{"id":10021418,"date":"2026-07-27T23:53:34","date_gmt":"2026-07-27T15:53:34","guid":{"rendered":"https:\/\/itmax.cn\/?p=10021418"},"modified":"2026-07-27T23:53:34","modified_gmt":"2026-07-27T15:53:34","slug":"%e6%b7%b1%e5%85%a5%e6%b5%85%e5%87%ba-rag%ef%bc%9a%e7%94%a8%e4%b8%80%e4%b8%aa%e5%8f%af%e8%bf%90%e8%a1%8c%e7%9a%84-demo-%e8%ae%b2%e9%80%8f%e5%ae%8c%e6%95%b4%e9%93%be%e8%b7%af","status":"publish","type":"post","link":"https:\/\/itmax.cn\/?p=10021418","title":{"rendered":"\u6df1\u5165\u6d45\u51fa RAG\uff1a\u7528\u4e00\u4e2a\u53ef\u8fd0\u884c\u7684 Demo \u8bb2\u900f\u5b8c\u6574\u94fe\u8def"},"content":{"rendered":"<h2>RAG \u662f\u4ec0\u4e48<\/h2>\n<p>RAG\uff08Retrieval-Augmented Generation\uff0c\u68c0\u7d22\u589e\u5f3a\u751f\u6210\uff09\u662f\u4e00\u79cd\u5148\u68c0\u7d22\u8d44\u6599\u3001\u518d\u751f\u6210\u7b54\u6848\u7684\u95ee\u7b54\u65b9\u6cd5\u3002\u7528\u6237\u63d0\u51fa\u95ee\u9898\u540e\uff0c\u7cfb\u7edf\u5148\u4ece\u5916\u90e8\u77e5\u8bc6\u5e93\u4e2d\u627e\u51fa\u76f8\u5173\u5185\u5bb9\uff0c\u518d\u628a\u95ee\u9898\u548c\u68c0\u7d22\u7ed3\u679c\u4e00\u8d77\u4ea4\u7ed9 LLM \u56de\u7b54\u3002<\/p>\n<p>\u68c0\u7d22\uff08Retrieval\uff09\u8d1f\u8d23\u4ece\u77e5\u8bc6\u5e93\u4e2d\u627e\u5230\u76f8\u5173\u8d44\u6599\uff0c\u751f\u6210\uff08Generation\uff09\u8d1f\u8d23\u6839\u636e\u8fd9\u4e9b\u8d44\u6599\u751f\u6210\u7b54\u6848\u3002\u5982\u679c\u7cfb\u7edf\u53ea\u5217\u51fa\u76f8\u4f3c\u6587\u6863\uff0c\u5b8c\u6210\u7684\u53ea\u662f\u68c0\u7d22\uff1b\u628a\u68c0\u7d22\u7ed3\u679c\u7ee7\u7eed\u4ea4\u7ed9 LLM \u751f\u6210\u56de\u7b54\uff0c\u624d\u6784\u6210 RAG\u3002<\/p>\n<p>\u68c0\u7d22\u5668\u4f1a\u6309\u76f8\u5173\u6027\u7ed9\u5019\u9009\u7247\u6bb5\u6392\u5e8f\uff0c\u901a\u5e38\u53ea\u628a\u6392\u540d\u6700\u9760\u524d\u7684 K \u6761\u4ea4\u7ed9 LLM\u3002\u8fd9\u7ec4\u7ed3\u679c\u79f0\u4e3a Top-K\uff0c\u5176\u4e2d K \u8868\u793a\u4fdd\u7559\u591a\u5c11\u4e2a\u6587\u672c\u7247\u6bb5\u3002<\/p>\n<pre style=\"display:none;\"><code class=\"mermaid\">flowchart LR\n    Q[\u7528\u6237\u95ee\u9898] --&gt; R[\u68c0\u7d22 Retrieval]\n    R --&gt; K[Top-K\uff1a\u76f8\u5173\u6027\u6700\u9ad8\u7684 K \u4e2a\u6587\u672c\u7247\u6bb5]\n    K --&gt; G[LLM \u751f\u6210 Generation]\n    G --&gt; A[\u5e26\u5f15\u7528\u7684\u56de\u7b54]<\/code><\/pre>\n<p>\u672c\u6587\u914d\u5408\u4e00\u4e2a\u53ef\u4ee5\u76f4\u63a5\u8fd0\u884c\u7684 Demo \u6765\u8bb2\u89e3\u8fd9\u5957\u6d41\u7a0b\u3002\u670d\u52a1\u7aef\u5148\u5728\u672c\u5730\u628a\u8d44\u6599\u8f6c\u6362\u6210\u53ef\u4f9b\u6bd4\u8f83\u7684\u5411\u91cf\uff0c\u518d\u7ed3\u5408\u8bed\u4e49\u548c\u5173\u952e\u8bcd\u4e24\u79cd\u65b9\u5f0f\u67e5\u627e\u76f8\u5173\u5185\u5bb9\uff0c\u6700\u540e\u8c03\u7528 <code>.env<\/code> \u6587\u4ef6\u4e2d\u914d\u7f6e\u7684 LLM \u751f\u6210\u5e26\u6765\u6e90\u7684\u56de\u7b54\u3002\u6587\u4e2d\u7684\u516c\u53f8\u3001\u5236\u5ea6\u3001\u91d1\u989d\u548c\u9519\u8bef\u7801\u5747\u4e3a mock \u6570\u636e\uff0c\u7528\u6237\u4e5f\u53ef\u4ee5\u4e0a\u4f20\u81ea\u5df1\u7684\u6587\u6863\uff0c\u9a8c\u8bc1\u79c1\u6709\u8d44\u6599\u7684\u68c0\u7d22\u6548\u679c\u3002<\/p>\n<h2>\u4e3a\u4ec0\u4e48\u9700\u8981 RAG<\/h2>\n<p>LLM \u53ef\u4ee5\u7ec4\u7ec7\u51fa\u6d41\u7545\u7684\u6587\u5b57\uff0c\u4f46\u5982\u679c\u5e94\u7528\u6ca1\u6709\u4e3a\u5b83\u63a5\u5165\u5916\u90e8\u8d44\u6599\u6216\u641c\u7d22\u5de5\u5177\uff0c\u5b83\u53ea\u80fd\u4f9d\u636e\u8bad\u7ec3\u6570\u636e\u548c\u5f53\u524d\u4e0a\u4e0b\u6587\u56de\u7b54\u95ee\u9898\u3002\u516c\u53f8\u7684\u5185\u90e8\u5236\u5ea6\u3001\u56e2\u961f\u6587\u6863\u548c\u4e2a\u4eba\u7b14\u8bb0\u8fd9\u4e9b\u8d44\u6599\u901a\u5e38\u4e0d\u4f1a\u51fa\u73b0\u5728\u516c\u5f00\u8bad\u7ec3\u6570\u636e\u4e2d\u3002<\/p>\n<p>\u4f8b\u5982\u7528\u6237\u95ee\uff1a<\/p>\n<blockquote><p>\u6253\u8f66\u8d39\u8d85\u8fc7\u516b\u767e\u5757\u5e94\u8be5\u600e\u4e48\u529e\uff1f<\/p><\/blockquote>\n<p>\u5982\u679c\u53ea\u628a\u8fd9\u53e5\u8bdd\u53d1\u7ed9 LLM\uff0c\u6a21\u578b\u6700\u591a\u6839\u636e\u5e38\u8bc6\u731c\u6d4b\u201c\u53ef\u80fd\u9700\u8981\u9886\u5bfc\u5ba1\u6279\u201d\u3002\u7531\u4e8e\u5f53\u524d\u4e0a\u4e0b\u6587\u91cc\u6ca1\u6709\u516c\u53f8\u7684\u62a5\u9500\u989d\u5ea6\u3001\u5ba1\u6279\u4eba\u548c\u586b\u5355\u8981\u6c42\uff0c\u8fd9\u4e2a\u56de\u7b54\u5373\u4f7f\u78b0\u5de7\u6b63\u786e\uff0c\u4e5f\u5f88\u96be\u8bc1\u660e\u4f9d\u636e\u6765\u81ea\u54ea\u4efd\u5236\u5ea6\u3002<\/p>\n<p>\u63a5\u5165 RAG \u540e\uff0c\u7cfb\u7edf\u4e0d\u4f1a\u76f4\u63a5\u751f\u6210\u7b54\u6848\uff0c\u800c\u662f\u5148\u4ece\u77e5\u8bc6\u5e93\u4e2d\u627e\u5230\u76f8\u5173\u8d44\u6599\u3002\u4f8b\u5982\uff0c\u8fd9\u4e2a\u95ee\u9898\u5e94\u8be5\u547d\u4e2d\u4e0b\u9762\u7684\u5236\u5ea6\u6761\u6b3e\uff1a<\/p>\n<pre><code class=\"text\">\u5458\u5de5\u56e0\u516c\u4ea7\u751f\u7684\u5e02\u5185\u4ea4\u901a\u8d39\u7528\uff0c\u6bcf\u6708\u62a5\u9500\u9650\u989d\u4e3a 800 \u5143\u3002\n\u8d85\u51fa\u9650\u989d\u7684\u90e8\u5206\u9700\u8981\u90e8\u95e8\u8d1f\u8d23\u4eba\u5ba1\u6279\uff0c\u5e76\u5728\u62a5\u9500\u5355\u4e2d\u8bf4\u660e\u4e1a\u52a1\u539f\u56e0\u3002<\/code><\/pre>\n<p>\u68c0\u7d22\u5b8c\u6210\u540e\uff0cLLM \u518d\u6839\u636e\u8fd9\u6bb5\u5236\u5ea6\u6761\u6b3e\u56de\u7b54\u3002\u79c1\u6709\u8d44\u6599\u4e0d\u9700\u8981\u5199\u5165\u6a21\u578b\u53c2\u6570\uff0c\u6587\u6863\u53d1\u751f\u53d8\u5316\u65f6\uff0c\u91cd\u65b0\u5904\u7406\u53d1\u751f\u53d8\u5316\u7684\u5185\u5bb9\u5e76\u66f4\u65b0\u7d22\u5f15\u5373\u53ef\u3002\u68c0\u7d22\u7ed3\u679c\u8fd8\u4fdd\u7559\u4e86\u6765\u6e90\uff0c\u56e0\u6b64\u7b54\u6848\u53ef\u4ee5\u9644\u4e0a\u5f15\u7528\uff0c\u7528\u6237\u80fd\u591f\u56de\u5230\u539f\u6587\u6838\u5bf9\u3002<\/p>\n<p>\u4e0d\u8fc7\uff0cRAG \u4e0d\u80fd\u201c\u5f7b\u5e95\u6d88\u706d\u5e7b\u89c9\u201d\u3002\u5982\u679c\u6b63\u786e\u7684\u8d44\u6599\u6ca1\u6709\u8fdb\u5165 Top-K\uff0c\u6216\u8005\u6587\u6863\u5df2\u7ecf\u8fc7\u671f\u3001\u5206\u5757\u5207\u65ad\u4e86\u5173\u952e\u6761\u4ef6\uff0c\u7cfb\u7edf\u4ecd\u7136\u4f1a\u7ed9\u51fa\u9519\u8bef\u7b54\u6848\u3002\u56e0\u6b64\uff0c\u68c0\u7d22\u7ed3\u679c\u548c\u6700\u7ec8\u56de\u7b54\u90fd\u9700\u8981\u5355\u72ec\u8bc4\u4f30\uff0c\u4e0d\u80fd\u56e0\u4e3a\u7b54\u6848\u5e26\u6709\u5f15\u7528\u5c31\u8ba4\u4e3a\u5b83\u4e00\u5b9a\u6b63\u786e\u3002<\/p>\n<h2>RAG \u7684\u5de5\u4f5c\u539f\u7406<\/h2>\n<p>\u6587\u7ae0\u5f00\u5934\u7684\u201c\u68c0\u7d22\u201d\u548c\u201c\u751f\u6210\u201d\u63cf\u8ff0\u7684\u662f\u4e00\u6b21\u95ee\u7b54\uff1a\u7cfb\u7edf\u5148\u627e\u8d44\u6599\uff0c\u518d\u751f\u6210\u7b54\u6848\u3002\u4f46\u5728\u7528\u6237\u63d0\u95ee\u4e4b\u524d\uff0c\u7cfb\u7edf\u8fd8\u8981\u9884\u5148\u5904\u7406\u77e5\u8bc6\u5e93\u3002\u56e0\u6b64\uff0c\u4ece\u7cfb\u7edf\u8fd0\u884c\u987a\u5e8f\u6765\u770b\uff0cRAG \u5305\u542b\u5efa\u7acb\u7d22\u5f15\uff08Indexing\uff09\u548c\u5904\u7406\u67e5\u8be2\uff08Querying\uff09\u4e24\u4e2a\u9636\u6bb5\uff1b\u68c0\u7d22\u4e0e\u751f\u6210\u90fd\u53d1\u751f\u5728\u67e5\u8be2\u9636\u6bb5\u3002<\/p>\n<p>\u5efa\u7acb\u7d22\u5f15\u65f6\uff0c\u7cfb\u7edf\u8bfb\u53d6\u6587\u6863\uff0c\u6309\u6807\u9898\u3001\u6bb5\u843d\u548c\u957f\u5ea6\u5207\u6210\u591a\u4e2a chunk\uff08\u6587\u672c\u7247\u6bb5\uff09\uff0c\u518d\u7528 Embedding \u6a21\u578b\u4e3a\u6bcf\u4e2a\u7247\u6bb5\u751f\u6210\u5411\u91cf\u3002\u7cfb\u7edf\u4f1a\u540c\u65f6\u4fdd\u5b58\u539f\u6587\u3001\u5411\u91cf\u3001\u6765\u6e90\u3001\u6807\u9898\u548c\u539f\u6587\u4f4d\u7f6e\uff0c\u4e4b\u540e\u4e0d\u5fc5\u628a\u6574\u4efd\u6587\u6863\u9010\u5b57\u4ea4\u7ed9\u641c\u7d22\u6a21\u5757\u3002<\/p>\n<p>\u5904\u7406\u67e5\u8be2\u65f6\uff0cDemo \u5206\u522b\u6267\u884c\u5411\u91cf\u68c0\u7d22\u548c BM25 \u5173\u952e\u8bcd\u68c0\u7d22\uff1a\u524d\u8005\u6309\u8bed\u4e49\u76f8\u4f3c\u5ea6\u6392\u5e8f\uff0c\u540e\u8005\u6309\u8bcd\u9879\u5339\u914d\u7a0b\u5ea6\u6392\u5e8f\u3002\u4e24\u6761\u5206\u652f\u72ec\u7acb\u641c\u7d22\u540c\u4e00\u6279 chunk\uff0c\u5f97\u5230\u7684\u987a\u5e8f\u53ef\u80fd\u4e0d\u540c\u3002\u7a0b\u5e8f\u628a\u4e24\u8def\u5206\u6570\u6362\u7b97\u5230\u53ef\u5408\u5e76\u7684\u8303\u56f4\uff0c\u8ba1\u7b97\u65b0\u7684\u603b\u5206\uff0c\u518d\u628a\u6700\u7ec8 Top-K \u8fde\u540c\u6765\u6e90\u7f16\u53f7\u4ea4\u7ed9 LLM\u3002<\/p>\n<p>\u540e\u9762\u7684\u7ae0\u8282\u4f1a\u4f9d\u6b21\u62c6\u5f00\u8fd9\u6761\u94fe\u8def\uff1a\u5148\u770b\u6587\u6863\u600e\u6837\u5206\u5757\u3001Embedding \u600e\u6837\u751f\u6210\uff0c\u518d\u770b\u5411\u91cf\u4e0e BM25 \u5982\u4f55\u6392\u5e8f\u548c\u878d\u5408\u3002<\/p>\n<p>\u8fd9\u4e2a Demo \u7684 Embedding \u6a21\u578b\u8f93\u51fa 384 \u7ef4\u5411\u91cf\u3002\u5411\u91cf\u4e2d\u7684\u6bcf\u4e2a\u4f4d\u7f6e\u79f0\u4e3a\u4e00\u7ef4\uff1a\u7b2c\u4e00\u4e2a\u4f4d\u7f6e\u662f\u7b2c 1 \u7ef4\uff0c\u7b2c\u4e8c\u4e2a\u4f4d\u7f6e\u662f\u7b2c 2 \u7ef4\uff0c\u4f9d\u6b64\u7c7b\u63a8\u3002\u8ba1\u7b97\u6587\u672c\u76f8\u4f3c\u5ea6\u65f6\uff0c\u7a0b\u5e8f\u4f1a\u4f7f\u7528\u5168\u90e8 384 \u7ef4\uff1b\u5355\u72ec\u4e00\u7ef4\u6ca1\u6709\u201c\u4ea4\u901a\u8d39\u201d\u6216\u201c\u62a5\u9500\u201d\u8fd9\u6837\u7684\u56fa\u5b9a\u542b\u4e49\u3002<\/p>\n<p>\u7d22\u5f15\u6587\u4ef6\u4e2d\u7684\u4e00\u6761\u8bb0\u5f55\u5982\u4e0b\u3002\u5b8c\u6574\u7684 <code>vector<\/code> \u5305\u542b 384 \u7ef4\u30020.024451 \u662f\u7b2c 1 \u7ef4\u7684\u503c\uff0c-0.015882 \u662f\u7b2c 2 \u7ef4\u7684\u503c\u3002\u6587\u7ae0\u7701\u7565\u4e86\u5176\u4f59 378 \u7ef4\uff0c\u5b9e\u9645\u7d22\u5f15\u4f1a\u4fdd\u5b58\u5b8c\u6574\u5411\u91cf\u3002<\/p>\n<pre><code class=\"jsonc\">{\n  &quot;id&quot;: &quot;\u661f\u6cb3\u79d1\u6280\u5dee\u65c5\u4e0e\u4ea4\u901a\u62a5\u9500\u5236\u5ea6-27-coxv8b&quot;,\n  &quot;content&quot;: &quot;\u5458\u5de5\u56e0\u516c\u4ea7\u751f\u7684\u5e02\u5185\u4ea4\u901a\u8d39\u7528\uff0c\u6bcf\u6708\u62a5\u9500\u9650\u989d\u4e3a 800 \u5143\u3002\u8d85\u51fa\u9650\u989d\u7684\u90e8\u5206\u9700\u8981\u90e8\u95e8\u8d1f\u8d23\u4eba\u5ba1\u6279\uff0c\u5e76\u5728\u62a5\u9500\u5355\u4e2d\u8bf4\u660e\u4e1a\u52a1\u539f\u56e0\u3002&quot;,\n  &quot;source&quot;: &quot;\u661f\u6cb3\u79d1\u6280\u5dee\u65c5\u4e0e\u4ea4\u901a\u62a5\u9500\u5236\u5ea6&quot;,\n  &quot;heading&quot;: &quot;\u5e02\u5185\u4ea4\u901a\u8d39\u7528&quot;,\n  &quot;startOffset&quot;: 27,\n  &quot;vector&quot;: [0.024451, -0.015882, -0.075545, -0.058644, 0.048382, 0.016459 \/* \u5176\u4f59 378 \u7ef4\u7701\u7565 *\/]\n}<\/code><\/pre>\n<p>\u8fd9\u6761\u8bb0\u5f55\u4e2d\u7684 <code>content<\/code> \u4fdd\u5b58\u539f\u59cb\u6587\u672c\uff0c\u4e4b\u540e\u4f1a\u4f5c\u4e3a\u4e0a\u4e0b\u6587\u4ea4\u7ed9 LLM\u3002\u6765\u6e90\u548c\u6807\u9898\u7528\u4e8e\u751f\u6210\u5f15\u7528\uff0c\u8d77\u59cb\u4f4d\u7f6e\u7528\u4e8e\u5b9a\u4f4d\u539f\u6587\u3002<code>vector<\/code> \u5219\u53ea\u53c2\u4e0e\u76f8\u4f3c\u5ea6\u8ba1\u7b97\uff0c\u4e0d\u4f1a\u76f4\u63a5\u5c55\u793a\u7ed9\u6700\u7ec8\u7528\u6237\u3002<\/p>\n<pre style=\"display:none;\"><code class=\"mermaid\">flowchart TB\n    subgraph Indexing[\u7d22\u5f15\u9636\u6bb5]\n        D[\u5185\u7f6e\u8d44\u6599\u6216\u7528\u6237\u6587\u4ef6] --&gt; P[\u63d0\u53d6\u7eaf\u6587\u672c]\n        P --&gt; C[\u6309\u6807\u9898\u3001\u6bb5\u843d\u548c\u957f\u5ea6\u5206\u5757]\n        C --&gt; E1[\u672c\u5730 Embedding \u6a21\u578b]\n        E1 --&gt; S[(\u6587\u672c\u3001\u5411\u91cf\u4e0e\u5143\u6570\u636e)]\n        C --&gt; S\n    end\n\n    subgraph Querying[\u67e5\u8be2\u9636\u6bb5]\n        Q[\u7528\u6237\u95ee\u9898] --&gt; E2[\u95ee\u9898\u5411\u91cf]\n        Q --&gt; B2[\u63d0\u53d6\u67e5\u8be2\u8bcd]\n        E2 --&gt; V[\u6309\u5411\u91cf\u76f8\u4f3c\u7a0b\u5ea6\u6392\u5e8f]\n        B2 --&gt; B[\u626b\u63cf\u539f\u6587\u5e76\u8ba1\u7b97 BM25]\n        V --&gt; M[\u628a\u4e24\u8def\u5206\u6570\u6362\u7b97\u540e\u5408\u5e76]\n        B --&gt; M\n        M --&gt; K[Top-K]\n        K --&gt; L[LLM]\n        L --&gt; A[\u7b54\u6848\u4e0e\u6765\u6e90]\n    end\n\n    S --&gt; V\n    S --&gt; B<\/code><\/pre>\n<p>\u7d22\u5f15\u3001\u68c0\u7d22\u548c\u751f\u6210\u89e3\u51b3\u7684\u95ee\u9898\u4e0d\u540c\uff1a\u7d22\u5f15\u628a\u8d44\u6599\u52a0\u5de5\u6210\u53ef\u641c\u7d22\u7684\u6570\u636e\uff0c\u68c0\u7d22\u5bfb\u627e\u4e0e\u5f53\u524d\u95ee\u9898\u76f8\u5173\u7684\u7247\u6bb5\uff0c\u751f\u6210\u518d\u628a\u8fd9\u4e9b\u7247\u6bb5\u7ec4\u7ec7\u6210\u81ea\u7136\u8bed\u8a00\u7b54\u6848\u3002\u8fd9\u4e09\u4e2a\u73af\u8282\u7684\u6545\u969c\u4e5f\u4e0d\u540c\u2014\u2014\u5206\u5757\u53ef\u80fd\u5207\u65ad\u5173\u952e\u6761\u4ef6\uff0c\u6392\u5e8f\u53ef\u80fd\u6f0f\u6389\u6b63\u786e\u7247\u6bb5\uff0cLLM \u5219\u53ef\u80fd\u8bef\u89e3\u8d44\u6599\u6216\u52a0\u5165\u8d44\u6599\u4e2d\u6ca1\u6709\u7684\u5185\u5bb9\u3002Demo \u56e0\u6b64\u4fdd\u7559\u4e86\u201c\u53ea\u770b\u6df7\u5408\u641c\u7d22\u201d\u89c6\u56fe\uff0c\u8ba9\u8bfb\u8005\u5148\u68c0\u67e5\u68c0\u7d22\u7ed3\u679c\uff0c\u518d\u68c0\u67e5\u751f\u6210\u5185\u5bb9\u3002<\/p>\n<h2>\u5206\u5757\u7b56\u7565\uff1a\u6807\u9898\u3001\u957f\u5ea6\u548c\u91cd\u53e0\u5185\u5bb9<\/h2>\n<p>\u5982\u679c\u628a\u4e00\u6574\u672c\u624b\u518c\u538b\u6210\u4e00\u4e2a\u5411\u91cf\uff0c\u91cc\u9762\u4e0d\u540c\u7ae0\u8282\u7684\u8bed\u4e49\u4f1a\u4e92\u76f8\u7a00\u91ca\u3002\u68c0\u7d22\u5373\u4f7f\u547d\u4e2d\u4e86\u6574\u672c\u624b\u518c\uff0c\u4e5f\u65e0\u6cd5\u7cbe\u786e\u5b9a\u4f4d\u7b54\u6848\u6240\u5728\u7684\u6bb5\u843d\u3002\u628a\u5168\u6587\u653e\u8fdb\u4e0a\u4e0b\u6587\u8fd8\u4f1a\u5360\u7528\u66f4\u591a token\u3002<\/p>\n<h3>\u5e38\u89c1\u505a\u6cd5<\/h3>\n<p>\u56e0\u6b64\uff0c\u5efa\u7acb\u7d22\u5f15\u524d\u8981\u5148\u505a\u6587\u672c\u5206\u5757\uff08Chunking\uff09\u3002\u5728\u5b9e\u9645\u9879\u76ee\u4e2d\uff0c\u6ca1\u6709\u9002\u5408\u6240\u6709\u6587\u6863\u7684\u56fa\u5b9a\u5207\u6cd5\uff0c\u901a\u5e38\u4f1a\u6309\u4e0b\u9762\u7684\u987a\u5e8f\u5904\u7406\uff1a<\/p>\n<ol>\n<li>\u5148\u5229\u7528\u6587\u6863\u7ed3\u6784\u3002Markdown \u548c HTML \u6309\u6807\u9898\u3001\u6bb5\u843d\u3001\u5217\u8868\u7b49\u8fb9\u754c\u5207\u5206\uff1bPDF\u3001DOCX \u548c\u8868\u683c\u5219\u5c3d\u91cf\u4fdd\u7559\u7ae0\u8282\u3001\u9875\u7801\u3001\u8868\u5934\u4e0e\u5355\u5143\u683c\u5173\u7cfb\u3002<\/li>\n<li>\u518d\u9650\u5236 chunk \u957f\u5ea6\u3002\u67d0\u4e2a\u7ae0\u8282\u4ecd\u7136\u8fc7\u957f\u65f6\uff0c\u6309\u6bb5\u843d\u3001\u53e5\u5b50\u3001\u8bcd\u6216 token \u9010\u7ea7\u5207\u5206\uff0c\u5c3d\u91cf\u4e0d\u4ece\u4e00\u53e5\u8bdd\u4e2d\u95f4\u622a\u65ad\u3002<\/li>\n<li>\u89c6\u60c5\u51b5\u4fdd\u7559\u91cd\u53e0\u5185\u5bb9\uff08overlap\uff09\uff0c\u4e5f\u5c31\u662f\u8ba9\u540e\u4e00\u5757\u5f00\u5934\u91cd\u590d\u524d\u4e00\u5757\u7ed3\u5c3e\u7684\u4e00\u5c0f\u6bb5\u6587\u5b57\u3002\u8fd9\u6837\u4e0d\u5bb9\u6613\u628a\u8de8\u8d8a\u5207\u5206\u4f4d\u7f6e\u7684\u4e00\u53e5\u8bdd\u6216\u4e00\u4e2a\u6761\u4ef6\u62c6\u6563\uff0c\u4f46\u91cd\u590d\u8fc7\u591a\u4f1a\u589e\u5927\u7d22\u5f15\uff0c\u4e5f\u53ef\u80fd\u8ba9\u641c\u7d22\u7ed3\u679c\u51fa\u73b0\u5185\u5bb9\u76f8\u4f3c\u7684\u7247\u6bb5\u3002<\/li>\n<li>\u4fdd\u5b58\u6587\u6863 ID\u3001\u6807\u9898\u5c42\u7ea7\u3001\u9875\u7801\u3001\u6743\u9650\u6807\u7b7e\u7b49\u5143\u6570\u636e\uff0c\u4ee5\u4f9b\u8fc7\u6ee4\u3001\u5f15\u7528\u548c\u56de\u6eaf\u4f7f\u7528\u3002<\/li>\n<\/ol>\n<p>\u524d\u4e24\u6b65\u662f\u6700\u5e38\u89c1\u7684\u57fa\u7840\u5207\u6cd5\uff1a\u5148\u6309\u7167\u6587\u6863\u539f\u6709\u7684\u6807\u9898\u548c\u6bb5\u843d\u5212\u5206\u5185\u5bb9\uff0c\u518d\u628a\u4ecd\u7136\u8fc7\u957f\u7684\u90e8\u5206\u7ee7\u7eed\u5207\u5c0f\u3002\u9047\u5230\u7ed3\u6784\u66f4\u590d\u6742\u7684\u8d44\u6599\uff0c\u8fd8\u53ef\u4ee5\u4f7f\u7528\u4e0b\u9762\u8fd9\u4e9b\u65b9\u6cd5\uff1a<\/p>\n<ul>\n<li>\u7248\u9762\u611f\u77e5\u5207\u5206\u4f1a\u8bfb\u53d6\u6587\u5b57\u5728\u9875\u9762\u4e0a\u7684\u4f4d\u7f6e\uff0c\u907f\u514d\u628a\u5de6\u53f3\u5206\u680f\u7684\u5185\u5bb9\u6216\u8868\u683c\u4e2d\u4e0d\u540c\u5355\u5143\u683c\u7684\u6587\u5b57\u9519\u8bef\u5730\u62fc\u5728\u4e00\u8d77\u3002<\/li>\n<li>\u8bed\u4e49\u5206\u5757\u4f1a\u5224\u65ad\u76f8\u90bb\u53e5\u5b50\u662f\u5426\u4ecd\u5728\u8ba8\u8bba\u540c\u4e00\u4e3b\u9898\uff0c\u8bdd\u9898\u660e\u663e\u53d8\u5316\u65f6\u624d\u5207\u5f00\uff0c\u800c\u4e0d\u662f\u53ea\u6309\u56fa\u5b9a\u957f\u5ea6\u622a\u65ad\u3002<\/li>\n<li>\u7236\u5b50\u5757\u68c0\u7d22\u4f1a\u7528\u8f83\u77ed\u7684\u7247\u6bb5\u8fdb\u884c\u7cbe\u786e\u641c\u7d22\uff0c\u547d\u4e2d\u540e\u518d\u53d6\u56de\u5b83\u6240\u5c5e\u7684\u8f83\u957f\u7ae0\u8282\uff0c\u4f5c\u4e3a\u66f4\u5b8c\u6574\u7684\u751f\u6210\u4e0a\u4e0b\u6587\u3002<\/li>\n<\/ul>\n<p>\u8fd9\u4e9b\u65b9\u6cd5\u9700\u8981\u989d\u5916\u7684\u6587\u6863\u89e3\u6790\u6216\u68c0\u7d22\u6b65\u9aa4\uff0c\u662f\u5426\u4f7f\u7528\u8981\u770b\u8d44\u6599\u7684\u7ed3\u6784\u548c\u5b9e\u9645\u6d4b\u8bd5\u7ed3\u679c\u3002<\/p>\n<h3>Demo \u7684\u5b9e\u73b0<\/h3>\n<p>\u8fd9\u4e2a Demo \u4f7f\u7528\u4e86\u7b80\u5316\u7684\u5207\u5206\u65b9\u6cd5\uff1aMarkdown \u5148\u6309\u4e00\u5230\u4e09\u7ea7\u6807\u9898\u5206\u6bb5\uff0c\u8fc7\u957f\u7684\u6bb5\u843d\u518d\u6309\u957f\u5ea6\u5207\u5f00\uff0c\u5e76\u5c3d\u91cf\u5728\u7a7a\u884c\u6216\u6362\u884c\u5904\u7ed3\u675f\u3002\u4ee3\u7801\u4e2d\u7684 <code>chunkSize = 400<\/code> \u8868\u793a\u6bcf\u4e2a\u7247\u6bb5\u7684\u76ee\u6807\u4e0a\u9650\u7ea6\u4e3a 400 \u4e2a token\uff1b<code>chunkOverlap = 80<\/code> \u8868\u793a\u5207\u51fa\u4e0b\u4e00\u7247\u65f6\uff0c\u4f1a\u91cd\u590d\u4e0a\u4e00\u7247\u7ed3\u5c3e\u7ea6 80 \u4e2a token\u3002\u6bd4\u5982\u7b2c\u4e00\u7247\u63a5\u8fd1 400 \u4e2a token\uff0c\u7b2c\u4e8c\u7247\u4f1a\u4ece\u5927\u7ea6\u7b2c 321 \u4e2a token \u9644\u8fd1\u5f00\u59cb\uff0c\u56e0\u6b64\u4e24\u7247\u4e4b\u95f4\u4fdd\u7559\u7ea6 80 \u4e2a token \u7684\u76f8\u540c\u5185\u5bb9\u3002<\/p>\n<p>\u8fd9\u91cc\u7684 token \u6570\u662f\u4f30\u7b97\u503c\u3002\u4ee3\u7801\u7528\u201c\u6587\u672c\u5b57\u7b26\u6570\u9664\u4ee5 3.5\u201d\u505a\u8fd1\u4f3c\u8ba1\u7b97\uff0c\u6ca1\u6709\u8c03\u7528\u6a21\u578b\u5b9e\u9645\u4f7f\u7528\u7684\u5206\u8bcd\u8ba1\u6570\u5668\uff08tokenizer\uff09\u3002\u8fd9\u79cd\u5199\u6cd5\u8ba9\u793a\u4f8b\u4ee3\u7801\u66f4\u5bb9\u6613\u8fd0\u884c\uff0c\u4f46 