{"id":43671,"date":"2025-09-28T22:40:58","date_gmt":"2025-09-28T13:40:58","guid":{"rendered":"https:\/\/vitalify.jp\/blog\/?p=43671"},"modified":"2025-09-28T22:40:58","modified_gmt":"2025-09-28T13:40:58","slug":"ai%e9%96%8b%e7%99%ba%e3%81%af%e8%aa%b2%e9%a1%8c%e8%a7%a3%e6%b1%ba%e6%80%9d%e8%80%83%e3%81%8c%e3%81%be%e3%81%99%e3%81%be%e3%81%99%e5%bf%85%e8%a6%81%e3%81%ab%e3%81%aa%e3%82%8b","status":"publish","type":"post","link":"https:\/\/vitalify.jp\/blog\/2025\/09\/ai%e9%96%8b%e7%99%ba%e3%81%af%e8%aa%b2%e9%a1%8c%e8%a7%a3%e6%b1%ba%e6%80%9d%e8%80%83%e3%81%8c%e3%81%be%e3%81%99%e3%81%be%e3%81%99%e5%bf%85%e8%a6%81%e3%81%ab%e3%81%aa%e3%82%8b.html","title":{"rendered":"AI\u958b\u767a\u306f\u8ab2\u984c\u89e3\u6c7a\u601d\u8003\u304c\u307e\u3059\u307e\u3059\u5fc5\u8981\u306b\u306a\u308b"},"content":{"rendered":"

\u307e\u3044\u3069<\/p>\n

AI\u30bd\u30ea\u30e5\u30fc\u30b7\u30e7\u30f3\u4e8b\u696d\u90e8\u306e\u82b1\u5cf6\u3067\u3059\u3002<\/p>\n

\u6700\u8fd1\u3001\u30c1\u30e3\u30c3\u30c8\u30dc\u30c3\u30c8\u306e\u56de\u7b54\u3067\u3069\u3046\u3084\u3063\u305f\u3089\u6570\u5b57\u3092\u542b\u3081\u305f\u691c\u7d22\u306b\u5bfe\u3057\u3066\u826f\u3044\u56de\u7b54\u304c\u3067\u304d\u308b\u304b\u3092\u8003\u3048\u3066\u3044\u307e\u3059\u3002<\/p>\n

\u30d9\u30af\u30c8\u30eb\u691c\u7d22\u306f\u6570\u5b57\u304c\u82e6\u624b\u306a\u3093\u3067\u3059\u3002Revealing the Numeracy Gap: An Empirical Investigation of Text Embedding Models\u306e\u8ad6\u6587\u306e4.2 Main Results\u306b\u4ee5\u4e0b\u8a18\u8f09\u3055\u308c\u3066\u3044\u307e\u3059\u3002<\/p>\n

The retrieval accuracy, averaged over the 13 embedding models under evaluation, is merely 0.54, slightly above random guessing (0.5). We believe this is mainly because current embedding model training practices typically focus on capturing semantic similarities or differences at the sentence or paragraph level, with fine-grained but potentially important details, such as numbers as part of the text, often overlooked. This aligns with Liu et al. (2024), who find that embedding models generally fall short in distinguishing nuanced differences between texts.<\/p><\/blockquote>\n

