{"id":2544,"date":"2025-07-04T11:12:19","date_gmt":"2025-07-04T03:12:19","guid":{"rendered":"https:\/\/sanlangcode.com\/?p=2544"},"modified":"2025-07-04T11:12:19","modified_gmt":"2025-07-04T03:12:19","slug":"%e7%94%a8%e6%bb%91%e5%8a%a8%e7%aa%97%e5%8f%a3%e6%89%b9%e9%87%8f%e6%8e%a8%e7%90%86%ef%bc%8c%e6%8b%af%e6%95%91%e4%bd%a0%e7%9a%84%e5%90%91%e9%87%8f%e6%a3%80%e7%b4%a2%e6%85%a2-%e4%b8%80","status":"publish","type":"post","link":"https:\/\/sanlangcode.com\/index.php\/2025\/07\/04\/%e7%94%a8%e6%bb%91%e5%8a%a8%e7%aa%97%e5%8f%a3%e6%89%b9%e9%87%8f%e6%8e%a8%e7%90%86%ef%bc%8c%e6%8b%af%e6%95%91%e4%bd%a0%e7%9a%84%e5%90%91%e9%87%8f%e6%a3%80%e7%b4%a2%e6%85%a2-%e4%b8%80\/","title":{"rendered":"\u7528\u6ed1\u52a8\u7a97\u53e3\u6279\u91cf\u63a8\u7406\uff0c\u62ef\u6551\u4f60\u7684\u5411\u91cf\u68c0\u7d22\u6162\u2014\u2014\u4e00\u573a\u201c\u5757\u201d\u4e0a\u52a0\u901f\u7684\u5b9e\u6218"},"content":{"rendered":"\n
\n

\u201c\u660e\u660e GPU \u5f88\u5f3a\uff0c\u4e3a\u4ec0\u4e48\u68c0\u7d22\u8fd8\u662f\u50cf\u8717\u725b\uff1f\u201d
\u2014\u2014\u4e00\u4f4d\u88ab embedding \u62d6\u6162\u7684\u5f00\u53d1\u8005<\/p>\n<\/blockquote>\n\n\n\n

\u524d\u8a00<\/h2>\n\n\n\n

\u5728\u5927\u6a21\u578b+\u5411\u91cf\u6570\u636e\u5e93\u7684\u68c0\u7d22\u573a\u666f\u4e0b\uff0c\u4e8b\u5b9e\u9a8c\u8bc1<\/strong>\u7ecf\u5e38\u9700\u8981\u201c\u5728\u4e00\u5806\u6587\u6863\u5757\u91cc\u627e\u51fa\u548c\u4e8b\u5b9e\u6700\u76f8\u5173\u7684\u90a3\u51e0\u5757\u201d\u3002
\u4e00\u5f00\u59cb\u4f60\u53ef\u80fd\u5f88\u5929\u771f\uff1a<\/p>\n\n\n\n

\n

\u201c\u76f4\u63a5 embedding \u68c0\u7d22+reranker\uff0c\u4e0d\u5c31\u5b8c\u4e8b\u4e86\u5417\uff1f\u201d<\/p>\n<\/blockquote>\n\n\n\n

\u5f88\u5feb\u4f60\u4f1a\u53d1\u73b0\uff0c\u51c6\u786e\u7387\u8fd8\u884c\uff0c\u4f46\u6709\u4e9b\u4e8b\u5b9e\u662f\u6587\u6863\u5757\u7684\u5b50\u53e5\uff0cembedding \u68c0\u7d22\u5bb9\u6613\u6f0f\u6389<\/strong>\u3002\u4e8e\u662f\u4f60\u796d\u51fa\u4e86\u201c\u6ed1\u52a8\u7a97\u53e3\u201d\u5927\u6cd5\uff1a<\/p>\n\n\n\n

    \n
  • \u628a\u6bcf\u4e2a chunk \u518d\u5207\u6210\u591a\u4e2a\u7a97\u53e3\uff0c\u548c fact \u505a embedding \u76f8\u4f3c\u5ea6\uff0c\u7a97\u53e3\u547d\u4e2d\u5c31\u7b97\u76f8\u5173\u3002<\/li>\n<\/ul>\n\n\n\n