400 \u548c 80 \u53ea\u662f\u8fd9\u4e2a Demo \u7684\u9ed8\u8ba4\u503c\uff0c\u4e0d\u80fd\u7531\u6b64\u63a8\u65ad\u5b83\u4eec\u9002\u5408\u6240\u6709\u4e2d\u6587\u8d44\u6599\u3002<\/p>\n<p>\u9664\u4e86\u6587\u672c\u672c\u8eab\uff0c\u6bcf\u4e2a\u7247\u6bb5\u8fd8\u4fdd\u5b58\u6765\u6e90\u3001\u6807\u9898\u548c\u5b83\u5728\u539f\u6587\u4e2d\u7684\u8d77\u59cb\u4f4d\u7f6e\uff0c\u65b9\u4fbf\u5c55\u793a\u5f15\u7528\u5e76\u5b9a\u4f4d\u539f\u6587\u3002\u5f53\u524d Demo \u7684\u8d44\u6599\u8f83\u77ed\uff0c\u52a0\u5165\u590d\u6742\u7684 PDF \u7248\u9762\u89e3\u6790\u3001\u8bed\u4e49\u5206\u5757\u6216\u7236\u5b50\u5757\u68c0\u7d22\u672a\u5fc5\u80fd\u6539\u5584\u7ed3\u679c\uff0c\u8fd8\u4f1a\u589e\u52a0\u8bfb\u8005\u7406\u89e3\u6838\u5fc3\u68c0\u7d22\u6d41\u7a0b\u7684\u96be\u5ea6\u3002\u5982\u679c\u8d44\u6599\u662f\u6bcf\u9875\u5206\u6210\u5de6\u53f3\u51e0\u680f\u6392\u7248\u7684 PDF\u3001\u5305\u542b\u590d\u6742\u8868\u683c\uff0c\u6216\u8005\u662f\u4e00\u4efd\u5f88\u957f\u7684\u624b\u518c\uff0c\u518d\u6539\u7528\u80fd\u591f\u8bc6\u522b\u76f8\u5e94\u7ed3\u6784\u7684\u5207\u5206\u65b9\u6cd5\u66f4\u5408\u9002\u3002<\/p>\n<p>\u5728\u5b9e\u9645\u9879\u76ee\u4e2d\uff0c\u9700\u8981\u51c6\u5907\u4e00\u7ec4\u6d4b\u8bd5\u95ee\u9898\uff0c\u5e76\u63d0\u524d\u6807\u51fa\u6bcf\u4e2a\u95ee\u9898\u5e94\u8be5\u547d\u4e2d\u54ea\u6bb5\u539f\u6587\u3002\u7136\u540e\u5206\u522b\u8c03\u6574\u7247\u6bb5\u957f\u5ea6\u3001\u91cd\u53e0\u957f\u5ea6\u548c\u6807\u9898\u5207\u5206\u65b9\u5f0f\uff0c\u6bd4\u8f83\u54ea\u79cd\u8bbe\u7f6e\u66f4\u5bb9\u6613\u627e\u5230\u6b63\u786e\u5185\u5bb9\u3002\u5982\u679c\u8fd8\u6ca1\u6709\u8fd9\u7ec4\u6d4b\u8bd5\u95ee\u9898\uff0c\u53ef\u4ee5\u5148\u7528 400 \u548c 80 \u628a\u7cfb\u7edf\u8fd0\u884c\u8d77\u6765\uff1b\u5b83\u4eec\u662f\u5426\u5408\u9002\uff0c\u8981\u7b49\u5b9e\u9645\u68c0\u7d22\u7ed3\u679c\u9a8c\u8bc1\u540e\u624d\u80fd\u786e\u5b9a\u3002<\/p>\n<h2>Embedding \u5230\u5e95\u662f\u4ec0\u4e48<\/h2>\n<p>\u6587\u672c\u5411\u91cf\uff08Embedding\uff09\u662f\u7528\u5411\u91cf\u8868\u793a\u4e00\u6bb5\u6587\u5b57\uff0c\u4f7f\u7a0b\u5e8f\u53ef\u4ee5\u8ba1\u7b97\u4e24\u6bb5\u6587\u5b57\u5728\u8bed\u4e49\u4e0a\u7684\u76f8\u4f3c\u7a0b\u5ea6\uff1a<\/p>\n<pre><code class=\"text\">\u201c\u5dee\u65c5\u4ea4\u901a\u8d39\u5982\u4f55\u62a5\u9500\u201d\n    \u2193\n[0.021, -0.114, 0.307, ...]<\/code><\/pre>\n<p>\u7ecf\u8fc7\u8bad\u7ec3\u7684 Embedding \u6a21\u578b\u4f1a\u8ba9\u610f\u601d\u76f8\u8fd1\u7684\u6587\u5b57\u5f97\u5230\u76f8\u8fd1\u7684\u5411\u91cf\u3002\u56e0\u6b64\uff0c\u7528\u6237\u67e5\u201c\u6253\u8f66\u8d39\u8d85\u989d\u201d\uff0c\u5373\u4f7f\u539f\u6587\u5199\u7684\u662f\u201c\u5e02\u5185\u4ea4\u901a\u8d39\u7528\u8d85\u8fc7\u9650\u989d\u201d\uff0c\u5411\u91cf\u68c0\u7d22\u4ecd\u6709\u673a\u4f1a\u628a\u5b83\u4eec\u5339\u914d\u8d77\u6765\u3002<\/p>\n<p>\u8fd9\u4e2a Demo \u4f7f\u7528\u7684\u6a21\u578b\u4f1a\u4e3a\u6bcf\u6bb5\u6587\u5b57\u8f93\u51fa 384 \u7ef4\u5411\u91cf\u3002\u68c0\u7d22\u65f6\uff0c\u7a0b\u5e8f\u6bd4\u8f83\u95ee\u9898\u5411\u91cf\u548c\u7247\u6bb5\u5411\u91cf\u7684\u6574\u4f53\u65b9\u5411\uff1b\u4e24\u8005\u8d8a\u63a5\u8fd1\uff0c\u5bf9\u5e94\u6587\u672c\u7684\u8bed\u4e49\u76f8\u5173\u6027\u901a\u5e38\u8d8a\u9ad8\u3002\u5411\u91cf\u7684\u5404\u4e2a\u7ef4\u5ea6\u7531\u6a21\u578b\u8bad\u7ec3\u5f97\u5230\uff0c\u6ca1\u6709\u201c\u7b2c 7 \u7ef4\u56fa\u5b9a\u8868\u793a\u62a5\u9500\u201d\u8fd9\u6837\u7684\u4eba\u5de5\u5b9a\u4e49\u3002<\/p>\n<p>Embedding \u53ea\u8d1f\u8d23\u8868\u793a\u548c\u6bd4\u8f83\u6587\u672c\uff0c\u4e0d\u4f1a\u76f4\u63a5\u751f\u6210\u7b54\u6848\u3002\u4e24\u4e2a\u5411\u91cf\u63a5\u8fd1\uff0c\u53ea\u80fd\u8bf4\u660e\u4e24\u6bb5\u6587\u5b57\u5728\u5f53\u524d\u6a21\u578b\u751f\u6210\u7684\u5411\u91cf\u7a7a\u95f4\u4e2d\u76f8\u5173\uff0c\u4e0d\u80fd\u8bc1\u660e\u5b83\u4eec\u8868\u8fbe\u7684\u662f\u540c\u4e00\u4e2a\u4e8b\u5b9e\u3002<\/p>\n<p>\u7406\u89e3\u4e86 Embedding \u7684\u4f5c\u7528\u4ee5\u540e\uff0c\u8fd8\u8981\u533a\u5206\u5b83\u7684\u90e8\u7f72\u65b9\u5f0f\u3002\u8fd9\u91cc\u53ef\u4ee5\u4ece\u4e24\u4e2a\u7ef4\u5ea6\u6765\u770b\uff1a\u6a21\u578b\u8fd0\u884c\u5728\u54ea\u91cc\uff0c\u4ee5\u53ca\u5e94\u7528\u901a\u8fc7\u4ec0\u4e48\u65b9\u5f0f\u8c03\u7528\u5b83\u3002\u4e0b\u8868\u4e2d\u7684 ONNX \u662f\u4e00\u79cd\u901a\u7528\u7684\u6a21\u578b\u6587\u4ef6\u683c\u5f0f\uff0c\u7528\u6765\u4fdd\u5b58\u5df2\u7ecf\u8bad\u7ec3\u597d\u7684\u6a21\u578b\uff1b\u5e94\u7528\u901a\u8fc7\u76f8\u5e94\u7684\u8fd0\u884c\u5e93\u8bfb\u53d6\u6587\u4ef6\u5e76\u5b8c\u6210\u8ba1\u7b97\u3002<\/p>\n<table>\n<thead>\n<tr>\n<th>\u7ef4\u5ea6<\/th>\n<th>\u9009\u62e9<\/th>\n<th>\u542b\u4e49<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>\u90e8\u7f72\u4f4d\u7f6e<\/td>\n<td>\u4e91\u7aef<\/td>\n<td>\u6587\u672c\u53d1\u9001\u7ed9\u6258\u7ba1\u670d\u52a1\uff0c\u7531\u670d\u52a1\u8fd4\u56de\u5411\u91cf<\/td>\n<\/tr>\n<tr>\n<td>\u90e8\u7f72\u4f4d\u7f6e<\/td>\n<td>\u672c\u5730<\/td>\n<td>\u6a21\u578b\u5728\u7528\u6237\u8bbe\u5907\u6216\u81ea\u6709\u670d\u52a1\u5668\u8fd0\u884c<\/td>\n<\/tr>\n<tr>\n<td>\u5e38\u89c1\u5b9e\u73b0\u65b9\u5f0f<\/td>\n<td>\u4e91\u7aef Embedding API<\/td>\n<td>\u63a5\u5165\u7b80\u5355\uff0c\u9700\u8981\u8003\u8651\u8d39\u7528\u3001\u7f51\u7edc\u548c\u6570\u636e\u8fb9\u754c<\/td>\n<\/tr>\n<tr>\n<td>\u5e38\u89c1\u5b9e\u73b0\u65b9\u5f0f<\/td>\n<td>\u672c\u5730\u6a21\u578b\u5de5\u5177 Ollama<\/td>\n<td>Ollama \u7ba1\u7406\u6a21\u578b\uff0c\u5e76\u901a\u8fc7\u672c\u673a HTTP \u63a5\u53e3\u63d0\u4f9b Embedding<\/td>\n<\/tr>\n<tr>\n<td>\u5e38\u89c1\u5b9e\u73b0\u65b9\u5f0f<\/td>\n<td>\u672c\u5730 ONNX \u6a21\u578b<\/td>\n<td>\u6a21\u578b\u4fdd\u5b58\u5728 ONNX \u6587\u4ef6\u4e2d\uff0c\u7531\u5e94\u7528\u5185\u7684\u8fd0\u884c\u5e93\u8bfb\u53d6\u5e76\u8ba1\u7b97\u5411\u91cf<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>\u8868\u4e2d\u7684\u90e8\u7f72\u4f4d\u7f6e\u548c\u5b9e\u73b0\u65b9\u5f0f\u4e0d\u662f\u540c\u4e00\u5c42\u6982\u5ff5\uff1a\u524d\u8005\u63cf\u8ff0\u6a21\u578b\u4f4d\u4e8e\u4e91\u7aef\u8fd8\u662f\u672c\u5730\uff0c\u540e\u8005\u63cf\u8ff0\u5e94\u7528\u5982\u4f55\u8c03\u7528\u6a21\u578b\u3002Ollama \u50cf\u4e00\u4e2a\u5355\u72ec\u8fd0\u884c\u7684\u6a21\u578b\u670d\u52a1\uff0c\u5e94\u7528\u901a\u8fc7\u672c\u673a\u63a5\u53e3\u8bf7\u6c42\u5b83\u3002\u4f7f\u7528 ONNX \u6587\u4ef6\u65f6\uff0c\u8fd0\u884c\u6a21\u578b\u7684\u4ee3\u7801\u53ef\u4ee5\u653e\u5728\u5f53\u524d\u5e94\u7528\u5185\u90e8\uff0c\u4e0d\u5fc5\u53e6\u5916\u542f\u52a8 Ollama \u670d\u52a1\u3002<\/p>\n<h3>Demo \u5982\u4f55\u5728 Node.js \u4e2d\u751f\u6210\u5411\u91cf<\/h3>\n<p>\u8fd9\u4e2a Demo \u4f7f\u7528 <code>multilingual-e5-small<\/code> \u628a\u95ee\u9898\u548c\u77e5\u8bc6\u5e93\u7247\u6bb5\u8f6c\u6362\u6210\u5411\u91cf\u3002\u5b83\u662f\u4e00\u4e2a\u6a21\u578b\u540d\u79f0\uff0c\u5176\u4e2d multilingual \u8868\u793a\u652f\u6301\u591a\u79cd\u8bed\u8a00\uff0cE5 \u662f\u6a21\u578b\u7cfb\u5217\uff0csmall \u8868\u793a\u8fd9\u662f\u8be5\u7cfb\u5217\u4e2d\u4f53\u79ef\u8f83\u5c0f\u7684\u7248\u672c\u3002\u4ee3\u7801\u4f7f\u7528\u5b8c\u6574\u6807\u8bc6 <code>Xenova\/multilingual-e5-small<\/code>\uff0c\u524d\u9762\u7684 Xenova \u662f Hugging Face \u6a21\u578b\u4ed3\u5e93\u7684\u53d1\u5e03\u8005\u540d\u79f0\u3002<\/p>\n<p>\u9879\u76ee\u6ca1\u6709\u53e6\u5916\u542f\u52a8\u4e00\u4e2a\u6a21\u578b\u670d\u52a1\u3002Node.js \u670d\u52a1\u7aef\u7b2c\u4e00\u6b21\u521d\u59cb\u5316 Embedding \u65f6\uff0c<code>@huggingface\/transformers<\/code> \u4f1a\u4ece Hugging Face Hub \u4e0b\u8f7d\u6a21\u578b\u6743\u91cd\u3001Tokenizer \u548c\u914d\u7f6e\u6587\u4ef6\uff0c\u603b\u8ba1\u7ea6 135 MB\uff1b\u5176\u4e2d q8 ONNX \u6743\u91cd\u7ea6 118 MB\u3002\u8fd9\u4e9b\u6587\u4ef6\u7f13\u5b58\u5230\u9879\u76ee\u7684 <code>.cache\/transformers<\/code> \u76ee\u5f55\uff0c\u6a21\u578b\u518d\u88ab\u52a0\u8f7d\u5230\u5f53\u524d\u670d\u52a1\u8fdb\u7a0b\u4e2d\u3002\u540e\u7eed\u542f\u52a8\u4f1a\u4f18\u5148\u8bfb\u53d6\u7f13\u5b58\u3002\u6bcf\u6b21\u4f20\u5165\u95ee\u9898\u6216\u6587\u672c\u7247\u6bb5\uff0c\u670d\u52a1\u7aef\u90fd\u4f1a\u76f4\u63a5\u8c03\u7528\u5df2\u7ecf\u52a0\u8f7d\u7684\u6a21\u578b\u5e76\u5f97\u5230 384 \u7ef4\u5411\u91cf\u3002<\/p>\n<p>\u9879\u76ee\u52a0\u8f7d\u7684\u662f q8 \u91cf\u5316\u7248\u672c\u3002\u91cf\u5316\u662f\u628a\u6a21\u578b\u53c2\u6570\u4ece\u8f83\u9ad8\u7cbe\u5ea6\u538b\u7f29\u5230\u8f83\u4f4e\u7cbe\u5ea6\uff1bq8 \u8868\u793a\u8fd9\u91cc\u4f7f\u7528 8 \u4f4d\u6574\u6570\u4fdd\u5b58\u53c2\u6570\u3002\u8fd9\u6837\u53ef\u4ee5\u51cf\u5c0f\u6a21\u578b\u6587\u4ef6\u548c\u5185\u5b58\u5360\u7528\uff0c\u4ee3\u4ef7\u662f\u8ba1\u7b97\u7cbe\u5ea6\u53ef\u80fd\u6709\u5c11\u91cf\u635f\u5931\u3002<\/p>\n<p>\u4ece\u4e00\u6bb5\u6587\u672c\u5230\u6700\u7ec8\u5411\u91cf\uff0c\u4f1a\u7ecf\u8fc7\u4e0b\u9762\u51e0\u6b65\uff1a<\/p>\n<pre style=\"display:none;\"><code class=\"mermaid\">flowchart LR\n    T[\u95ee\u9898\u6216\u6587\u672c\u7247\u6bb5] --&gt; P[\u6dfb\u52a0 