Deepl\u3067\u7ffb\u8a33\u3057\u307e\u3057\u305f\u3002<\/p>\n

\u8a55\u4fa1\u5bfe\u8c61\u306e13\u306e\u57cb\u3081\u8fbc\u307f\u30e2\u30c7\u30eb\u3067\u5e73\u5747\u3057\u305f\u691c\u7d22\u7cbe\u5ea6\u306f\u308f\u305a\u304b0.54\u3067\u3042\u308a\u3001\u30e9\u30f3\u30c0\u30e0\u63a8\u6e2c\uff080.5\uff09\u3092\u308f\u305a\u304b\u306b\u4e0a\u56de\u308b\u7a0b\u5ea6\u3067\u3042\u308b\u3002\u3053\u308c\u306f\u4e3b\u306b\u3001\u73fe\u5728\u306e\u57cb\u3081\u8fbc\u307f\u30e2\u30c7\u30eb\u8a13\u7df4\u624b\u6cd5\u304c\u6587\u3084\u6bb5\u843d\u30ec\u30d9\u30eb\u3067\u306e\u610f\u5473\u7684\u985e\u4f3c\u6027\u30fb\u76f8\u9055\u306e\u6355\u6349\u306b\u91cd\u70b9\u3092\u7f6e\u304f\u4e00\u65b9\u3067\u3001\u30c6\u30ad\u30b9\u30c8\u5185\u306e\u6570\u5b57\u306a\u3069\u3001\u7d30\u304b\u304f\u3066\u91cd\u8981\u306a\u8a73\u7d30\u304c\u3057\u3070\u3057\u3070\u898b\u843d\u3068\u3055\u308c\u308b\u305f\u3081\u3068\u8003\u3048\u3089\u308c\u308b\u3002\u3053\u308c\u306fLiu et al. (2024)\u306e\u77e5\u898b\u3068\u4e00\u81f4\u3057\u3066\u304a\u308a\u3001\u540c\u7814\u7a76\u3067\u306f\u57cb\u3081\u8fbc\u307f\u30e2\u30c7\u30eb\u304c\u30c6\u30ad\u30b9\u30c8\u9593\u306e\u5fae\u5999\u306a\u5dee\u7570\u3092\u533a\u5225\u3059\u308b\u80fd\u529b\u306b\u6982\u3057\u3066\u4e0d\u8db3\u304c\u3042\u308b\u3068\u6307\u6458\u3055\u308c\u3066\u3044\u308b\u3002<\/p>\n

\u73fe\u5728\u306e\u57cb\u3081\u8fbc\u307f\u30e2\u30c7\u30eb\u8a13\u7df4\u624b\u6cd5\u304c\u6587\u3084\u6bb5\u843d\u30ec\u30d9\u30eb\u3067\u306e\u610f\u5473\u7684\u985e\u4f3c\u6027\u30fb\u76f8\u9055\u306e\u6355\u6349\u306b\u91cd\u70b9\u3092\u7f6e\u304f\u4e00\u65b9\u3067\u3001\u30c6\u30ad\u30b9\u30c8\u5185\u306e\u6570\u5b57\u306a\u3069\u3001\u7d30\u304b\u304f\u3066\u91cd\u8981\u306a\u8a73\u7d30\u304c\u3057\u3070\u3057\u3070\u898b\u843d\u3068\u3055\u308c\u308b<\/p><\/blockquote>\n

\u305d\u308c\u3092\u8a00\u3063\u3061\u3083\u3042\u304a\u3057\u3081\u3048\u3088\u3068\u3044\u3046\u611f\u3058\u3067\u3059\u3002\u3002\u3002\u3002<\/p>\n

\u30d9\u30af\u30c8\u30eb\u3068\u4e00\u7dd2\u306b\u4fdd\u5b58\u3055\u308c\u3066\u3044\u308b\u30e1\u30bf\u3067\u691c\u7d22\u3059\u308b\uff1f\uff1f\u3068\u3082\u601d\u3044\u307e\u3057\u305f\u304c\u3001<\/p>\n

\u3084\u306f\u308a\u3001\u305d\u308d\u305d\u308dMCP\u3092\u5c0e\u5165\u3057\u3001RAG\u3067\u306f\u306a\u304f\u3001\u691c\u7d22\u3092MCP\u3067\u5f37\u5316\u3057\u305f\u3044\u3068\u601d\u3044\u307e\u3057\u305f\u3002<\/p>\n

\u305d\u3057\u305f\u3089\u3001Vercel\u306e\u4ee5\u4e0b\u306e\u8a18\u4e8b\u3092\u898b\u3064\u3051\u307e\u3057\u305f\u3002<\/p>\n

The second wave of MCP: Building for LLMs, not developers<\/a><\/p>\n

MCP\u3068\u3044\u3048\u3070\u3060\u3044\u305f\u3044WEB\u691c\u7d22\u3057\u307e\u3059\u3068\u304b\u5929\u6c17\u306eAPI\u3092\u4f7f\u3044\u307e\u3057\u305f\u3068\u304bAPI\u306e\u7f6e\u304d\u63db\u3048\u306e\u4f8b\u3048\u8a71\u304c\u30e1\u30a4\u30f3\u3067\u3059\u3002<\/p>\n

