o2o-site-AEO/ontology/src/serving/match.service.ts

import { Inject, Injectable } from '@nestjs/common';
import { Sql, toVector } from '../db/db';
import { PG } from '../db/db.module';
import { EmbeddingProvider } from '../embedding/types';
import { normalizeKeyword } from '../keywords/normalize';
import { MerchantWithTaxonomy } from '../merchants/merchants.service';
import {
  AreaGroup, Lane, MerchantFacts, buildLanes, checkAmenity, detectAreaGroup,
  keywordAreaGroup, normalizeAmenities, violatesCapacity,
} from './match.rules';

// RRF 상수를 관례값 60 대신 20 으로 낮춘다. 60 이면 1위와 40위의 기여도 차이가 1.6배뿐이라
// 깊은 순위의 generic 키워드가 여러 레인에서 조금씩 쌓아 상위를 차지한다. 20 이면 2.9배로 벌어진다.
const RRF_K = 20;
const LANE_DEPTH = 50;    // 레인당 후보 깊이 — 깊을수록 generic 이 유리해진다
const LANE_FLOOR = 0.80;  // 이 코사인 미만은 그 레인에서 기여하지 않는다
// 매칭 후보로 인정하는 출처. 고정 데이터셋 정책상 LLM 생성물은 사전에 섞이면 안 된다.
const MATCH_SOURCES = ['dataset', 'nationwide', 'manual'];

interface Hit {
  id: string; canonical: string; intent: string; kind: string;
  category: string | null; aliases: string[]; score: number;
}

export interface MatchRow extends Hit {
  rrf: number;
  status: 'ok' | 'hold';
  holdReason?: string;
  lanes: Array<{ key: string; label: string; rank: number; score: number }>;
  linked: boolean;
}

@Injectable()
export class MatchService {
  constructor(
    @Inject(PG) private readonly sql: Sql,
    private readonly embedder: EmbeddingProvider,
  ) {}

  /** 속성별 서브 질의 → 가중 RRF 융합 → 사실 기반 필터 */
  async fusion(rawQuery: string, limit: number) {
    const query = rawQuery.trim();
    const merchant = await this.resolveMerchant(query);
    const facts = merchant ? toFacts(merchant) : null;

    const lanes: Lane[] = facts
      ? buildLanes(facts)
      : [{ key: 'free', label: '입력문', weight: 1.0, text: query }];

    const vectors = await this.embedder.embed(lanes.map((l) => l.text), 'query');

    // 레인별 검색.
    // 후보 풀을 업체 업종으로 좁힌다. 사전 전체를 뒤지면 '강남 미용실' 같은
    // 다른 업종 키워드가 후보에 섞인다 (실제로 섞여 있었다).
    const perLane = await Promise.all(
      vectors.map((v) => this.laneSearch(v, LANE_DEPTH, merchant?.industry_id ?? null)),
    );

    // 가중 RRF 융합
    const acc = new Map<string, { hit: Hit; rrf: number; lanes: MatchRow['lanes'] }>();
    perLane.forEach((hits, li) => {
      const lane = lanes[li];
      hits.forEach((hit, idx) => {
        if (hit.score < LANE_FLOOR) return;
        const rank = idx + 1;
        const contrib = lane.weight / (RRF_K + rank);
        const cur = acc.get(hit.id) ?? { hit, rrf: 0, lanes: [] };
        cur.rrf += contrib;
        cur.lanes.push({ key: lane.key, label: lane.label, rank, score: hit.score });
        if (hit.score > cur.hit.score) cur.hit = hit;
        acc.set(hit.id, cur);
      });
    });

    // 사실 기반 필터
    const kept: MatchRow[] = [];
    const excluded: Array<{ canonical: string; reason: string }> = [];
    for (const { hit, rrf, lanes: ls } of acc.values()) {
      if (facts) {
        if (violatesCapacity(hit.canonical, facts.capacityMax)) {
          excluded.push({ canonical: hit.canonical, reason: `최대 ${facts.capacityMax}인 — 단체 키워드` });
          continue;
        }
        const kwArea = keywordAreaGroup(hit.canonical);
        if (kwArea && facts.areaGroup && kwArea !== facts.areaGroup) {
          excluded.push({ canonical: hit.canonical, reason: `권역 불일치 — ${kwArea} (업체는 ${facts.areaGroup})` });
          continue;
        }
        const am = checkAmenity(hit.canonical, facts);
        if (!am.ok && !am.hold) {
          excluded.push({ canonical: hit.canonical, reason: `미보유 시설 — ${am.amenity}` });
          continue;
        }
        kept.push({
          ...hit, rrf,
          lanes: ls.sort((a, b) => a.rank - b.rank),
          status: am.ok ? 'ok' : 'hold',
          holdReason: am.ok ? undefined : `${am.amenity} 미확인 — 사업자 확인 필요`,
          linked: false,
        });
      } else {
        kept.push({ ...hit, rrf, lanes: ls.sort((a, b) => a.rank - b.rank), status: 'ok', linked: false });
      }
    }

    kept.sort((a, b) => b.rrf - a.rrf);
    const top = kept.slice(0, limit);