query \u6216 passage \u524d\u7f00]\n    P --&gt; Z[Tokenizer \u5207\u6210\u6a21\u578b\u53ef\u5904\u7406\u7684 token]\n    Z --&gt; E[E5 ONNX \u6a21\u578b\u8ba1\u7b97\u6bcf\u4e2a token \u7684\u8868\u793a]\n    E --&gt; M[\u6309 attention mask \u5bf9\u6709\u6548 token \u53d6\u5e73\u5747]\n    M --&gt; N[\u5c06\u5411\u91cf\u957f\u5ea6\u5f52\u4e00\u5316]\n    N --&gt; V[\u5f97\u5230 384 \u7ef4\u5411\u91cf]<\/code><\/pre>\n<p>\u6838\u5fc3\u4ee3\u7801\u4f4d\u4e8e <code>packages\/server\/src\/embedding.ts<\/code>\uff1a<\/p>\n<pre><code class=\"typescript\">const instance = await pipeline(\n  'feature-extraction',\n  'Xenova\/multilingual-e5-small',\n  {\n    dtype: 'q8',\n    revision: EMBEDDING_REVISION,\n  }\n);\n\nconst output = await instance(text, {\n  pooling: 'mean',\n  normalize: true,\n});<\/code><\/pre>\n<p>\u6a21\u578b\u63a5\u6536\u7684\u4e0d\u662f\u539f\u59cb\u5b57\u7b26\u4e32\uff0c\u800c\u662f Tokenizer \u751f\u6210\u7684\u6574\u6570 ID\u3002\u4ee5\u201c\u5dee\u65c5\u4ea4\u901a\u8d39\u5982\u4f55\u62a5\u9500\uff1f\u201d\u4e3a\u4f8b\uff0c\u5728\u672c\u6587\u8bb0\u5f55\u7684\u6a21\u578b revision \u548c Transformers.js 3.8.1 \u4e0b\uff0c\u5b9e\u9645\u5207\u5206\u7ed3\u679c\u5982\u4e0b\uff1a<\/p>\n<pre><code class=\"text\">query: \u5dee\u65c5\u4ea4\u901a\u8d39\u5982\u4f55\u62a5\u9500\uff1f\n    \u2193\ntoken\uff1a&lt;s&gt; | que | ry | : | \u7a7a\u683c\u6807\u8bb0 | \u5dee | \u65c5 | \u4ea4\u901a |\n       \u8d39 | \u5982\u4f55 | \u62a5 | \u9500 | ? | &lt;\/s&gt;\n\nID\uff1a[0, 41, 1294, 12, 6, 9470, 27783, 10766,\n    14479, 5733, 9001, 29906, 32, 2]<\/code><\/pre>\n<p>\u8fd9\u6b21\u8f93\u5165\u4e00\u5171\u5f97\u5230 14 \u4e2a token\u3002<code>&lt;s&gt;<\/code> \u548c <code>&lt;\/s&gt;<\/code> \u662f Tokenizer \u81ea\u52a8\u6dfb\u52a0\u7684\u8d77\u6b62\u6807\u8bb0\uff1b<code>query:<\/code> \u672c\u8eab\u88ab\u62c6\u6210 <code>que<\/code>\u3001<code>ry<\/code> \u548c <code>:<\/code>\uff0c\u5e76\u4e0d\u662f\u4e24\u4e2a\u7279\u6b8a token\u3002\u4e2d\u6587\u4e5f\u4e0d\u4e00\u5b9a\u6309\u5355\u4e2a\u6c49\u5b57\u5207\u5206\uff1a\u201c\u5dee\u201d\u201c\u65c5\u201d\u201c\u8d39\u201d\u5206\u522b\u5360\u4e00\u4e2a token\uff0c\u800c\u201c\u4ea4\u901a\u201d\u201c\u5982\u4f55\u201d\u5404\u81ea\u5408\u6210\u4e00\u4e2a token\u3002\u6362\u4e00\u6bb5\u6587\u5b57\u3001\u6a21\u578b revision \u6216 Tokenizer \u7248\u672c\uff0c\u5207\u5206\u7ed3\u679c\u90fd\u53ef\u80fd\u53d8\u5316\uff0c\u56e0\u6b64\u8fd9\u4e9b ID \u53ea\u7528\u4e8e\u89e3\u91ca\u5f53\u524d Demo\uff0c\u4e0d\u80fd\u5f53\u4f5c E5 \u7684\u56fa\u5b9a\u89c4\u5219\u3002<\/p>\n<p><code>pipeline(...)<\/code> \u8d1f\u8d23\u52a0\u8f7d ONNX \u6a21\u578b\uff0c<code>feature-extraction<\/code> \u8868\u793a\u53d6\u5f97\u6587\u672c\u8868\u793a\u800c\u4e0d\u662f\u751f\u6210\u56de\u7b54\u3002\u6a21\u578b\u5148\u4e3a 14 \u4e2a token \u5206\u522b\u751f\u6210\u4e00\u7ec4\u5305\u542b 384 \u4e2a\u6570\u5b57\u7684\u5411\u91cf\uff0c\u6240\u4ee5\u6b64\u65f6\u5171\u6709 14 \u7ec4\u5411\u91cf\u3002\u8fd9\u91cc\u7684\u201c\u6c60\u5316\u201d\uff08pooling\uff09\u5c31\u662f\u628a\u591a\u7ec4 token \u5411\u91cf\u5408\u5e76\u6210\u4e00\u4e2a\u80fd\u4ee3\u8868\u6574\u53e5\u8bdd\u7684\u5411\u91cf\u3002<code>pooling: 'mean'<\/code> \u4f7f\u7528\u7684\u662f\u5e73\u5747\u6c60\u5316\uff1aattention mask \u4f1a\u6807\u8bb0\u54ea\u4e9b\u4f4d\u7f6e\u662f\u771f\u5b9e\u8f93\u5165\u3001\u54ea\u4e9b\u53ea\u662f\u4e3a\u4e86\u7edf\u4e00\u957f\u5ea6\u800c\u8865\u9f50\u7684\u7a7a\u4f4d\uff0c\u7a0b\u5e8f\u5148\u5ffd\u7565\u8fd9\u4e9b\u7a7a\u4f4d\uff0c\u518d\u628a\u771f\u5b9e token \u7684\u5411\u91cf\u9010\u9879\u6c42\u5e73\u5747\u3002\u4f8b\u5982\uff0c14 \u7ec4\u5411\u91cf\u7684\u7b2c 1 \u4e2a\u6570\u5b57\u53d6\u5e73\u5747\u540e\u6210\u4e3a\u53e5\u5b50\u5411\u91cf\u7684\u7b2c 1 \u4e2a\u6570\u5b57\uff0c\u5176\u4f59 383 \u4e2a\u4f4d\u7f6e\u4e5f\u505a\u540c\u6837\u5904\u7406\uff0c\u6700\u7ec8\u5f97\u5230\u4e00\u7ec4 384 \u7ef4\u5411\u91cf\u3002<code>normalize: true<\/code> \u518d\u628a\u8fd9\u7ec4\u5411\u91cf\u7684\u957f\u5ea6\u8c03\u6574\u4e3a 1\uff0c\u4fbf\u4e8e\u540e\u7eed\u6bd4\u8f83\u65b9\u5411\u3002<\/p>\n<p>\u6bcf\u8f93\u5165\u4e00\u6bb5\u6587\u5b57\uff0c\u8fd9\u6bb5\u4ee3\u7801\u90fd\u4f1a\u8fd4\u56de\u4e00\u4e2a 384 \u7ef4\u5411\u91cf\u3002\u7d22\u5f15\u8fd8\u4f1a\u8bb0\u5f55\u6a21\u578b\u540d\u79f0\u548c\u7248\u672c\uff0c\u68c0\u7d22\u65f6\u53ea\u6bd4\u8f83\u7531\u540c\u4e00\u6a21\u578b\u3001\u540c\u4e00\u7248\u672c\u751f\u6210\u7684\u5411\u91cf\u3002<\/p>\n<p>E5 \u5728\u8bad\u7ec3\u65f6\u7528\u4e0d\u540c\u524d\u7f00\u533a\u5206\u201c\u7528\u6237\u8981\u67e5\u4ec0\u4e48\u201d\u548c\u201c\u53ef\u4f9b\u641c\u7d22\u7684\u8d44\u6599\u201d\uff0c\u56e0\u6b64\u4ee3\u7801\u9700\u8981\u5206\u522b\u6dfb\u52a0 <code>query:<\/code> \u548c <code>passage:<\/code>\uff1a<\/p>\n<pre><code class=\"typescript\">embedQuery(question)  \/\/ \u5b9e\u9645\u8f93\u5165\uff1aquery: ${question}\nembedPassage(chunk)   \/\/ \u5b9e\u9645\u8f93\u5165\uff1apassage: ${chunk}<\/code><\/pre>\n<p><code>query:<\/code> \u6807\u8bb0\u540e\u9762\u7684\u6587\u5b57\u662f\u7528\u6237\u95ee\u9898\uff0c<code>passage:<\/code> \u6807\u8bb0\u540e\u9762\u7684\u6587\u5b57\u662f\u77e5\u8bc6\u5e93\u7247\u6bb5\u3002\u8fd9\u4e24\u4e2a\u524d\u7f00\u53ea\u53c2\u4e0e\u5411\u91cf\u8ba1\u7b97\uff0c\u9875\u9762\u4ecd\u7136\u5c55\u793a\u539f\u6587\u3002E5 \u7684\u6a21\u578b\u8bf4\u660e\u8981\u6c42\u68c0\u7d22\u4efb\u52a1\u6309\u8fd9\u79cd\u683c\u5f0f\u8f93\u5165\uff0c\u7701\u7565\u524d\u7f00\u4f1a\u4f7f\u8f93\u5165\u683c\u5f0f\u4e0e\u8bad\u7ec3\u65b9\u5f0f\u4e0d\u4e00\u81f4\uff0c\u5e76\u53ef\u80fd\u964d\u4f4e\u68c0\u7d22\u6548\u679c\u3002<\/p>\n<h2>\u5411\u91cf\u68c0\u7d22\u662f\u600e\u6837\u6392\u5e8f\u7684<\/h2>\n<p>Demo \u4f7f\u7528\u201c\u4f59\u5f26\u76f8\u4f3c\u5ea6\u201d\u6bd4\u8f83\u95ee\u9898\u5411\u91cf\u548c\u6587\u6863\u5411\u91cf\u3002\u5b83\u5173\u6ce8\u4e24\u4e2a\u5411\u91cf\u7684\u65b9\u5411\u662f\u5426\u63a5\u8fd1\u3002A \u548c B \u5206\u522b\u4ee3\u8868\u4e24\u4e2a\u5411\u91cf\uff0c\u7a0b\u5e8f\u4f1a\u5bf9\u4e24\u8005\u5bf9\u5e94\u7684\u7ef4\u5ea6\u8fdb\u884c\u8ba1\u7b97\uff1a<\/p>\n<pre><code class=\"text\">cos(A, B) = (A \u00b7 B) \/ (||A|| \u00d7 ||B||)<\/code><\/pre>\n<p>\u8fd9\u9879\u8ba1\u7b97\u4f1a\u5f97\u5230 <code>-1<\/code> \u5230 <code>1<\/code> \u4e4b\u95f4\u7684\u5206\u6570\u3002\u5206\u6570\u8d8a\u63a5\u8fd1 <code>1<\/code>\uff0c\u4e24\u6bb5\u6587\u5b57\u5728\u5f53\u524d\u6a21\u578b\u4e0b\u7684\u76f8\u5173\u6027\u8d8a\u9ad8\uff1b\u5206\u6570\u8f83\u4f4e\uff0c\u76f8\u5173\u6027\u4e5f\u8d8a\u5f31\u3002<\/p>\n<p>\u4f59\u5f26\u76f8\u4f3c\u5ea6\u4e0d\u662f\u552f\u4e00\u7684\u5411\u91cf\u6bd4\u8f83\u65b9\u5f0f\u3002\u5e38\u89c1\u65b9\u6cd5\u8fd8\u6709\u70b9\u79ef\u548c\u6b27\u6c0f\u8ddd\u79bb\uff1a\u70b9\u79ef\u540c\u65f6\u53d7\u65b9\u5411\u548c\u5411\u91cf\u957f\u5ea6\u5f71\u54cd\uff0c\u6b27\u6c0f\u8ddd\u79bb\u8861\u91cf\u4e24\u4e2a\u5411\u91cf\u5728\u7a7a\u95f4\u4e2d\u76f8\u9694\u591a\u8fdc\u3002\u5177\u4f53\u5e94\u8be5\u4f7f\u7528\u54ea\u4e00\u79cd\uff0c\u8981\u770b Embedding \u6a21\u578b\u7684\u8bad\u7ec3\u65b9\u5f0f\u548c\u7d22\u5f15\u914d\u7f6e\u3002\u8fd9\u4e2a Demo \u5df2\u901a\u8fc7 <code>normalize: true<\/code> \u628a\u5411\u91cf\u957f\u5ea6\u7edf\u4e00\u4e3a 1\uff1b\u5728\u8fd9\u79cd\u60c5\u51b5\u4e0b\uff0c\u6309\u4f59\u5f26\u76f8\u4f3c\u5ea6\u6216\u70b9\u79ef\u4ece\u9ad8\u5230\u4f4e\u6392\u5217\uff0c\u4e0e\u6309\u6b27\u6c0f\u8ddd\u79bb\u4ece\u4f4e\u5230\u9ad8\u6392\u5217\uff0c\u4f1a\u5f97\u5230\u76f8\u540c\u7684\u5148\u540e\u987a\u5e8f\uff0c\u56e0\u6b64\u4ee3\u7801\u9009\u62e9\u4e86\u5bb9\u6613\u89e3\u91ca\u7684\u4f59\u5f26\u76f8\u4f3c\u5ea6\u3002<\/p>\n<p>\u4e0d\u8fc7\uff0c<code>0.86<\/code> \u4e0d\u8868\u793a\u201c\u7b54\u6848\u6709 86% \u6982\u7387\u6b63\u786e\u201d\u3002\u76f8\u4f3c\u5ea6\u53ea\u662f\u5f53\u524d\u6a21\u578b\u4e0b\u7684\u6392\u5e8f\u4fe1\u53f7\uff0c\u4e0d\u540c\u6a21\u578b\u3001\u8bed\u6599\u548c\u67e5\u8be2\u7684\u5206\u6570\u5206\u5e03\u90fd\u53ef\u80fd\u4e0d\u540c\u3002<\/p>\n<p>\u6b64\u5916\uff0c\u5411\u91cf\u68c0\u7d22\u867d\u7136\u80fd\u5339\u914d\u540c\u4e49\u8868\u8fbe\uff0c\u4f46\u4e0d\u64c5\u957f\u4fdd\u8bc1\u7cbe\u786e\u6807\u8bc6\u7b26\u5b8c\u5168\u4e00\u81f4\u3002<code>ERR_FIN_403<\/code> \u548c <code>ERR_FIN_404<\/code> \u53ea\u6709\u6700\u540e\u4e00\u4f4d\u6570\u5b57\u4e0d\u540c\uff1b\u5982\u679c\u4e24\u8005\u6240\u5728\u7247\u6bb5\u7684\u5176\u4ed6\u6587\u5b57\u4e5f\u5f88\u76f8\u4f3c\uff0c\u7eaf\u5411\u91cf\u68c0\u7d22\u53ef\u80fd\u628a\u5b83\u4eec\u6392\u5f97\u5f88\u8fd1\u3002\u4e3a\u4e86\u8ba9\u9519\u8bef\u7801\u7684\u5b8c\u6574\u5b57\u7b26\u4e32\u53c2\u4e0e\u6392\u5e8f\uff0cDemo \u8fd8\u52a0\u5165\u4e86 BM25\u3002<\/p>\n<h2>FTS\u3001\u5012\u6392\u7d22\u5f15\u4e0e BM25<\/h2>\n<p>\u524d\u9762\u63d0\u5230 BM25 \u6839\u636e\u67e5\u8be2\u8bcd\u5339\u914d\u60c5\u51b5\u6253\u5206\uff0c\u4f46\u5177\u4f53\u600e\u4e48\u627e\u5230\u54ea\u4e9b\u7247\u6bb5\u5305\u542b\u67e5\u8be2\u8bcd\uff1f\u6570\u636e\u91cf\u8f83\u5927\u7684\u751f\u4ea7\u7cfb\u7edf\u901a\u5e38\u4e0d\u4f1a\u50cf