o2o-site-AEO/ontology/scripts/ingest-nationwide.ts

/** 전국 지역별 펜션 키워드를 pgvector 에 적재한다. */
import { readFileSync } from 'node:fs';
import { createSql, toVector } from '../src/db/db';
import { canonicalizeKeyword, isBanned, normalizeKeyword } from '../src/keywords/normalize';
import { LocalEmbeddingProvider } from '../src/embedding/local.provider';
import { MockEmbeddingProvider } from '../src/embedding/mock.provider';
import { env } from '../src/config/env';

const SOURCE = 'nationwide';

interface Item {
  sido: string; region: string; regionKey: string; regionType: string;
  keyword: string; kind: string; intent: string; category: string; tier: string; relevance: number;
}

async function main() {
  const sql = createSql();
  const embedder =
    env.embedding.provider === 'mock' ? new MockEmbeddingProvider() : new LocalEmbeddingProvider();

  const ds = JSON.parse(readFileSync('data/nationwide-pension-keywords.json', 'utf8'));
  const items: Item[] = ds.items;
  const regions = JSON.parse(readFileSync('data/regions.json', 'utf8')).regions as
    Array<{ sido: string; name: string; key: string }>;
  console.log(`📦 ${items.length}건 / ${ds.regionCount}개 지역 · 임베딩 ${embedder.name}`);

  // 1) 지역 계층 심기 (시도 → 시군).
  const nodes = new Map<string, string>();
  for (const r of regions) {
    const sidoKey = r.key.split('.').slice(0, -1).join('.');
    nodes.set(sidoKey, r.sido);
    nodes.set(r.key, r.name);
  }
  nodes.set('kr', '대한민국');
  for (const [key, name] of [...nodes].sort((a, b) => a[0].length - b[0].length)) {
    await sql`INSERT INTO region (id, path, name) VALUES (${key}, ${key}::ltree, ${name})
              ON CONFLICT (id) DO UPDATE SET name = EXCLUDED.name`;
  }
  console.log(`   🗺  지역 노드 ${nodes.size}개 등록`);

  // 2) 어휘 중복 정리 — 키는 (지역, 정규화 키워드)
  const byKey = new Map<string, { item: Item; norm: string }>();
  let banned = 0;
  for (const it of items) {
    const canonical = canonicalizeKeyword(it.keyword);
    const n = normalizeKeyword(canonical);
    if (!n || isBanned(canonical)) { banned++; continue; }
    const k = `${it.regionKey}|${n}`;
    if (!byKey.has(k)) byKey.set(k, { item: { ...it, keyword: canonical }, norm: n });
  }
  const uniq = [...byKey.values()];
  console.log(`   어휘 중복제거 → ${uniq.length}건 (금칙어 ${banned})`);

  // 3) 임베딩
  const t0 = Date.now();
  const vecs = await embedder.embed(uniq.map((u) => u.item.keyword), 'passage');
  console.log(`   임베딩 ${vecs.length}건 · ${embedder.dimensions}차원 · ${Date.now() - t0}ms`);

  // 4) 적재.
  let inserted = 0, updated = 0;
  await sql.begin(async (tx) => {
    for (let i = 0; i < uniq.length; i++) {
      const { item, norm } = uniq[i];
      const res = await tx<Array<{ inserted: boolean }>>`
        INSERT INTO keyword
          (canonical, normalized, locale, aliases, intent, kind, category, source,
           industry_id, region_id, embedding)
        VALUES (${item.keyword}, ${norm}, 'ko-KR', ${[]},
                ${item.intent}::keyword_intent, ${item.kind}, ${item.category}, ${SOURCE},
                'stay.pension', ${item.kind === 'tag' ? null : item.regionKey},
                ${toVector(vecs[i])}::vector)
        ON CONFLICT (normalized, locale) DO UPDATE SET
          canonical = EXCLUDED.canonical, intent = EXCLUDED.intent, kind = EXCLUDED.kind,
          category = EXCLUDED.category, source = EXCLUDED.source,
          region_id = COALESCE(keyword.region_id, EXCLUDED.region_id),
          embedding = EXCLUDED.embedding, updated_at = now()
        RETURNING (xmax = 0) AS inserted`;
      res[0]?.inserted ? inserted++ : updated++;
      if (i % 500 === 0) process.stdout.write(`\r   적재 ${i}/${uniq.length}`);
    }
  });
  console.log(`\r   ✅ 신규 ${inserted} · 갱신 ${updated}                    `);

  // 5) 이 출처 안에서만 잔여 정리
  const wanted = uniq.map((u) => u.norm);
  const stale = await sql`
    DELETE FROM keyword WHERE source = ${SOURCE} AND NOT (normalized = ANY(${wanted})) RETURNING id`;
  console.log(`   🧹 이전 판본 잔여 ${stale.length}건 삭제`);

  const counts = await sql<Array<{ source: string; n: number }>>`
    SELECT source, count(*)::int AS n FROM keyword GROUP BY source ORDER BY n DESC`;
  console.log('   📚 사전 현황: ' + counts.map((c) => `${c.source} ${c.n}`).join(' · '));
  await sql.end();
}

main().catch((e) => { console.error('❌', e); process.exit(1); });