MCP\u306f\u30c4\u30fc\u30eb\u306a\u306e\u3067\u3001\u5c0f\u3055\u304f\u4f5c\u308b\u306e\u304c\u6b63\u89e3\u304b\u306a\u3041\u3068\u601d\u3063\u3066\u3044\u305f\u3093\u3067\u3059\u304c\u3001Vercel\u66f0\u304f\u3001\u305d\u3046\u3058\u3083\u306a\u3044\u3068\u3002<\/p>\n

MCP works best when tools handle complete user intentions rather than exposing individual API operations.<\/p><\/blockquote>\n

MCP\u306f\u3001\u500b\u3005\u306eAPI\u64cd\u4f5c\u3092\u30c4\u30fc\u30eb\u3068\u4f7f\u7528\u3059\u308b\u306e\u3067\u306f\u306a\u304f\u3001\u30c4\u30fc\u30eb\u304c\u30e6\u30fc\u30b6\u30fc\u306e\u610f\u56f3\u5168\u4f53\u3092\u51e6\u7406\u3059\u308b\u5834\u5408\u306b\u6700\u3082\u52b9\u679c\u3092\u767a\u63ee\u3059\u308b\u3002<\/p>\n

\u7406\u7531\u3068\u3057\u3066\u306fLLM\u304c\u30b3\u30f3\u30c6\u30ad\u30b9\u30c8\u3084\u72b6\u614b\u306e\u7ba1\u7406\u304c\u3042\u308a\u3001\u4f1a\u8a71\u3092\u8e0f\u307e\u3048\u3001\u9069\u5207\u306a\u9806\u756a\u3067\u9069\u5207\u306a\u30c4\u30fc\u30eb\u3092\u4f7f\u3046\u3053\u3068\u304c\u96e3\u3057\u3044\u3068\u3002<\/p>\n

\u7d0d\u5f97\u3002<\/p>\n

1\u3064\u306e\u30ef\u30fc\u30af\u30d5\u30ed\u30fc\u3068\u3057\u3066\u5404\u51e6\u7406\u3092\u5185\u5305\u3057\u3001\u5236\u5fa1\u3055\u308c\u305f\u72b6\u614b\u3067\u30c4\u30fc\u30eb\u3068\u3057\u3066\u4f7f\u3046\u3002<\/p>\n

\u3053\u306e\u8a18\u4e8b\u304c\u306a\u304b\u3063\u305f\u3089\u3001Google Drive\u3001DB\u691c\u7d22\u3092\u5225\u3005\u3067MCP\u3067\u4f5c\u308b\u3053\u3068\u3092\u8003\u3048\u3066\u3044\u305f\u3068\u601d\u3044\u307e\u3059\u3002<\/p>\n

\u691c\u7d22\u3092\u4e00\u3064\u306e\u30ef\u30fc\u30af\u30d5\u30ed\u30fc\u3068\u3057\u3066\u305d\u306e\u4e2d\u3067\u9069\u5207\u306a\u3082\u306e\u3092\u9078\u3076\u3088\u3046\u5b9f\u88c5\u3057\u305f\u65b9\u304c\u826f\u3044\u3002<\/p>\n

MCP\u306e\u958b\u767a\u306e\u7c92\u5ea6\u306f\u610f\u8b58\u3057\u305f\u3044\u3068\u601d\u3044\u307e\u3057\u305f\u3002<\/p>\n

LLM\u304c\u4f55\u3092\u4f7f\u3046\u304b\u3092\u5224\u65ad\u3059\u308b\u305f\u3081\u306eMCP\u306etool\u306edescription\u3082\u91cd\u8981\u3067\u3059\u3002<\/p>\n

\u4e00\u3064\u306eAPI\u306e\u8aac\u660e\u3067\u306f\u306a\u304f\u3001\u30e6\u30fc\u30b6\u30fc\u306e\u8ab2\u984c\u3001\u76ee\u7684\u3092\u89e3\u6c7a\u3059\u308b\u30c4\u30fc\u30eb\u306edescription\u3001\u5207\u308a\u5206\u3051\u304c\u5fc5\u8981\u3060\u3068\u601d\u3044\u307e\u3057\u305f\u3002<\/p>\n

AI\u306e\u958b\u767a\u306f\u30d7\u30ed\u30f3\u30d7\u30c8\u3082\u91cd\u8981\u3067\u3042\u308a\u3001description\u3082\u91cd\u8981\u3002<\/p>\n