    \u7ed3\u679c\u2014\u2014\u51c6\u786e\u7387\u63d0\u5347\u4e86\uff0c\u901f\u5ea6\u5374\u6162\u5230\u6000\u7591\u4eba\u751f<\/strong>\u3002<\/p>\n\n\n\n

    \n\n\n\n

    1. \u6734\u7d20\u6ed1\u52a8\u7a97\u53e3\uff1a\u4e00\u5200\u5207\uff0c\u6162\u5f97\u4f18\u96c5<\/h2>\n\n\n\n

    \u6700\u521d\u7684\u4ee3\u7801\u5927\u6982\u957f\u8fd9\u6837\uff1a<\/p>\n\n\n\n

    for chunk in all_chunks:\n    for i in range(0, len(chunk) - window_size + 1, step):\n        window = chunk[i:i+window_size]\n        window_emb = embedding_model.encode([window])\n        score = cosine_similarity(fact_emb, window_emb)\n        if score > threshold:\n            # \u547d\u4e2d<\/code><\/pre>\n\n\n\n
    \u95ee\u9898\uff1a<\/strong><\/p>\n\n\n\n
      \n
    • \u6bcf\u4e2a chunk \u90fd\u6ed1\u7a97\uff0c\u7a97\u53e3\u6570\u91cf\u7206\u70b8\u3002<\/li>\n\n\n\n
    • embedding \u63a8\u7406\u4e00\u7a97\u4e00\u63a8\uff0cGPU\/CPU \u8d44\u6e90\u6d6a\u8d39\u3002<\/li>\n\n\n\n
    • \u7ed3\u679c\uff1a\u51c6\u786e\u7387\u9ad8\uff0c\u901f\u5ea6\u611f\u4eba\u3002<\/li>\n<\/ul>\n\n\n\n
      \n\n\n\n
      2. \u4f18\u5316\u4e00\uff1a\u53ea\u5bf9 Top-N Chunk \u6ed1\u7a97<\/h2>\n\n\n\n
      \u601d\u8def\uff1a<\/strong>
      \u5148\u7528 embedding \u68c0\u7d22+reranker\uff0c\u6311\u51fa\u6700\u76f8\u5173\u7684\u524d N \u4e2a chunk\uff0c\u518d\u6ed1\u7a97\u3002<\/p>\n\n\n\n
      semantic_results = retriever.hybrid_search_db(query=fact, top_k_embedding=5)\ntop_chunks = [r[\"document\"] for r in semantic_results[:2]]  # \u53ea\u6ed1\u524d2\u4e2a<\/code><\/pre>\n\n\n\n
      \u6548\u679c\uff1a<\/strong><\/p>\n\n\n\n
        \n
      • \u53ea\u5bf9\u6700\u6709\u5e0c\u671b\u7684\u5757\u6ed1\u7a97\uff0c\u7a97\u53e3\u6570\u9aa4\u51cf\u3002<\/li>\n\n\n\n
      • \u901f\u5ea6\u63d0\u5347\uff0c\u51c6\u786e\u7387\u57fa\u672c\u4e0d\u53d8\u3002<\/li>\n<\/ul>\n\n\n\n
        \n\n\n\n
        3. \u4f18\u5316\u4e8c\uff1a\u6ed1\u7a97\u6279\u91cf\u63a8\u7406\uff0cGPU \u5403\u9971\u996d<\/h2>\n\n\n\n
        \u601d\u8def\uff1a<\/strong>
        \u628a\u6240\u6709\u7a97\u53e3\u6587\u672c\u6536\u96c6\u8d77\u6765\uff0c\u4e00\u6b21\u6027 batch \u63a8\u7406 embedding\u3002<\/p>\n\n\n\n
        all_windows = []\nwindow_info = []\nfor chunk in top_chunks:\n    for i in range(0, len(chunk) - window_size + 1, window_size):\n        window = chunk[i:i+window_size]\n        all_windows.append(window)\n        window_info.append((chunk, i, i+window_size))\n\nwindow_embs = embedding_model.encode(all_windows)  # \u4e00\u6b21\u6027\u63a8\u7406<\/code><\/pre>\n\n\n\n
        \u6548\u679c\uff1a<\/strong><\/p>\n\n\n\n
          \n
        • embedding \u63a8\u7406\u901f\u5ea6\u63d0\u5347\u6570\u500d\u3002<\/li>\n\n\n\n
