * (1 - config.b + config.b * (len / this.averageDocLength)); return idf * (numerator / denominator); }) .reduce((sum, v) => sum + v, 0); const totalTF = Object.values(tf).reduce((sum, v) => sum + v, 0); return { id: chunk.id, chunkId: chunk.chunkId, score, totalTF, }; }); } private escapeRegExp(string: string) { return string.replace(/[.*+?^${}()|[\]\\]/g, "\\$&"); } } export interface BM25Result { id: number; chunkId: number; score: number; }