Demo \u90a3\u6837\u9010\u4e2a\u626b\u63cf\uff0c\u800c\u662f\u9760\u5168\u6587\u68c0\u7d22\uff08Full-Text Search\uff0cFTS\uff09\u52a0\u901f\u3002FTS \u8d1f\u8d23\u5207\u5206\u67e5\u8be2\u8bcd\u3001\u901a\u8fc7\u5012\u6392\u7d22\u5f15\u627e\u5230\u5305\u542b\u8fd9\u4e9b\u8bcd\u7684\u7247\u6bb5\uff0c\u518d\u5bf9\u7ed3\u679c\u6392\u5e8f\u3002BM25 \u53ea\u662f\u5176\u4e2d\u7684\u4e00\u79cd\u6392\u5e8f\u7b97\u6cd5\u3002<\/p>\n<p>\u8fd9\u91cc\u7684\u201c\u5207\u5206\u67e5\u8be2\u8bcd\u201d\u4e0d\u662f\u524d\u9762 Embedding \u6a21\u578b\u4f7f\u7528\u7684 Tokenizer\uff0c\u800c\u662f Demo \u4e3a BM25 \u5355\u72ec\u5b9e\u73b0\u7684\u7b80\u5355\u5206\u8bcd\u89c4\u5219\u3002\u5b83\u4f1a\u628a\u8fde\u7eed\u4e2d\u6587\u62c6\u6210\u76f8\u90bb\u7684\u4e24\u4e2a\u5b57\uff0c\u4f8b\u5982\u201c\u6253\u8f66\u8d39\u201d\u4f1a\u5f97\u5230\u201c\u6253\u8f66\u201d\u548c\u201c\u8f66\u8d39\u201d\uff1b\u82f1\u6587\u3001\u6570\u5b57\u548c\u4e0b\u5212\u7ebf\u7ec4\u6210\u7684 <code>ERR_FIN_403<\/code> \u5219\u4f1a\u4fdd\u7559\u4e3a\u4e00\u4e2a\u8bcd\u9879\u3002\u67e5\u8be2\u548c\u6587\u6863\u90fd\u4f7f\u7528\u8fd9\u5957\u89c4\u5219\uff0c\u7a0b\u5e8f\u624d\u80fd\u7edf\u8ba1\u540c\u4e00\u4e2a\u8bcd\u9879\u51fa\u73b0\u4e86\u51e0\u6b21\u3002<\/p>\n<p>\u751f\u4ea7\u73af\u5883\u7684 FTS \u5f15\u64ce\u901a\u5e38\u6709\u81ea\u5df1\u7684 tokenizer\uff0c\u4f46\u201c\u81ea\u5e26 tokenizer\u201d\u4e0d\u7b49\u4e8e\u81ea\u52a8\u9002\u5408\u4e2d\u6587\u3002\u4f8b\u5982 SQLite FTS5 \u63d0\u4f9b <code>unicode61<\/code>\u3001<code>trigram<\/code> \u7b49\u5185\u7f6e\u65b9\u6848\uff0c\u4e5f\u5141\u8bb8\u63a5\u5165\u81ea\u5b9a\u4e49 tokenizer\uff1b\u9879\u76ee\u4ecd\u8981\u6839\u636e\u4e2d\u6587\u8bcd\u8bed\u3001\u9519\u8bef\u7801\u548c\u77ed\u8bed\u67e5\u8be2\u7684\u5b9e\u9645\u6548\u679c\u9009\u62e9\u914d\u7f6e\u3002\u6574\u4f53\u601d\u8def\u4e0d\u53d8\uff1a\u5148\u628a\u6587\u6863\u548c\u67e5\u8be2\u6309\u540c\u4e00\u89c4\u5219\u8f6c\u6362\u6210\u8bcd\u9879\uff0c\u518d\u901a\u8fc7\u5012\u6392\u7d22\u5f15\u5b9a\u4f4d\u5019\u9009\u7247\u6bb5\u3002<\/p>\n<p>FTS \u53ef\u4ee5\u6539\u7528\u5176\u4ed6\u6392\u5e8f\u7b97\u6cd5\uff08\u5982 PostgreSQL FTS \u7684 <code>ts_rank<\/code>\uff09\uff0cBM25 \u4e5f\u53ef\u4ee5\u8131\u79bb FTS\uff1a\u672c Demo \u5c31\u6ca1\u6709 FTS \u548c\u5012\u6392\u7d22\u5f15\uff0c\u800c\u662f\u5728\u7528\u6237\u6bcf\u6b21\u67e5\u8be2\u65f6\u904d\u5386\u6240\u6709\u7247\u6bb5\u5e76\u9010\u4e2a\u8ba1\u7b97 BM25 \u5206\u6570\u3002<\/p>\n<p>FTS \u5148\u627e\u51fa\u5019\u9009\u7247\u6bb5\uff0cBM25 \u518d\u4e3a\u8fd9\u4e9b\u5019\u9009\u6253\u5206\u6392\u5e8f\u3002\u5728 Elasticsearch \u7b49\u7cfb\u7edf\u4e2d\uff0c\u8fd9\u4e24\u6b65\u7531\u4e00\u6b21\u641c\u7d22\u8bf7\u6c42\u5b8c\u6210\u3002<\/p>\n<p>\u5012\u6392\u7d22\u5f15\u662f FTS \u52a0\u901f\u67e5\u627e\u7684\u6838\u5fc3\uff0c\u5b83\u628a\u201c\u7247\u6bb5\u5305\u542b\u54ea\u4e9b\u8bcd\u201d\u7684\u6b63\u5411\u5173\u7cfb\u53cd\u8fc7\u6765\uff0c\u5b58\u6210\u201c\u8bcd\u51fa\u73b0\u5728\u54ea\u4e9b\u7247\u6bb5\u201d\u3002<\/p>\n<p>\u4ee5\u67e5\u8be2 <code>ERR_FIN_403<\/code> \u4e3a\u4f8b\uff0cFTS \u7684\u5904\u7406\u6d41\u7a0b\u5982\u4e0b\uff1a<\/p>\n<pre style=\"display:none;\"><code class=\"mermaid\">flowchart LR\n    Q[&quot;\u67e5\u8be2\uff1aERR_FIN_403&quot;] --&gt; T[&quot;\u5206\u8bcd\uff1aerr_fin_403&quot;]\n    T --&gt; I[&quot;\u67e5\u5012\u6392\u7d22\u5f15\uff1aerr_fin_403 -&gt; \u7247\u6bb5 5&quot;]\n    I --&gt; C[&quot;\u8df3\u8fc7\u5176\u4ed6\u7247\u6bb5\uff0c\u76f4\u63a5\u62ff\u5230\u7247\u6bb5 5&quot;]\n    C --&gt; B[&quot;BM25 \u8ba1\u7b97\u7247\u6bb5 5 \u7684\u76f8\u5173\u6027\u5206\u6570&quot;]\n    B --&gt; R[&quot;\u6309\u5206\u6570\u6392\u5e8f&quot;]<\/code><\/pre>\n<p>\u4ee5 Demo \u7684\u8d44\u6599\u4e3a\u4f8b\uff0c\u5012\u6392\u7d22\u5f15\u5c55\u5f00\u540e\u5982\u4e0b\uff08\u5b9e\u9645\u4fdd\u5b58\u7684\u662f\u7247\u6bb5 ID\uff0c\u8fd9\u91cc\u5c55\u5f00\u4e3a\u6807\u9898\u548c\u539f\u6587\u65b9\u4fbf\u6838\u5bf9\uff09\uff1a<\/p>\n<pre><code class=\"text\">err_fin_403 -&gt; \u62a5\u9500\u5165\u53e3\uff08&quot;\u7cfb\u7edf\u9519\u8bef\u7801 ERR_FIN_403&quot;\uff09\n\u62a5\u9500        -&gt; \u5e02\u5185\u4ea4\u901a\u8d39\u7528\uff08&quot;\u6bcf\u6708\u62a5\u9500\u9650\u989d\u4e3a 800 \u5143&quot;\uff09\n              \u62a5\u9500\u5165\u53e3\uff08&quot;\u5728\u9996\u9875\u9009\u62e9\u8d39\u7528\u62a5\u9500&quot;\uff09\n\u9650\u989d        -&gt; \u5e02\u5185\u4ea4\u901a\u8d39\u7528\uff08&quot;\u6bcf\u6708\u62a5\u9500\u9650\u989d\u4e3a 800 \u5143&quot;\uff09<\/code><\/pre>\n<p>\u7d22\u5f15\u5efa\u7acb\u65f6\u7edf\u4e00\u8f6c\u5c0f\u5199\uff0c\u6240\u4ee5\u539f\u6587\u7684 <code>ERR_FIN_403<\/code> \u53ef\u4ee5\u6309 <code>err_fin_403<\/code> \u67e5\u5230\u3002\u7528\u6237\u67e5\u8be2 <code>ERR_FIN_403<\/code> \u65f6\uff0cFTS \u76f4\u63a5\u4ece\u7d22\u5f15\u4e2d\u53d6\u51fa\u5bf9\u5e94\u7684\u7247\u6bb5\uff0c\u4e0d\u5fc5\u9010\u4e2a\u68c0\u67e5\u5176\u4ed6\u7247\u6bb5\u3002\u5982\u679c\u67e5\u8be2\u5305\u542b\u591a\u4e2a\u8bcd\u3001\u547d\u4e2d\u4e86\u591a\u4e2a\u7247\u6bb5\uff0cBM25 \u518d\u8ba1\u7b97\u6bcf\u4e2a\u5019\u9009\u7684\u76f8\u5173\u6027\u5206\u6570\u5e76\u5b8c\u6210\u6392\u5e8f\u3002SQLite FTS5 \u548c Elasticsearch \u7684\u5168\u6587\u68c0\u7d22\u53ef\u4ee5\u4f7f\u7528 BM25\uff0cPostgreSQL FTS \u5219\u63d0\u4f9b <code>ts_rank<\/code> \u548c <code>ts_rank_cd<\/code> \u7b49\u5176\u4ed6\u6392\u5e8f\u51fd\u6570\u3002<\/p>\n<p>\u76f8\u6bd4\u4e4b\u4e0b\uff0cDemo \u6ca1\u6709\u5b9e\u73b0 FTS\uff0c\u4e5f\u6ca1\u6709\u5efa\u7acb\u5012\u6392\u7d22\u5f15\uff0c\u800c\u662f\u91c7\u7528\u4e0b\u9762\u8fd9\u6761\u6d41\u7a0b\uff1a<\/p>\n<pre style=\"display:none;\"><code class=\"mermaid\">flowchart LR\n    Q[\u67e5\u8be2] --&gt; S[\u904d\u5386\u5168\u90e8 chunk]\n    S --&gt; B[\u9010\u4e2a\u8ba1\u7b97 BM25 \u5206\u6570]\n    B --&gt; R[\u6309\u5206\u6570\u6392\u5e8f]<\/code><\/pre>\n<p>Demo \u91c7\u7528\u5168\u91cf\u626b\u63cf\uff0c\u9996\u5148\u662f\u56e0\u4e3a\u5f53\u524d\u6570\u636e\u91cf\u5f88\u5c0f\uff0c\u9010\u4e2a\u68c0\u67e5\u7247\u6bb5\u4e0d\u4f1a\u9020\u6210\u660e\u663e\u5ef6\u8fdf\u3002\u5176\u6b21\uff0c\u6587\u7ae0\u9700\u8981\u76f4\u63a5\u5c55\u793a\u5206\u8bcd\u3001\u8bcd\u9879\u7edf\u8ba1\u548c BM25 \u5206\u6570\uff1b\u63a5\u5165 FTS \u5f15\u64ce\u540e\uff0c\u8fd9\u4e9b\u6b65\u9aa4\u4f1a\u88ab\u5c01\u88c5\u5728\u6570\u636e\u5e93\u6216\u641c\u7d22\u670d\u52a1\u5185\u90e8\u3002\u9879\u76ee\u4e5f\u4e0d\u5fc5\u4e3a\u6559\u5b66\u4ee3\u7801\u589e\u52a0\u4e00\u5957\u5012\u6392\u7d22\u5f15\u7684\u589e\u5220\u66f4\u65b0\u903b\u8f91\u3002<\/p>\n<p>\u8fd9\u6837\u5199\u8ba9\u8bfb\u8005\u53ef\u4ee5\u76f4\u63a5\u770b\u5230 BM25 \u7684\u8ba1\u7b97\u8fc7\u7a0b\uff0c\u4f46\u4e5f\u5e26\u6765\u4e86\u9650\u5236\uff1a\u6587\u672c\u7247\u6bb5\u589e\u591a\u540e\uff0c\u9010\u4e2a\u626b\u63cf\u4f1a\u8d8a\u6765\u8d8a\u6162\uff0cDemo \u4e5f\u6ca1\u6709\u77ed\u8bed\u67e5\u8be2\u3001\u524d\u7f00\u67e5\u8be2\u548c\u7ed3\u679c\u9ad8\u4eae\u7b49\u5b8c\u6574\u7684\u5168\u6587\u68c0\u7d22\u80fd\u529b\u3002\u8fdb\u5165\u751f\u4ea7\u73af\u5883\u6216\u6570\u636e\u91cf\u660e\u663e\u589e\u52a0\u540e\uff0c\u53ef\u4ee5\u6539\u7528 SQLite FTS5\u3001PostgreSQL FTS \u6216 Elasticsearch\u3002<\/p>\n<p>\u8fd9\u4e9b FTS \u7cfb\u7edf\u4f1a\u8fd4\u56de\u4e00\u4efd\u6309\u5173\u952e\u8bcd\u76f8\u5173\u6027\u6392\u5217\u7684\u7ed3\u679c\uff0c\u5411\u91cf\u68c0\u7d22\u5219\u8fd4\u56de\u53e6\u4e00\u4efd\u6309\u8bed\u4e49\u76f8\u4f3c\u5ea6\u6392\u5217\u7684\u7ed3\u679c\u3002\u6240\u8c13\u201c\u628a\u5173\u952e\u8bcd\u7ed3\u679c\u4e0e\u5411\u91cf\u7ed3\u679c\u878d\u5408\u201d\uff0c\u5c31\u662f\u628a\u540c\u4e00\u4e2a\u7247\u6bb5\u5728\u4e24\u8def\u68c0\u7d22\u4e2d\u7684\u5206\u6570\u5408\u5e76\uff0c\u518d\u751f\u6210\u4e00\u4efd\u65b0\u7684\u603b\u6392\u540d\uff1b\u5e76\u4e0d\u662f\u5148\u7528 FTS \u67e5\u5b8c\uff0c\u518d\u53ea\u5bf9 FTS \u7684\u7ed3\u679c\u505a\u5411\u91cf\u68c0\u7d22\u3002<\/p>\n<p>\u4e86\u89e3 FTS \u548c BM25 \u7684\u5206\u5de5\u4ee5\u540e\uff0c\u518d\u770b BM25 \u5982\u4f55\u8ba1\u7b97\u5173\u952e\u8bcd\u76f8\u5173\u5ea6\u3002\u5b83\u4f1a\u7efc\u5408\u4e0b\u9762\u4e09\u4e2a\u56e0\u7d20\uff1a<\/p>\n<ul>\n<li>\u8bcd\u9891\uff08Term Frequency\uff0cTF\uff09\uff1a\u7edf\u8ba1\u67e5\u8be2\u8bcd\u5728\u5f53\u524d\u7247\u6bb5\u4e2d\u51fa\u73b0\u51e0\u6b21\u3002\u4e00\u4e2a\u7247\u6bb5\u591a\u6b21\u51fa\u73b0\u67e5\u8be2\u8bcd\uff0c\u901a\u5e38\u6bd4\u53ea\u51fa\u73b0\u4e00\u6b21\u66f4\u76f8\u5173\uff1b\u4f46\u4ece\u7b2c\u5341\u6b21\u589e\u52a0\u5230\u7b2c\u5341\u4e00\u6b21\uff0c\u4e0d\u4f1a\u83b7\u5f97\u4e0e\u7b2c\u4e00\u6b21\u51fa\u73b0\u540c\u6837\u591a\u7684\u52a0\u5206\u3002<\/li>\n<li>\u9006\u6587\u6863\u9891\u7387\uff08Inverse Document Frequency\uff0cIDF\uff09\uff1a\u7edf\u8ba1\u4e00\u4e2a\u67e5\u8be2\u8bcd\u51fa\u73b0\u5728\u591a\u5c11\u4e2a\u7247\u6bb5\u4e2d\u3002\u5305\u542b\u5b83\u7684\u7247\u6bb5\u8d8a\u5c11\uff0c\u8fd9\u4e2a\u8bcd\u8d8a\u80fd\u5e2e\u52a9\u5b9a\u4f4d\uff0cIDF \u6743\u91cd\u5c31\u8d8a\u9ad8\u3002\u4f8b\u5982 <code>ERR_FIN_403<\/code> \u53ea\u51fa\u73b0\u5728\u5c11\u6570\u7247\u6bb5\u4e2d\uff0c\u6743\u91cd\u8f83\u9ad8\uff1b\u201c\u62a5\u9500\u201d\u51fa\u73b0\u5728\u5f88\u591a\u7247\u6bb5\u4e2d\uff0c\u65e0\u6cd5\u5355\u72ec\u5b9a\u4f4d\u67d0\u4e00\u6bb5\uff0c\u6240\u4ee5\u201c\u62a5\u9500\u201d\u7684\u6743\u91cd\u8f83\u4f4e\u3002\u8fd9\u4e24\u4e2a\u4f8b\u5b50\u5206\u522b\u8bf4\u660e\u9ad8\u6743\u91cd\u548c\u4f4e\u6743\u91cd\u3002<\/li>\n<li>\u6587\u6863\u957f\u5ea6\u5f52\u4e00\u5316\uff1a\u540c\u4e00\u4e2a\u8bcd\u5728\u4e00\u6bb5\u5f88\u77ed\u7684\u8bf4\u660e\u4e2d\u51fa\u73b0\u4e00\u6b21\uff0c\u901a\u5e38\u6bd4\u5b83\u5728\u4e00\u672c\u5f88\u957f\u7684\u624b\u518c\u4e2d\u5076\u7136\u51fa\u73b0\u4e00\u6b21\u66f4\u80fd\u8bf4\u660e\u4e3b\u9898\u76f8\u5173\u3002BM25 \u4f1a\u6839\u636e\u5f53\u524d\u7247\u6bb5\u957f\u5ea6\u548c\u5e73\u5747\u7247\u6bb5\u957f\u5ea6\u4fee\u6b63\u5206\u6570\uff0c\u907f\u514d\u957f\u7247\u6bb5\u4ec5\u4ec5\u56e0\u4e3a\u5305\u542b\u7684\u8bcd\u591a\u5c31\u5360\u4f18\u52bf\u3002<\/li>\n<\/ul>\n<p>\u628a\u8fd9\u4e09\u4e2a\u56e0\u7d20\u5199\u6210\u516c\u5f0f\u5982\u4e0b\uff1a<\/p>\n<pre><code class=\"text\">score(D, Q) = \u03a3 IDF(q) \u00d7 f(q,D) \u00d7 (k1 + 1)\n                         -------------------------------\n                         f(q,D) + k1 \u00d7 (1 - b + b \u00d7 |D|\/avgdl)<\/code><\/pre>\n<p>\u516c\u5f0f\u4e2d\u7684\u7b26\u53f7\u542b\u4e49\u5982\u4e0b\uff1a<\/p>\n<ul>\n<li><code>D<\/code> \u662f\u5f53\u524d\u7247\u6bb5\uff0c<code>Q<\/code> \u662f\u7528\u6237\u67e5\u8be2\u3002<\/li>\n<li><code>q<\/code> \u662f\u67e5\u8be2\u4e2d\u7684\u4e00\u4e2a\u8bcd\uff0c<code>f(q,D)<\/code> \u662f\u8fd9\u4e2a\u8bcd\u5728\u5f53\u524d\u7247\u6bb5\u4e2d\u7684\u8bcd\u9891\u3002<\/li>\n<li><code>|D|\/avgdl<\/code> \u662f\u5f53\u524d\u7247\u6bb5\u957f\u5ea6\u4e0e\u5e73\u5747\u7247\u6bb5\u957f\u5ea6\u7684\u6bd4\u503c\u3002<\/li>\n<li><code>k1<\/code> \u63a7\u5236\u8bcd\u9891\u589e\u52a0\u5230\u591a\u5c11\u6b21\u540e\u9010\u6e10\u505c\u6b62\u660e\u663e\u52a0\u5206\u3002<\/li>\n<li><code>b<\/code> \u63a7\u5236\u7247\u6bb5\u957f\u5ea6\u5bf9\u5206\u6570\u7684\u5f71\u54cd\u7a0b\u5ea6\u3002<\/li>\n<\/ul>\n<p>Demo \u4f7f\u7528 <code>k1 = 1.2<\/code>\u3001<code>b = 0.75<\/code>\u3002\u8fd9\u662f BM25 \u7684\u5e38\u89c1\u9ed8\u8ba4\u914d\u7f6e\uff0c\u5b9e\u9645\u9879\u76ee\u4ecd\u8981\u6839\u636e\u68c0\u7d22\u8bc4\u4f30\u7ed3\u679c\u8c03\u6574\u3002<\/p>\n<p>BM25 \u53ea\u80fd\u7edf\u8ba1\u5df2\u7ecf\u5207\u5206\u597d\u7684\u8bcd\uff0c\u56e0\u6b64\u8ba1\u7b97\u524d\u8981\u5148\u5206\u8bcd\u3002\u82f1\u6587\u53e5\u5b50\u901a\u5e38\u6709\u7a7a\u683c\uff0c\u4f8b\u5982\u201ctravel expense\u201d\u53ef\u4ee5\u76f4\u63a5\u5207\u6210\u201ctravel\u201d\u548c\u201cexpense\u201d\u3002\u4e2d\u6587\u53e5\u5b50\u6ca1\u6709\u7a7a\u683c\uff0c\u201c\u6253\u8f66\u8d39\u201d\u7a76\u7adf\u5e94\u8be5\u5207\u6210\u201c\u6253\u8f66\u8d39\u201d\u201c\u6253\u8f66 + \u8d39\u201d\u8fd8\u662f\u201c\u6253 + \u8f66\u8d39\u201d\uff0c\u9700\u8981\u5206\u8bcd\u5668\u51b3\u5b9a\u3002<\/p>\n<p>\u4e3a\u4e86\u907f\u514d\u989d\u5916\u5f15\u5165\u4e2d\u6587\u5206\u8bcd\u5e93\uff0c\u672c\u9879\u76ee\u91c7\u7528\u4e86\u4e00\u4e2a\u7b80\u5355\u89c4\u5219\uff1a\u8fde\u7eed\u4e2d\u6587\u6bcf\u76f8\u90bb\u4e24\u4e2a\u5b57\u7ec4\u6210\u4e00\u4e2a\u4e8c\u5143\u8bcd\u7ec4\uff0c\u5e76\u5411\u540e\u79fb\u52a8\u4e00\u4e2a\u5b57\u7ee7\u7eed\u5207\u5206\u3002\u4f8b\u5982\uff1a<\/p>\n<pre><code class=\"text\">\u6253\u8f66\u8d39\u8d85\u8fc7\u516b\u767e\u5757\u5e94\u8be5\u600e\u4e48\u529e\uff1f\n    \u2193\n\u6253\u8f66 | \u8f66\u8d39 | \u8d39\u8d85 | \u8d85\u8fc7 | \u8fc7\u516b | \u516b\u767e |\n\u767e\u5757 | \u5757\u5e94 | \u5e94\u8be5 | \u8be5\u600e | \u600e\u4e48 | \u4e48\u529e<\/code><\/pre>\n<p>\u8fd9\u91cc\u7684\u201c\u4e8c\u5143\u201d\u5c31\u662f\u201c\u4e24\u4e2a\u5b57\u7b26\u201d\uff0c\u6240\u4ee5\u201c\u6253\u8f66\u8d39\u201d\u4f1a\u5f97\u5230\u201c\u6253\u8f66\u201d\u548c\u201c\u8f66\u8d39\u201d\u3002\u82f1\u6587\u3001\u6570\u5b57\u548c\u4e0b\u5212\u7ebf\u8fde\u7eed\u7ec4\u6210\u7684\u5185\u5bb9\u5219\u4f5c\u4e3a\u4e00\u4e2a\u5b8c\u6574\u8bcd\u9879\uff0c\u5e76\u7edf\u4e00\u8f6c\u6210\u5c0f\u5199\u3002<code>ERR_FIN_403<\/code> \u4f1a\u53d8\u6210\u201cerr_fin_403\u201d\uff0c\u4e0d\u4f1a\u88ab\u62c6\u6210\u201cerr\u201d\u201cfin\u201d\u548c\u201c403\u201d\u3002<\/p>\n<p>\u8fd9\u79cd\u5207\u6cd5\u5bb9\u6613\u5b9e\u73b0\uff0c\u4e5f\u80fd\u5b8c\u6574\u5339\u914d\u9519\u8bef\u7801\uff0c\u4f46\u5b83\u4e0d\u8bc6\u522b\u8bcd\u4e49\u548c\u8bcd\u8bed\u8fb9\u754c\u3002\u4f8b\u5982\uff0c\u201c\u8d39\u8d85\u201d\u53ea\u662f\u76f8\u90bb\u5b57\u7b26\uff0c\u4e0d\u662f\u4e00\u4e2a\u6709\u610f\u4e49\u7684\u4e2d\u6587\u8bcd\uff1b\u201c\u8d85\u8fc7\u201d\u51fa\u73b0\u5728\u4e0d\u540c\u4e3b\u9898\u7684\u7247\u6bb5\u4e2d\u65f6\uff0c\u4e5f\u53ef\u80fd\u4ea7\u751f\u4e0d\u5408\u7406\u7684\u9ad8\u5206\u3002\u540e\u6587\u7684\u5c0f\u8282\u201c\u4e00\u6b21\u5b9e\u9645\u7684\u68c0\u7d22\u8c03\u8bd5\u201d\u4e2d\u7684\u5b9e\u6d4b\u7ed3\u679c\u5c31\u51fa\u73b0\u4e86\u8fd9\u4e2a\u95ee\u9898\uff1a\u7531\u4e8e\u4e2d\u6587\u4e8c\u5143\u8bcd\u7ec4\u628a\u201c\u8d85\u8fc7\u201d\u62c6\u6210\u4e86\u4e00\u4e2a\u6709\u6548\u8bcd\u9879\uff0c\u201c\u4f4f\u5bbf\u6807\u51c6\u201d\u7247\u6bb5\u83b7\u5f97\u4e86\u6700\u9ad8\u7684 BM25 \u5206\u6570\uff0c\u751a\u81f3\u538b\u8fc7\u4e86\u5411\u91cf\u68c0\u7d22\u6392\u5728\u7b2c\u4e00\u7684\u6b63\u786e\u7247\u6bb5\u3002<\/p>\n<h2>\u5411\u91cf\u4e0e BM25 \u7684\u6df7\u5408\u68c0\u7d22<\/h2>\n<p>\u5411\u91cf\u68c0\u7d22\u4e0e BM25 \u89e3\u51b3\u7684\u662f\u4e24\u7c7b\u4e0d\u540c\u95ee\u9898\uff1a<\/p>\n<table>\n<thead>\n<tr>\n<th>\u67e5\u8be2<\/th>\n<th>\u66f4\u6709\u4ef7\u503c\u7684\u4fe1\u53f7<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>\u201c\u6253\u8f66\u8d39\u8d85\u989d\u600e\u4e48\u529e\u201d\u5339\u914d\u201c\u5e02\u5185\u4ea4\u901a\u8d39\u7528\u8d85\u8fc7\u9650\u989d\u201d<\/td>\n<td>\u5411\u91cf\u8bed\u4e49<\/td>\n<\/tr>\n<tr>\n<td><code>ERR_FIN_403<\/code>\u3001\u5408\u540c\u53f7\u3001\u51fd\u6570\u540d<\/td>\n<td>BM25 \u7cbe\u786e\u8bcd\u9879<\/td>\n<\/tr>\n<tr>\n<td>\u7528\u6237\u548c\u539f\u6587\u4f7f\u7528\u4e0d\u540c\u8bed\u8a00\u6216\u540c\u4e49\u8bcd<\/td>\n<td>\u591a\u8bed\u8a00\u5411\u91cf\u6a21\u578b<\/td>\n<\/tr>\n<tr>\n<td>\u5c11\u89c1\u7f16\u53f7\u5728\u957f\u6587\u6863\u4e2d\u53ea\u51fa\u73b0\u4e00\u6b21<\/td>\n<td>BM25 \u7684 IDF<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>\u5411\u91cf\u68c0\u7d22\u548c BM25 \u4f1a\u7ed9\u540c\u4e00\u4e2a\u7247\u6bb5\u7b97\u51fa\u4e24\u79cd\u5206\u6570\uff0c\u4f46\u8fd9\u4e24\u79cd\u5206\u6570\u4e0d\u80fd\u76f4\u63a5\u76f8\u52a0\u3002\u4f59\u5f26\u76f8\u4f3c\u5ea6\u901a\u5e38\u5c0f\u4e8e 1\uff0cBM25 \u5206\u6570\u5374\u53ef\u80fd\u662f 2\u30016 \u6216\u66f4\u5927\u3002\u5982\u679c\u76f4\u63a5\u8ba1\u7b97 <code>0.86 + 6.2<\/code>\uff0cBM25 \u4ec5\u4ec5\u56e0\u4e3a\u6570\u503c\u8303\u56f4\u66f4\u5927\u5c31\u4f1a\u63a7\u5236\u6700\u7ec8\u6392\u540d\u3002<\/p>\n<p>Demo \u5148\u8ba9\u4e24\u6761\u5206\u652f\u5404\u53d6 <code>2 \u00d7 Top-K<\/code> \u4e2a\u5019\u9009\u3002\u9ed8\u8ba4 <code>Top-K = 5<\/code>\uff0c\u6240\u4ee5\u5411\u91cf\u68c0\u7d22\u548c BM25 \u6700\u591a\u5404\u4ea4\u51fa 10 \u4e2a\u7247\u6bb5\u3002\u7a0b\u5e8f\u6309\u7167\u7247\u6bb5 ID \u5408\u5e76\u4e24\u4efd\u5217\u8868\uff1a\u540c\u4e00\u4e2a\u7247\u6bb5\u5982\u679c\u4e24\u8fb9\u90fd\u51fa\u73b0\uff0c\u53ea\u4fdd\u7559\u4e00\u6761\u8bb0\u5f55\u5e76\u540c\u65f6\u4fdd\u5b58\u4e24\u8def\u5206\u6570\uff1b\u53ea\u5728\u4e00\u8def\u51fa\u73b0\u7684\u7247\u6bb5\uff0c\u53e6\u4e00\u8def\u5206\u6570\u8bb0\u4e3a 0\u3002\u8fd9\u4efd\u53bb\u91cd\u540e\u7684\u5217\u8868\u5c31\u662f\u4e24\u8def\u5019\u9009\u7684\u5e76\u96c6\u3002<\/p>\n<p>\u63a5\u4e0b\u6765\uff0c\u7a0b\u5e8f\u628a\u6bcf\u4e00\u8def\u7684\u6700\u9ad8\u5206\u6362\u7b97\u6210 <code>1<\/code>\uff0c\u5176\u4f59\u5206\u6570\u6309\u76f8\u540c\u6bd4\u4f8b\u7f29\u5c0f\u3002\u8fd9\u4e2a\u8fc7\u7a0b\u79f0\u4e3a\u5f52\u4e00\u5316\uff1a<\/p>\n<pre><code class=\"text\">vectorNorm = max(0, vectorScore) \/ maxVectorScore\nbm25Norm   = bm25Score \/ maxBm25Score\nfinalScore = 0.7 \u00d7 vectorNorm + 0.3 \u00d7 bm25Norm<\/code><\/pre>\n<p>\u7136\u540e\u518d\u6309\u7167\u9ed8\u8ba4\u6743\u91cd\u8ba1\u7b97\u603b\u5206\uff1a\u5411\u91cf\u5206\u5360 <code>70%<\/code>\uff0cBM25 \u5206\u5360 <code>30%<\/code>\u3002\u8fd9\u91cc\u7684\u6743\u91cd\u8868\u793a\u4e24\u8def\u5206\u6570\u5bf9\u6392\u5e8f\u7684\u5f71\u54cd\u5927\u5c0f\uff0c\u4e0d\u662f\u7b54\u6848\u6b63\u786e\u7387\u3002<\/p>\n<p>\u7a0b\u5e8f\u4f1a\u5bf9\u5019\u9009\u5217\u8868\u4e2d\u7684\u6bcf\u4e2a\u7247\u6bb5\u6267\u884c\u8fd9\u5957\u8ba1\u7b97\uff0c\u6309\u6700\u7ec8\u5206\u91cd\u65b0\u6392\u5e8f\uff0c\u518d\u8fd4\u56de\u524d K \u6761\u3002\u4e0b\u4e00\u8282\u4f1a\u7528\u4e00\u7ec4\u5b9e\u9645\u7ed3\u679c\u8bf4\u660e\uff1a\u5411\u91cf\u5206\u8f83\u9ad8\u7684\u6b63\u786e\u7247\u6bb5\uff0c\u4e3a\u4ec0\u4e48\u53ef\u80fd\u88ab\u540c\u65f6\u83b7\u5f97\u4e24\u8def\u52a0\u5206\u7684\u9519\u8bef\u7247\u6bb5\u8d85\u8fc7\u3002<\/p>\n<p>\u6838\u5fc3\u5408\u5e76\u903b\u8f91\u53ef\u4ee5\u6982\u62ec\u4e3a\uff1a<\/p>\n<pre><code class=\"typescript\">const vectorContribution = normalizedVectorScore * vectorWeight;\nconst keywordContribution = normalizedKeywordScore * keywordWeight;\nconst finalScore = vectorContribution + keywordContribution;<\/code><\/pre>\n<p>\u9875\u9762\u4f1a\u540c\u65f6\u5c55\u793a\u539f\u59cb\u5206\u3001\u5f52\u4e00\u5316\u5206\u3001\u52a0\u6743\u8d21\u732e\u548c\u6700\u7ec8\u5206\uff0c\u65b9\u4fbf\u68c0\u67e5\u67d0\u4e2a\u7247\u6bb5\u4e3a\u4ec0\u4e48\u6392\u5728\u5f53\u524d\u4f4d\u7f6e\u3002\u6743\u91cd\u8bbe\u4e3a 0 \u65f6\uff0c\u5bf9\u5e94\u7684\u68c0\u7d22\u5206\u652f\u4e0d\u4f1a\u6267\u884c\u3002\u751f\u4ea7\u7cfb\u7edf\u8fd8\u53ef\u4ee5\u4f7f\u7528\u6821\u51c6\u540e\u7684\u5206\u6570\uff0c\u6216\u8005\u6539\u7528\u53ea\u53c2\u8003\u4e24\u8def\u540d\u6b21\u7684\u5012\u6570\u6392\u540d\u878d\u5408\uff08Reciprocal Rank Fusion\uff0cRRF\uff09\u3002<\/p>\n<h2>RAG \u4e00\u5b9a\u9700\u8981\u5411\u91cf\u6570\u636e\u5e93\u5417<\/h2>\n<p>\u4e0d\u4e00\u5b9a\u3002RAG \u9700\u8981\u7684\u662f\u68c0\u7d22\u80fd\u529b\uff0c\u65e2\u4e0d\u5f3a\u5236\u4f7f\u7528\u4e13\u7528\u5411\u91cf\u6570\u636e\u5e93\uff0c\u4e5f\u4e0d\u5f3a\u5236\u53ea\u7528\u5411\u91cf\u68c0\u7d22\u3002BM25\u3001SQL\u3001\u77e5\u8bc6\u56fe\u8c31\u548c\u641c\u7d22 API \u90fd\u53ef\u4ee5\u627f\u62c5 Retrieval\u3002<\/p>\n<p>\u8fd9\u4e2a Demo \u628a\u539f\u6587\u3001384 \u7ef4\u5411\u91cf\u548c\u6765\u6e90\u4fe1\u606f\u4fdd\u5b58\u5728 <code>data\/vectors.json<\/code>\u3002\u670d\u52a1\u542f\u52a8\u540e\uff0cNode.js \u5c06\u5168\u90e8\u8bb0\u5f55\u8bfb\u5165\u5185\u5b58\uff1b\u67e5\u8be2\u65f6\uff0c\u5411\u91cf\u5206\u652f\u548c BM25 \u5206\u652f\u90fd\u626b\u63cf\u8fd9\u6279\u8bb0\u5f55\uff0c\u518d\u5408\u5e76\u4e24\u8def\u7ed3\u679c\u3002<code>vectors.json<\/code> \u53ea\u8d1f\u8d23\u6301\u4e45\u5316\uff0c\u672c\u8eab\u6ca1\u6709\u641c\u7d22\u80fd\u529b\u3002<\/p>\n<p>\u8fd9\u79cd\u65b9\u5f0f\u9002\u5408\u5c0f\u578b Demo\uff0c\u56e0\u4e3a\u6570\u636e\u7ed3\u6784\u548c\u8ba1\u7b97\u8fc7\u7a0b\u90fd\u80fd\u76f4\u63a5\u770b\u5230\u3002\u4ee3\u4ef7\u662f chunk \u8d8a\u591a\uff0c\u542f\u52a8\u65f6\u8bfb\u53d6\u7684\u6570\u636e\u8d8a\u591a\uff0c\u6bcf\u6b21\u67e5\u8be2\u8981\u6bd4\u8f83\u7684\u8bb0\u5f55\u4e5f\u8d8a\u591a\u3002\u662f\u5426\u5347\u7ea7\u5230\u5e26\u7d22\u5f15\u7684\u5b58\u50a8\u65b9\u6848\uff0c\u5e94\u7531\u6570\u636e\u91cf\u3001\u5ef6\u8fdf\u3001\u5e76\u53d1\u548c\u6743\u9650\u8fc7\u6ee4\u8981\u6c42\u51b3\u5b9a\u3002<\/p>\n<table>\n<thead>\n<tr>\n<th>\u65b9\u6848<\/th>\n<th>\u9002\u5408\u573a\u666f<\/th>\n<th>\u67e5\u8be2\u65b9\u5f0f<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>\u5185\u5b58\u6570\u7ec4\u6216 JSON + \u5185\u5b58<\/td>\n<td>\u5355\u5143\u6d4b\u8bd5\u3001\u5c0f\u578b Demo<\/td>\n<td>\u5e94\u7528\u8fdb\u7a0b\u5168\u91cf\u626b\u63cf<\/td>\n<\/tr>\n<tr>\n<td>\u6d4f\u89c8\u5668 IndexedDB<\/td>\n<td>\u7eaf\u672c\u5730\u4e2a\u4eba\u8d44\u6599\u5de5\u5177<\/td>\n<td>\u6d4f\u89c8\u5668\u626b\u63cf\uff0c\u6216\u53e6\u52a0\u7d22\u5f15\u5e93<\/td>\n<\/tr>\n<tr>\n<td>SQLite + sqlite-vec<\/td>\n<td>\u684c\u9762\u5e94\u7528\u3001\u5355\u673a\u670d\u52a1<\/td>\n<td>SQLite \u8ba1\u7b97\u5e76\u8fd4\u56de Top-K<\/td>\n<\/tr>\n<tr>\n<td>PostgreSQL + pgvector<\/td>\n<td>\u5df2\u6709\u4e1a\u52a1\u6570\u636e\u5e93\uff0c\u9700\u8981\u4e8b\u52a1\u548c\u5b57\u6bb5\u8fc7\u6ee4<\/td>\n<td>\u6570\u636e\u5e93\u6267\u884c\u5411\u91cf\u641c\u7d22<\/td>\n<\/tr>\n<tr>\n<td>Qdrant\u3001Milvus\u3001Weaviate\u3001Pinecone \u7b49<\/td>\n<td>\u5927\u89c4\u6a21\u3001\u9ad8\u5e76\u53d1\u3001\u72ec\u7acb\u68c0\u7d22\u670d\u52a1<\/td>\n<td>\u670d\u52a1\u7ba1\u7406\u7d22\u5f15\u3001\u8fc7\u6ee4\u4e0e\u5206\u7247<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>\u4f8b\u5982\uff0c\u4f7f\u7528 pgvector \u540e\uff0c\u53ef\u4ee5\u8ba9\u6570\u636e\u5e93\u5148\u6309\u79df\u6237\u8fc7\u6ee4\uff0c\u518d\u5b8c\u6210\u76f8\u4f3c\u5ea6\u6392\u5e8f\uff1a<\/p>\n<pre><code class=\"sql\">SELECT id, content, source\nFROM chunks\nWHERE tenant_id = $1\nORDER BY embedding &lt;=&gt; $2\nLIMIT 5;<\/code><\/pre>\n<p>\u5e94\u7528\u63d0\u4ea4\u67e5\u8be2\u5411\u91cf\u540e\u53ea\u63a5\u6536\u524d 5 \u6761\uff0c\u4e0d\u5fc5\u628a\u6574\u5f20\u5411\u91cf\u8868\u4f20\u7ed9 Node.js\u3002\u5f53\u524d Demo \u53ea\u6709\u5c11\u91cf\u8d44\u6599\uff0c\u4f7f\u7528 JSON \u548c\u5185\u5b58\u626b\u63cf\u5df2\u7ecf\u8db3\u591f\uff0c\u4e5f\u66f4\u9002\u5408\u5c55\u793a\u6df7\u5408\u68c0\u7d22\u7684\u5185\u90e8\u8fc7\u7a0b\u3002<\/p>\n<h2>\u4e00\u6b21\u5b9e\u9645\u7684\u68c0\u7d22\u8c03\u8bd5<\/h2>\n<p>\u4e3a\u4e86\u68c0\u67e5\u4e0a\u8ff0\u878d\u5408\u65b9\u5f0f\u7684\u5b9e\u9645\u6548\u679c\uff0c\u4e0b\u9762\u76f4\u63a5\u770b\u53ea\u5305\u542b 14 \u4e2a\u5185\u7f6e chunk \u65f6\u7684\u9875\u9762\u8fd0\u884c\u7ed3\u679c\u3002\u67e5\u8be2\u4f7f\u7528\u9875\u9762\u9ed8\u8ba4\u53c2\u6570\uff1a<code>Top-K = 5<\/code>\uff0c\u5411\u91cf\/BM25 \u6743\u91cd\u4e3a <code>0.7\/0.3<\/code>\u3002<\/p>\n<p>\u8fd9\u91cc\u6cbf\u7528\u524d\u9762\u7684\u67e5\u8be2\uff1a<\/p>\n<blockquote><p>\u6253\u8f66\u8d39\u8d85\u8fc7\u516b\u767e\u5757\u5e94\u8be5\u600e\u4e48\u529e\uff1f<\/p><\/blockquote>\n<p>\u5b8c\u6574\u7684 Top-5 \u5982\u4e0b\uff1a<\/p>\n<table>\n<thead>\n<tr>\n<th align=\"right\">\u540d\u6b21<\/th>\n<th>\u7247\u6bb5<\/th>\n<th align=\"right\">\u4f59\u5f26\u539f\u59cb\u5206<\/th>\n<th align=\"right\">\u5411\u91cf\u5f52\u4e00\u5316<\/th>\n<th align=\"right\">BM25 \u539f\u59cb\u5206<\/th>\n<th align=\"right\">BM25 \u5f52\u4e00\u5316<\/th>\n<th align=\"right\">\u6700\u7ec8\u878d\u5408\u5206<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td align=\"right\">1<\/td>\n<td>\u4f4f\u5bbf\u6807\u51c6<\/td>\n<td align=\"right\">0.843634<\/td>\n<td align=\"right\">0.960091<\/td>\n<td align=\"right\">2.613690<\/td>\n<td align=\"right\">1.000000<\/td>\n<td align=\"right\">0.972064<\/td>\n<\/tr>\n<tr>\n<td align=\"right\">2<\/td>\n<td>\u7968\u636e\u671f\u9650<\/td>\n<td align=\"right\">0.878703<\/td>\n<td align=\"right\">1.000000<\/td>\n<td align=\"right\">2.048268<\/td>\n<td align=\"right\">0.783669<\/td>\n<td align=\"right\">0.935101<\/td>\n<\/tr>\n<tr>\n<td align=\"right\">3<\/td>\n<td>\u5e02\u5185\u4ea4\u901a\u8d39\u7528<\/td>\n<td align=\"right\">0.872662<\/td>\n<td align=\"right\">0.993126<\/td>\n<td align=\"right\">0<\/td>\n<td align=\"right\">0<\/td>\n<td align=\"right\">0.695188<\/td>\n<\/tr>\n<tr>\n<td align=\"right\">4<\/td>\n<td>\u62a5\u9500\u5165\u53e3<\/td>\n<td align=\"right\">0.854296<\/td>\n<td align=\"right\">0.972224<\/td>\n<td align=\"right\">0<\/td>\n<td align=\"right\">0<\/td>\n<td align=\"right\">0.680557<\/td>\n<\/tr>\n<tr>\n<td align=\"right\">5<\/td>\n<td>\u52a0\u6743\u878d\u5408<\/td>\n<td align=\"right\">0.838667<\/td>\n<td align=\"right\">0.954437<\/td>\n<td align=\"right\">0<\/td>\n<td align=\"right\">0<\/td>\n<td align=\"right\">0.668106<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p><img decoding=\"async\"   referrerpolicy=\"no-referrer\" src=\"https:\/\/segmentfault.com\/img\/remote\/1460000048055389\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" title=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\"><\/p>\n<p>\u7ed3\u679c\u5e76\u4e0d\u6f02\u4eae\uff1a\u6b63\u786e\u7684\u201c\u5e02\u5185\u4ea4\u901a\u8d39\u7528\u201d\u6392\u5728\u7b2c 3\u3002\u201c\u4f4f\u5bbf\u6807\u51c6\u201d\u540c\u65f6\u542b\u6709\u201c\u8d85\u51fa\u6807\u51c6\u201d\uff0c\u4e2d\u6587\u4e8c\u5143\u8bcd\u7ec4\u4f7f\u5b83\u501f\u52a9\u76f8\u8fd1\u8bcd\u9879\u83b7\u5f97\u6700\u9ad8 BM25 \u5206\uff1b\u5f52\u4e00\u5316\u4e4b\u540e\uff0c\u8fd9\u4e2a\u7247\u6bb5\u53c8\u5f97\u5230 0.3 \u7684\u5173\u952e\u8bcd\u8d21\u732e\uff0c\u6700\u7ec8\u8d85\u8fc7\u4e86\u76f8\u5173\u6027\u66f4\u9ad8\u7684\u4ea4\u901a\u7247\u6bb5\u3002<\/p>\n<p>BM25 \u7684\u539f\u59cb\u5206\u6570\u4f1a\u968f\u7740\u7d22\u5f15\u5185\u5bb9\u53d8\u5316\uff0c\u56e0\u4e3a IDF \u9700\u8981\u7edf\u8ba1\u4e00\u4e2a\u8bcd\u51fa\u73b0\u5728\u5168\u90e8 chunk \u4e2d\u7684\u6bd4\u4f8b\u3002\u8fd9\u91cc\u7684 <code>2.613690<\/code> \u548c <code>2.048268<\/code> \u53ea\u5bf9\u5e94\u5f53\u524d 14 \u4e2a\u5185\u7f6e chunk\uff1b\u5982\u679c\u52a0\u5165\u7528\u6237\u4e0a\u4f20\u7684\u6587\u6863\uff0c\u5373\u4f7f\u67e5\u8be2\u548c\u53c2\u6570\u4e0d\u53d8\uff0cBM25 \u539f\u59cb\u5206\u548c\u6700\u7ec8\u6392\u540d\u4e5f\u53ef\u80fd\u53d8\u5316\u3002<\/p>\n<p>\u8fd9\u4e2a\u7ed3\u679c\u8bf4\u660e\uff0c\u4e0d\u80fd\u6839\u636e\u4e00\u6761\u56de\u7b54\u662f\u5426\u901a\u987a\u6765\u9a8c\u6536 RAG\u3002\u5f53\u524d Top-5 \u4ecd\u5305\u542b\u6b63\u786e\u8d44\u6599\uff0c\u6240\u4ee5 LLM \u8fd8\u6709\u673a\u4f1a\u56de\u7b54\u6b63\u786e\uff1b\u4f46\u566a\u58f0\u5df2\u7ecf\u8fdb\u5165\u4e0a\u4e0b\u6587\uff0c\u6362\u4e00\u4e2a