        • \u8d44\u6e90\u5229\u7528\u7387\u9ad8\uff0c\u54cd\u5e94\u65f6\u95f4\u5927\u5e45\u7f29\u77ed\u3002<\/li>\n<\/ul>\n\n\n\n
          \n\n\n\n
          4. \u4f18\u5316\u4e09\uff1a\u53ea\u8f93\u51fa\u6bcf\u4e2a Chunk \u7684\u6700\u4f73\u7a97\u53e3<\/h2>\n\n\n\n
          \u601d\u8def\uff1a<\/strong>
          \u540c\u4e00\u4e2a chunk \u53ef\u80fd\u591a\u4e2a\u7a97\u53e3\u547d\u4e2d\uff0c\u4f46\u4f60\u53ea\u5173\u5fc3\u201c\u6700\u4f73\u547d\u4e2d\u201d\u5373\u53ef\u3002<\/p>\n\n\n\n
          chunk_best = {}\nfor idx, (chunk, start, end) in enumerate(window_info):\n    score = cosine_similarity(fact_emb, window_embs[idx])\n    if score >= threshold:\n        if chunk not in chunk_best or score > chunk_best[chunk][\"window_similarity\"]:\n            chunk_best[chunk] = {\n                \"chunk\": chunk,\n                \"window_similarity\": score,\n                \"window_start\": start,\n                \"window_end\": end,\n            }\n# \u53ea\u8f93\u51fa\u6700\u4f73\nfor best in chunk_best.values():\n    print(f\"[\u6ed1\u52a8\u7a97\u53e3\u6700\u4f73\u5339\u914d] fact: '{fact}' \u5339\u914d\u5230 chunk\uff0cwindow_similarity={best['window_similarity']:.4f}\")<\/code><\/pre>\n\n\n\n
          \u6548\u679c\uff1a<\/strong><\/p>\n\n\n\n
            \n
          • \u65e5\u5fd7\u7b80\u6d01\uff0c\u5206\u6790\u65b9\u4fbf\u3002<\/li>\n\n\n\n
          • \u7ed3\u679c\u66f4\u805a\u7126\uff0c\u907f\u514d\u91cd\u590d\u3002<\/li>\n<\/ul>\n\n\n\n
            \n\n\n\n
            5. \u7ec8\u6781\u5408\u5e76\uff1a\u76f8\u5173\u5757\u53bb\u91cd\uff0c\u5206\u6570\u4f18\u5148<\/h2>\n\n\n\n
            \u601d\u8def\uff1a<\/strong>
            \u65e0\u8bba\u662f semantic\u3001contains \u8fd8\u662f window_embedding\uff0c\u6bcf\u4e2a chunk \u53ea\u4fdd\u7559\u5206\u6570\u6700\u9ad8\u7684\u90a3\u79cd\u7c7b\u578b<\/strong>\uff0c\u6700\u7ec8\u53ea\u8fd4\u56de top_k \u4e2a\u3002<\/p>\n\n\n\n
            all_relevant = {}\nfor c in semantic_chunks + contains_chunks + window_chunks:\n    key = c[\"chunk\"]\n    new_score = c.get(\"reranker_score\", c.get(\"window_similarity\", -1))\n    old_score = all_relevant[key].get(\"reranker_score\", all_relevant[key].get(\"window_similarity\", -1)) if key in all_relevant else -1\n    if key not in all_relevant or new_score > old_score:\n        all_relevant[key] = c\nsorted_chunks = sorted(\n    all_relevant.values(),\n    key=lambda x: x.get(\"reranker_score\", x.get(\"window_similarity\", -1)),\n    reverse=True\n)[:top_k]<\/code><\/pre>\n\n\n\n
            \u6548\u679c\uff1a<\/strong><\/p>\n\n\n\n
              \n
            • \u7ed3\u679c\u66f4\u7cbe\u70bc\uff0c\u76f8\u5173\u6027\u66f4\u9ad8\u3002<\/li>\n\n\n\n
            • \u7528\u6237\u4f53\u9a8c up up up\uff01<\/li>\n<\/ul>\n\n\n\n
              \n\n\n\n
