o2o-site-ontology/scripts/ingest-dataset.ts

/**
 * data/gunsan-pension-keywords.json 을 pgvector 에 적재한다.
 *   npx tsx scripts/ingest-dataset.ts
 *
 * 정책: 주기 수집 없음. 고정 데이터셋 1회 적재.
 * 중복제거는 어휘 단계(정규화 완전일치)만 자동 병합하고,
 * 벡터 유사도는 자동 병합하지 않고 "검토 목록"으로만 뽑는다. (이유는 README 참조)
 */
import { readFileSync } from 'node:fs';
import { createSql, toVector } from '../src/db/db';
import { normalizeKeyword, canonicalizeKeyword, isBanned } from '../src/keywords/normalize';
import { LocalEmbeddingProvider } from '../src/embedding/local.provider';
import { MockEmbeddingProvider } from '../src/embedding/mock.provider';
import { env } from '../src/config/env';

type Item = { keyword: string; intent: string; kind: string; category: string; relevance: number };

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

  const raw = JSON.parse(readFileSync('data/gunsan-pension-keywords.json', 'utf8'));
  const items: Item[] = raw.items;
  console.log(`📦 데이터셋 ${items.length}건  ·  임베딩 ${embedder.name}`);

  // 1) 파일 내 어휘 중복 정리
  const byNorm = new Map<string, { item: Item; aliases: string[] }>();
  let banned = 0;
  for (const it of items) {
    const canonical = canonicalizeKeyword(it.keyword);
    const norm = normalizeKeyword(it.keyword);
    if (!norm || isBanned(canonical)) { banned++; continue; }
    const hit = byNorm.get(norm);
    if (hit) {
      if (!hit.aliases.includes(canonical) && hit.item.keyword !== canonical) hit.aliases.push(canonical);
      if (it.relevance > hit.item.relevance) hit.item = it;
    } else {
      byNorm.set(norm, { item: { ...it, keyword: canonical }, aliases: [] });
    }
  }
  const uniq = [...byNorm.entries()];
  console.log(`   어휘 중복제거 → ${uniq.length}건 (병합 ${items.length - uniq.length - banned}, 금칙어 ${banned})`);

  // 2) 임베딩 (배치)
  const t0 = Date.now();
  const vecs = await embedder.embed(uniq.map(([, v]) => v.item.keyword), 'passage');
  console.log(`   임베딩 ${vecs.length}건 · ${embedder.dimensions}차원 · ${Date.now() - t0}ms`);

  // 3) 적재
  const region = 'kr.jeonbuk.gunsan';
  const industry = 'stay.pension';
  let inserted = 0, updated = 0;
  await sql.begin(async (tx) => {
    for (let i = 0; i < uniq.length; i++) {
      const [norm, v] = 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 (${v.item.keyword}, ${norm}, 'ko-KR', ${v.aliases},
                ${v.item.intent}::keyword_intent, ${v.item.kind}, ${v.item.category}, 'dataset',
                ${industry}, ${region}, ${toVector(vecs[i])}::vector)
        ON CONFLICT (normalized, locale) DO UPDATE SET
          canonical = EXCLUDED.canonical, aliases = EXCLUDED.aliases, intent = EXCLUDED.intent,
          kind = EXCLUDED.kind, category = EXCLUDED.category, source = EXCLUDED.source,
          embedding = EXCLUDED.embedding, updated_at = now()
        RETURNING (xmax = 0) AS inserted`;
      res[0]?.inserted ? inserted++ : updated++;
      if (i % 100 === 0) process.stdout.write(`\r   적재 ${i}/${uniq.length}`);
    }
  });
  console.log(`\r   ✅ 신규 ${inserted} · 갱신 ${updated}                    `);

  // 4) 벡터 근접쌍 — 자동 병합하지 않고 검토 목록으로만
  const near = await sql<Array<{ a: string; b: string; sim: number }>>`
    SELECT k1.canonical AS a, k2.canonical AS b, 1 - (k1.embedding <=> k2.embedding) AS sim
    FROM keyword k1 JOIN keyword k2
      ON k1.id < k2.id AND k1.embedding <=> k2.embedding < 0.02
    WHERE k1.source = 'dataset' AND k2.source = 'dataset'
    ORDER BY sim DESC LIMIT 15`;
  console.log(`\n🔍 벡터 근접쌍 검토 목록 (cos ≥ 0.98, 자동 병합 안 함) — 상위 ${near.length}`);
  for (const n of near) console.log(`   ${Number(n.sim).toFixed(4)}  ${n.a}  ↔  ${n.b}`);

  await sql.end();
}

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