\u6700\u8fd1\u3001\u30d7\u30ed\u30f3\u30d7\u30c8\u3082AI\u3067\u4f5c\u308b\u3068\u3044\u3046\u6d41\u308c\u304c\u3042\u308a\u3001\u30d7\u30ed\u30f3\u30d7\u30c8\u3082AI\u304c\u3084\u3063\u3066\u304f\u308c\u308b\u3068\u306a\u3063\u3066\u3044\u307e\u3059\u304c\u3001\u3042\u308a\u304d\u305f\u308a\u306e\u6307\u793a\u3060\u3068\u3042\u308a\u304d\u305f\u308a\u306e\u30d7\u30ed\u30f3\u30d7\u30c8\u304c\u3067\u304d\u3042\u304c\u308a\u307e\u3059\u3002<\/p>\n

\u3061\u3083\u3093\u3068\u30e6\u30fc\u30b6\u30fc\u306e\u8ab2\u984c\u3001\u76ee\u7684\u3092\u8e0f\u307e\u3048\u305f\u4e0a\u3067\u30d7\u30ed\u30f3\u30d7\u30c8\u3001\u6307\u793a\u3092\u3059\u308b\u3053\u3068\u304c\u5927\u4e8b\u3002<\/p>\n

\u8ab2\u984c\u3092\u30d2\u30a2\u30ea\u30f3\u30b0\u3057\u3001\u3053\u306e\u8ab2\u984c\u3092\u89e3\u6c7a\u3059\u308b\u8a00\u8a9e\u5316\u3001\u601d\u8003\u304c\u307e\u3059\u307e\u3059\u91cd\u8981\u306b\u306a\u3063\u3066\u304d\u3066\u3044\u308b\u3068\u611f\u3058\u3066\u3044\u307e\u3059\u3002<\/p>\n

\u305d\u3082\u305d\u3082\u3001AI\u958b\u767a\u306f\u30d7\u30ed\u30b0\u30e9\u30df\u30f3\u30b0\u8a00\u8a9e\u3060\u3051\u3067\u306f\u306a\u304f\u3001\u30d7\u30ed\u30f3\u30d7\u30c8\u7b49\u306e\u81ea\u7136\u8a00\u8a9e\u306e\u5408\u308f\u305b\u6280\u306a\u306e\u3067\u3001\u5f53\u305f\u308a\u524d\u3068\u3044\u3048\u3070\u5f53\u305f\u308a\u524d\u306a\u306e\u3067\u3059\u304c\u3002\u3053\u306e2\u3064\u3092\u540c\u6642\u306b\u4f7f\u3063\u3066\u958b\u767a\u3092\u884c\u3046\u3068\u3044\u3046\u306e\u304c\u96e3\u3057\u3044\u3002\u30d7\u30ed\u30b0\u30e9\u30e0\u3092\u5909\u3048\u308b\u304b\uff1f\u30d7\u30ed\u30f3\u30d7\u30c8\u3092\u5909\u3048\u308b\u304b\uff1f\u3068\u305d\u3046\u3044\u3048\u3070\u3053\u306e\u5224\u65ad\u306f\u4eca\u307e\u3067\u306e\u958b\u767a\u3067\u306f\u306a\u3044\u306a\u3041\u3068\u3002\u3002\u3002<\/p>\n

AI\u3082\u4f7f\u3046\u4eba\u3067\u5dee\u304c\u51fa\u3066\u3044\u308b\u306e\u306f\u75db\u611f\u3057\u3001\u3053\u3053\u3082\u5927\u304d\u304f\u9055\u3046\u306e\u3067\u306f\u3068\u601d\u3044\u307e\u3057\u305f\u3002<\/p>\n

\u5f15\u7528\u8a18\u4e8b<\/p>\n

Revealing the Numeracy Gap: An Empirical Investigation of Text Embedding Models<\/a><\/p>\n

The second wave of MCP: Building for LLMs, not developers<\/a><\/p>\n

\n
<\/div><\/fb:like><\/div>\n
<\/g:plusone><\/div>\n
\"\u3053\u306e\u30a8\u30f3\u30c8\u30ea\u30fc\u3092\u306f\u3066\u306a\u30d6\u30c3\u30af\u30de\u30fc\u30af\u306b\u8ffd\u52a0\"<\/a>