              6. \u603b\u7ed3<\/h2>\n\n\n\n
                \n
              • \u6ed1\u52a8\u7a97\u53e3<\/strong>\u8ba9\u4e8b\u5b9e\u9a8c\u8bc1\u66f4\u7cbe\u51c6\uff0c\u4f46\u8981\u7528\u5f97\u5de7\uff0c\u522b\u8ba9\u5b83\u62d6\u6162\u7cfb\u7edf\u3002<\/li>\n\n\n\n
              • \u6279\u91cf\u63a8\u7406+\u53ea\u6ed1 Top-N+\u53ea\u4fdd\u7559\u6700\u4f73\u7a97\u53e3+\u6700\u7ec8\u53bb\u91cd\u5408\u5e76<\/strong>\uff0c\u8ba9\u4f60\u7684\u68c0\u7d22\u53c8\u5feb\u53c8\u51c6\u3002<\/li>\n\n\n\n
              • \u4ee3\u7801\u4f18\u5316\u4e0d\u662f\u7384\u5b66\uff0c\u662f\u4e00\u573a\u201c\u5757\u201d\u4e0a\u52a0\u901f\u7684\u7406\u6027\u9769\u547d\uff01<\/li>\n<\/ul>\n\n\n\n
                \n\n\n\n
                \n
                \u201c\u4f18\u5316\u4e0d\u662f\u4e00\u8e74\u800c\u5c31\uff0c\u800c\u662f\u6bcf\u4e00\u884c\u4ee3\u7801\u7684\u7cbe\u96d5\u7ec6\u7422\u3002\u201d
                \u2014\u2014\u4f60\uff0c\u4f18\u5316\u5b8c\u6ed1\u52a8\u7a97\u53e3\u540e<\/p>\n<\/blockquote>\n\n\n\n
                \n\n\n\n
                \u6b22\u8fce\u70b9\u8d5e\u3001\u6536\u85cf\u3001\u8f6c\u53d1\uff0c\u8ba9\u66f4\u591a\u4eba\u5c11\u8e29 embedding \u68c0\u7d22\u7684\u5751\uff01<\/strong> \ud83d\ude80<\/p>\n\n\n\n
                \n\n\n\n
                \u5982\u9700\u5b8c\u6574\u4ee3\u7801\u6216\u6709\u7591\u95ee\uff0c\u8bc4\u8bba\u533a\u89c1\uff01<\/p>\n","protected":false},"excerpt":{"rendered":"
                \u201c\u660e\u660e GPU \u5f88\u5f3a\uff0c\u4e3a\u4ec0\u4e48\u68c0\u7d22...<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[4],"tags":[],"class_list":["post-2544","post","type-post","status-publish","format-standard","hentry","category-4"],"_links":{"self":[{"href":"https:\/\/sanlangcode.com\/index.php\/wp-json\/wp\/v2\/posts\/2544","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/sanlangcode.com\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/sanlangcode.com\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/sanlangcode.com\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/sanlangcode.com\/index.php\/wp-json\/wp\/v2\/comments?post=2544"}],"version-history":[{"count":1,"href":"https:\/\/sanlangcode.com\/index.php\/wp-json\/wp\/v2\/posts\/2544\/revisions"}],"predecessor-version":[{"id":2545,"href":"https:\/\/sanlangcode.com\/index.php\/wp-json\/wp\/v2\/posts\/2544\/revisions\/2545"}],"wp:attachment":[{"href":"https:\/\/sanlangcode.com\/index.php\/wp-json\/wp\/v2\/media?parent=2544"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/sanlangcode.com\/index.php\/wp-json\/wp\/v2\/categories?post=2544"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/sanlangcode.com\/index.php\/wp-json\/wp\/v2\/tags?post=2544"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}