「こういうデータがあったのか」を発見させる：非エンジニア向けデータカタログUX設計

glossary:  - business_term: "顧客生涯価値 (LTV)"    aliases: ["LTV", "顧客 LTV", "Lifetime Value", "生涯価値"]    definition: |      ある顧客が初回購入から最後の購入までに支払った合計金額。      キャンセル・返金は減算後。離反後の追加購入は含む。    formula: |      SUM(amount_jpy - refund_jpy) GROUP BY customer_id    primary_table: fact_user_revenue    primary_column: cumulative_amount    owners: [team:commerce]    related_terms: ["RFM", "ARPU", "離反率"]
  - business_term: "離反率 (Churn Rate)"    aliases: ["Churn", "解約率", "退会率"]    definition: |      期初に在籍していた顧客のうち、期末までに離反した割合。      離反の定義: 90 日間購入が無い状態。    formula: "...省略..."    primary_table: dim_customers    primary_column: churn_flag    owners: [team:retention]

import anthropicfrom anthropic import Anthropicimport numpy as np
client = Anthropic()# 各テーブル/カラム/グロッサリを1つの「ドキュメント」として extends してから埋め込むdocs = []for term, info in glossary.items():    docs.append({        "id": f"glossary:{term}",        "text": f"業務語彙: {term}\n別名: {', '.join(info['aliases'])}\n定義: {info['definition']}\nテーブル: {info['primary_table']}",    })for table_name, table in metadata_tables.items():    docs.append({        "id": f"table:{table_name}",        "text": f"テーブル: {table_name}\n説明: {table['description']}\nカラム: {', '.join(table['columns'].keys())}",    })
# 埋め込み（実装は省略 — Voyage AI / Cohere / OpenAI などを使う）embeddings = embed([d["text"] for d in docs])
def discover(question: str, top_k: int = 5):    q_emb = embed([question])[0]    sims = np.dot(embeddings, q_emb)    top_idx = np.argsort(sims)[::-1][:top_k]    return [docs[i] for i in top_idx]
# 例print(discover("売上の月次データはどこ？"))# → glossary:LTV, table:fact_user_revenue, table:fact_orders, ...

WITH refs AS (  SELECT    referenced_tables AS rt,    user_email,  FROM `region-us`.INFORMATION_SCHEMA.JOBS_BY_PROJECT,       UNNEST(referenced_tables) AS referenced_tables  WHERE creation_time >= TIMESTAMP_SUB(CURRENT_TIMESTAMP(), INTERVAL 7 DAY)    AND state = 'DONE')SELECT  CONCAT(rt.project_id, '.', rt.dataset_id, '.', rt.table_id) AS table_full,  COUNT(DISTINCT user_email) AS unique_users,  COUNT(*) AS query_countFROM refsGROUP BY 1ORDER BY unique_users DESCLIMIT 10;