    // 레인별 상위 — SEO 페이지 배분은 평평한 순위가 아니라 이쪽을 쓴다.
    // (주력 키워드는 유형 레인 1위, 주변 여행 페이지는 위치 레인 상위)
    const keptById = new Map(kept.map((k) => [k.id, k]));
    const byLane = lanes.map((lane, li) => ({
      key: lane.key, label: lane.label, weight: lane.weight, text: lane.text,
      items: perLane[li]
        .map((h) => keptById.get(h.id))
        .filter((x): x is MatchRow => Boolean(x))
        .slice(0, 8),
    }));

    await this.markLinked([...top, ...byLane.flatMap((l) => l.items)], merchant?.id ?? null);

    return {
      mode: 'fusion' as const,
      input: query,
      resolved: merchant ? publicMerchant(merchant) : null,
      facts: facts && {
        areaGroup: facts.areaGroup,
        capacityMax: facts.capacityMax,
        amenities: [...facts.amenities],
        unverified: [...facts.unverified],
      },
      lanes: lanes.map((l, i) => ({
        ...l, top: perLane[i][0]?.canonical ?? null, topScore: perLane[i][0]?.score ?? null,
      })),
      embeddingProvider: this.embedder.name,
      total: await this.dictionarySize(),
      matches: top,
      byLane,
      excluded: excluded.slice(0, 40),
      excludedTotal: excluded.length,
    };
  }

  private async laneSearch(
    embedding: number[], limit: number, industryId: string | null,
  ): Promise<Hit[]> {
    const vec = toVector(embedding);
    const rows = await this.sql<Hit[]>`
      SELECT id, canonical, intent, kind, category, aliases,
             1 - (embedding <=> ${vec}::vector) AS score
      FROM keyword
      WHERE embedding IS NOT NULL
        AND source = ANY(${MATCH_SOURCES})
        AND (${industryId}::text IS NULL OR industry_id IS NULL OR industry_id = ${industryId})
      ORDER BY embedding <=> ${vec}::vector
      LIMIT ${limit}`;
    return rows.map((r) => ({ ...r, score: Number(r.score) }));
  }

  private async markLinked(rows: MatchRow[], merchantId: string | null) {
    if (!merchantId || rows.length === 0) return;
    const ids = rows.map((r) => r.id);
    const linked = await this.sql<Array<{ keyword_id: string }>>`
      SELECT keyword_id FROM merchant_keyword
      WHERE merchant_id = ${merchantId} AND keyword_id = ANY(${ids}::uuid[])`;
    const set = new Set(linked.map((l) => l.keyword_id));
    for (const r of rows) r.linked = set.has(r.id);
  }

  async resolveMerchant(query: string) {
    const norm = normalizeKeyword(query);
    if (!norm) return null;
    const rows = await this.sql<Array<MerchantWithTaxonomy & { sim: number }>>`
      SELECT m.*,
             i.name AS industry_name, i.path::text AS industry_path,
             r.name AS region_name,   r.path::text AS region_path,
             similarity(regexp_replace(lower(m.name), '\\s', '', 'g'), ${norm}) AS sim
      FROM merchant m
      LEFT JOIN industry i ON i.id = m.industry_id
      LEFT JOIN region   r ON r.id = m.region_id
      WHERE regexp_replace(lower(m.name), '\\s', '', 'g') = ${norm}
         OR m.external_id = ${query}
         OR similarity(regexp_replace(lower(m.name), '\\s', '', 'g'), ${norm}) >= 0.45
      ORDER BY sim DESC NULLS LAST
      LIMIT 1`;
    return rows[0] ?? null;
  }

  async dictionarySize() {
    const [row] = await this.sql<Array<{ n: number }>>`
      SELECT count(*)::int AS n FROM keyword
      WHERE embedding IS NOT NULL AND source = ANY(${MATCH_SOURCES})`;
    return row?.n ?? 0;
  }
}

function str(v: unknown): string[] {
  return Array.isArray(v) ? v.filter((x): x is string => typeof x === 'string') : [];
}

function toFacts(m: MerchantWithTaxonomy): MerchantFacts {
  const p = (m.profile ?? {}) as Record<string, unknown>;
  const features = str(p['features']);
  const services = str(p['services']);
  const nearby = str(p['nearby']);
  const address = typeof p['address'] === 'string' ? p['address'] : null;
  const cap = p['capacity'] as { max?: number } | undefined;

  const areaSource = [address ?? '', ...nearby, m.description].join(' ');
  const areaGroup: AreaGroup | null = detectAreaGroup(areaSource);

  return {
    name: m.name,
    region: m.region_name,
    industry: m.industry_name,
    description: m.description,
    address,
    areaGroup,
    capacityMax: typeof cap?.max === 'number' ? cap.max : null,
    services, features,
    audiences: str(p['audiences']),
    nearby,
    amenities: normalizeAmenities([...features, ...services]),
    unverified: new Set(str(p['unverified'])),
    signals: collectSignals(p),
  };
}

/**
 * 고객 언어 신호를 모은다.
 *   hashtags     : ["#군산독채", "#군산감성숙소", ...]   인스타 등
 *   reviewSignals: [{ term: "바베큐", count: 47 }, ...]  리뷰 원문이 아닌 빈도 집계
 * 리뷰 원문은 받지 않는다 (저작권·개인정보). 빈도만으로 충분하다.
 */
function collectSignals(p: Record<string, unknown>): string[] {
  const tags = str(p['hashtags']).map((t) => t.replace(/^#/, '').trim()).filter(Boolean);
  const raw = Array.isArray(p['reviewSignals']) ? p['reviewSignals'] : [];
  const reviews = raw
    .filter((r): r is { term: string; count: number } =>
      Boolean(r) && typeof (r as any).term === 'string')
    .sort((a, b) => (b.count ?? 0) - (a.count ?? 0))
    .map((r) => r.term.trim())
    .filter(Boolean);
  return [...new Set([...reviews, ...tags])];
}

function publicMerchant(m: MerchantWithTaxonomy) {
  return {
    id: m.id, externalId: m.external_id, name: m.name,
    region: m.region_name, industry: m.industry_name,
    description: m.description, siteUrl: m.site_url, profile: m.profile ?? {},
  };
}