Top-K \u6216\u6362\u4e00\u6279\u6587\u6863\u5c31\u53ef\u80fd\u6f0f\u6389\u7b54\u6848\u3002\u9488\u5bf9\u8fd9\u4e2a\u95ee\u9898\uff0c\u53ef\u4ee5\u4ece\u4ee5\u4e0b\u51e0\u4e2a\u65b9\u5411\u6539\u8fdb\uff1a<\/p>\n<ul>\n<li>\u7528\u66f4\u5408\u9002\u7684\u4e2d\u6587\u5206\u8bcd\u5668\uff0c\u51cf\u5c11\u8de8\u8bcd\u8fb9\u754c\u7684\u4f2a\u547d\u4e2d\u3002<\/li>\n<li>\u964d\u4f4e\u5f53\u524d\u8bed\u6599\u4e0a\u7684 BM25 \u6743\u91cd\uff0c\u6216\u5728\u9a8c\u8bc1\u96c6\u4e0a\u91cd\u65b0\u8c03\u53c2\u3002<\/li>\n<li>\u5148\u627e\u51fa\u66f4\u591a\u5019\u9009\uff0c\u518d\u589e\u52a0\u91cd\u6392\u6a21\u578b\uff08reranker\uff09\uff0c\u8ba9\u5b83\u5bf9\u5019\u9009\u505a\u4e00\u6b21\u66f4\u7cbe\u7ec6\u7684\u6392\u5e8f\u3002<\/li>\n<li>\u8c03\u6574\u79cd\u5b50\u8d44\u6599\u548c\u5206\u5757\uff0c\u907f\u514d\u6807\u9898\u6216\u77ed\u7247\u6bb5\u4ea7\u751f\u5f02\u5e38\u9ad8\u5206\u3002<\/li>\n<li>\u4f7f\u7528\u5012\u6570\u6392\u540d\u878d\u5408\uff08Reciprocal Rank Fusion\uff0cRRF\uff09\u7b49\u66f4\u770b\u91cd\u6392\u540d\u3001\u8f83\u5c11\u53d7\u539f\u59cb\u5206\u6570\u5f71\u54cd\u7684\u5408\u5e76\u65b9\u6cd5\uff0c\u5e76\u7528\u8bc4\u4f30\u96c6\u9a8c\u8bc1\u3002<\/li>\n<\/ul>\n<p>\u8fd9\u4e00\u8282\u5148\u628a\u4e00\u4e2a\u6392\u5e8f\u5f02\u5e38\u89e3\u91ca\u6e05\u695a\u3002\u5176\u4ed6\u4e09\u4e2a\u9875\u9762\u793a\u4f8b\u9898\u4f1a\u653e\u5230\u540e\u9762\u7684\u8bc4\u4f30\u7ae0\u8282\u4e2d\u7edf\u4e00\u6bd4\u8f83\uff0c\u907f\u514d\u5728\u4e0d\u540c\u7d22\u5f15\u6761\u4ef6\u4e0b\u91cd\u590d\u6d4b\u8bd5\u3002<\/p>\n<h2>\u628a\u68c0\u7d22\u7ed3\u679c\u4ea4\u7ed9 LLM<\/h2>\n<p>\u68c0\u7d22\u5b8c\u6210\u540e\uff0c\u670d\u52a1\u7aef\u4f1a\u7ed9\u672c\u6b21\u56de\u7b54\u4e2d\u7684\u6bcf\u4e2a\u53bb\u91cd\u7247\u6bb5\u5206\u914d\u7f16\u53f7\uff0c\u518d\u628a\u4e0b\u9762\u8fd9\u6837\u7684\u6587\u672c\u4ea4\u7ed9 LLM\u3002\u6309\u7167\u524d\u9762\u7684\u5b9e\u9645\u6392\u540d\uff0c\u201c\u5e02\u5185\u4ea4\u901a\u8d39\u7528\u201d\u662f <code>[\u6765\u6e90 3]<\/code>\uff1a<\/p>\n<pre><code class=\"text\">[\u6765\u6e90 1]\n\u6587\u4ef6\uff1a\u661f\u6cb3\u79d1\u6280\u5dee\u65c5\u4e0e\u4ea4\u901a\u62a5\u9500\u5236\u5ea6\n\u6807\u9898\uff1a\u4f4f\u5bbf\u6807\u51c6\n\u5185\u5bb9\uff1a\u4e0a\u6d77\u5730\u533a\u4f4f\u5bbf\u6807\u51c6\u2026\u2026\n\n[\u6765\u6e90 3]\n\u6587\u4ef6\uff1a\u661f\u6cb3\u79d1\u6280\u5dee\u65c5\u4e0e\u4ea4\u901a\u62a5\u9500\u5236\u5ea6\n\u6807\u9898\uff1a\u5e02\u5185\u4ea4\u901a\u8d39\u7528\n\u5185\u5bb9\uff1a\u5458\u5de5\u56e0\u516c\u4ea7\u751f\u7684\u5e02\u5185\u4ea4\u901a\u8d39\u7528\u2026\u2026<\/code><\/pre>\n<p>\u8fd9\u4e9b\u7247\u6bb5\u5199\u5165\u63d0\u793a\u8bcd\u540e\uff0c\u7cfb\u7edf\u4f1a\u8981\u6c42 LLM \u53ea\u6839\u636e\u8d44\u6599\u9648\u8ff0\u4e8b\u5b9e\uff0c\u5e76\u4e3a\u7ed3\u8bba\u6807\u6ce8 <code>[\u6765\u6e90 N]<\/code>\u3002\u5982\u679c\u8d44\u6599\u4e0d\u8db3\uff0cLLM \u5e94\u8bf4\u660e\u7f3a\u5c11\u4ec0\u4e48\uff0c\u4e0d\u80fd\u4f7f\u7528\u6a21\u578b\u8bb0\u5fc6\u8865\u5168\u3002\u6b64\u5916\uff0c\u77e5\u8bc6\u5e93\u6587\u672c\u6309\u4e0d\u53ef\u4fe1\u6570\u636e\u5904\u7406\uff0c\u5176\u4e2d\u5373\u4f7f\u51fa\u73b0\u201c\u5ffd\u7565\u7cfb\u7edf\u89c4\u5219\u201d\u4e4b\u7c7b\u7684\u63d0\u793a\u8bcd\uff0c\u4e5f\u4e0d\u80fd\u8986\u76d6\u7cfb\u7edf\u6307\u4ee4\u3002<\/p>\n<p>\u5728\u8fd9\u4e9b\u7ea6\u675f\u4e0b\uff0c\u6a21\u578b\u53ef\u80fd\u8fd4\u56de\u4e0b\u9762\u7684\u7b54\u6848\uff1a<\/p>\n<blockquote><p>\u8d85\u51fa\u6bcf\u6708 800 \u5143\u9650\u989d\u7684\u90e8\u5206\uff0c\u9700\u8981\u90e8\u95e8\u8d1f\u8d23\u4eba\u5ba1\u6279\uff0c\u5e76\u5728\u62a5\u9500\u5355\u4e2d\u8bf4\u660e\u4e1a\u52a1\u539f\u56e0\u3002[\u6765\u6e90 3]<\/p><\/blockquote>\n<p>\u4e0e\u6b64\u540c\u65f6\uff0c\u524d\u7aef\u8fd8\u4f1a\u6536\u5230\u672c\u8f6e\u7d2f\u8ba1\u68c0\u7d22\u5230\u7684\u53bb\u91cd\u6765\u6e90\u3001\u5382\u5546\u4e0e\u6a21\u578b\u540d\uff0c\u4ee5\u53ca Agent \u6b65\u9aa4\u6570\u548c\u68c0\u7d22\u6b21\u6570\u3002\u8fd9\u91cc\u7684 <code>sources<\/code> \u662f\u201c\u66fe\u4ea4\u7ed9\u6a21\u578b\u7684\u5019\u9009\u8d44\u6599\u201d\uff0c\u4e0d\u7b49\u4e8e\u6a21\u578b\u5728\u7b54\u6848\u4e2d\u5b9e\u9645\u5f15\u7528\u4e86\u6bcf\u4e00\u6761\uff1b\u5f53\u524d\u4ee3\u7801\u4e5f\u6ca1\u6709\u81ea\u52a8\u9a8c\u8bc1 <code>[\u6765\u6e90 N]<\/code> \u662f\u5426\u5b58\u5728\u6216\u662f\u5426\u652f\u6491\u76f8\u90bb\u7ed3\u8bba\u3002\u7528\u6237\u53ef\u4ee5\u624b\u52a8\u6838\u5bf9\uff0c\u751f\u4ea7\u7cfb\u7edf\u5219\u5e94\u628a\u5f15\u7528\u6709\u6548\u6027\u52a0\u5165\u540e\u6587\u7684\u81ea\u52a8\u8bc4\u4f30\u4e0e\u76d1\u63a7\u3002<\/p>\n<h2>\u7528\u81ea\u5df1\u7684\u8d44\u6599\u9a8c\u8bc1<\/h2>\n<p>\u5185\u7f6e\u6570\u636e\u53ea\u80fd\u8bc1\u660e\u7a0b\u5e8f\u53ef\u4ee5\u8fd0\u884c\uff0c\u65e0\u6cd5\u4ee3\u8868\u7528\u6237\u81ea\u5df1\u7684\u8d44\u6599\u4e5f\u80fd\u5f97\u5230\u540c\u6837\u6548\u679c\u3002\u56e0\u6b64\uff0c\u8fd8\u8981\u4e0a\u4f20\u81ea\u5df1\u7684\u6587\u6863\u6765\u68c0\u67e5\u5206\u5757\u548c\u68c0\u7d22\u7ed3\u679c\u3002\u9875\u9762\u652f\u6301 Markdown\u3001TXT\u3001\u6587\u672c\u578b PDF \u548c DOCX \u56db\u79cd\u683c\u5f0f\u3002<\/p>\n<p>\u670d\u52a1\u7aef\u89e3\u6790\u6587\u4ef6\u3001\u63d0\u53d6\u6587\u672c\u5e76\u5206\u5757\uff0c\u518d\u8c03\u7528\u672c\u5730 ONNX \u6a21\u578b\u751f\u6210\u5411\u91cf\u5e76\u5199\u5165\u7d22\u5f15\u3002PDF \u5fc5\u987b\u5e26\u6709\u53ef\u4ee5\u76f4\u63a5\u590d\u5236\u548c\u641c\u7d22\u7684\u6587\u672c\u5c42\uff1b\u53ea\u6709\u56fe\u7247\u7684\u626b\u63cf\u4ef6\u9700\u8981\u5148\u505a\u5149\u5b66\u5b57\u7b26\u8bc6\u522b\uff08Optical Character Recognition\uff0cOCR\uff09\uff0c\u628a\u56fe\u7247\u4e2d\u7684\u6587\u5b57\u8bc6\u522b\u51fa\u6765\u3002<\/p>\n<pre style=\"display:none;\"><code class=\"mermaid\">sequenceDiagram\n    actor U as \u7528\u6237\n    participant UI as \u7f51\u9875\n    participant S as \u5e94\u7528\u670d\u52a1\u7aef\n    participant M as \u672c\u5730 ONNX\n    participant I as \u672c\u5730\u7d22\u5f15\n\n    U-&gt;&gt;UI: \u9009\u62e9\u6587\u4ef6\n    UI-&gt;&gt;S: \u4e0a\u4f20\n    S-&gt;&gt;S: \u63d0\u53d6\u6587\u672c\u5e76\u5206\u5757\n    S-&gt;&gt;M: \u6279\u91cf\u751f\u6210\u5411\u91cf\n    M--&gt;&gt;S: 384 \u7ef4\u5411\u91cf\n    S-&gt;&gt;I: \u6587\u672c\u3001\u5411\u91cf\u4e0e\u6765\u6e90\n    I--&gt;&gt;UI: \u6587\u6863\u4e0e chunk \u7edf\u8ba1<\/code><\/pre>\n<h2>\u4ece\u56fa\u5b9a RAG \u5230\u6a21\u578b\u81ea\u4e3b\u68c0\u7d22<\/h2>\n<p>\u524d\u9762\u7684\u6d41\u7a0b\u7531\u5e94\u7528\u9884\u5148\u51b3\u5b9a\u68c0\u7d22\u7b56\u7565\uff1a\u6536\u5230\u95ee\u9898\u540e\u56fa\u5b9a\u6267\u884c\u6df7\u5408\u68c0\u7d22\uff0c\u53d6\u51fa Top-K\uff0c\u518d\u8c03\u7528 LLM\u3002\u68c0\u7d22\u5668\u3001\u53c2\u6570\u548c\u8c03\u7528\u6b21\u6570\u90fd\u5199\u5728\u5e94\u7528\u4ee3\u7801\u4e2d\uff0c\u8fd9\u91cc\u79f0\u4e3a\u56fa\u5b9a RAG\u3002<\/p>\n<p>\u6a21\u578b\u81ea\u4e3b\u68c0\u7d22\uff08Agentic Search\uff09\u628a\u90e8\u5206\u63a7\u5236\u6743\u4ea4\u7ed9\u6a21\u578b\u3002\u5728\u5b8c\u6574\u7684 Agent \u7cfb\u7edf\u4e2d\uff0c\u6a21\u578b\u53ef\u4ee5\u6839\u636e\u5f53\u524d\u7ed3\u679c\u51b3\u5b9a\u63a5\u4e0b\u6765\u4f7f\u7528\u54ea\u79cd\u641c\u7d22\u5de5\u5177\u3001\u641c\u7d22\u4ec0\u4e48\u5185\u5bb9\u4ee5\u53ca\u662f\u5426\u7ee7\u7eed\u3002\u8fd9\u91cc\u63cf\u8ff0\u7684\u662f\u4e00\u79cd\u901a\u7528\u67b6\u6784\uff0c\u4e0d\u4ee3\u8868\u5f53\u524d Demo \u5df2\u7ecf\u5b9e\u73b0\u4e86\u6240\u6709\u8fd9\u4e9b\u80fd\u529b\u3002<\/p>\n<pre style=\"display:none;\"><code class=\"mermaid\">flowchart LR\n    subgraph Classic[&quot;\u56fa\u5b9a RAG&quot;]\n        Q1[&quot;\u95ee\u9898&quot;] --&gt; R1[&quot;\u56fa\u5b9a\u68c0\u7d22\u5668&quot;]\n        R1 --&gt; K1[&quot;Top-K&quot;]\n        K1 --&gt; L1[&quot;LLM \u751f\u6210&quot;]\n    end\n\n    subgraph Agentic[&quot;\u901a\u7528\u7684\u591a\u5de5\u5177 Agentic Search&quot;]\n        Q2[&quot;\u95ee\u9898&quot;] --&gt; A[&quot;LLM Agent&quot;]\n        A --&gt; G1[&quot;Grep \/ SQL \/ Vector Search&quot;]\n        G1 --&gt; G2[&quot;\u8bfb\u53d6\u7ed3\u679c&quot;]\n        G2 --&gt; A\n        A --&gt; O[&quot;\u56de\u7b54&quot;]\n    end<\/code><\/pre>\n<p>\u4e24\u79cd\u65b9\u5f0f\u90fd\u4f1a\u7528\u5916\u90e8\u8d44\u6599\u589e\u5f3a\u751f\u6210\uff0c\u533a\u522b\u5728\u4e8e\u8c01\u51b3\u5b9a\u68c0\u7d22\u6b65\u9aa4\u3002\u76f8\u5173\u672f\u8bed\u5e76\u672a\u5b8c\u5168\u7edf\u4e00\uff0c\u540c\u4e00\u79cd\u7ed3\u6784\u4e5f\u53ef\u80fd\u88ab\u79f0\u4e3a Agentic RAG \u6216 Agentic Retrieval\uff1b\u672c\u6587\u53ea\u7528\u201c\u56fa\u5b9a RAG\u201d\u548c\u201cAgentic Search\u201d\u533a\u5206\u63a7\u5236\u6d41\u3002<\/p>\n<p>\u5f53\u524d Demo \u7684\u80fd\u529b\u8981\u7b80\u5355\u5f97\u591a\u3002<code>\/api\/ask<\/code> \u5148\u7528\u56fa\u5b9a\u53c2\u6570\u6267\u884c\u6df7\u5408\u68c0\u7d22\uff0c\u518d\u628a\u7ed3\u679c\u4ea4\u7ed9 LLM\uff1b\u540c\u65f6\uff0c\u5b83\u628a\u540c\u4e00\u5957\u6df7\u5408\u68c0\u7d22\u5305\u88c5\u6210\u4e00\u4e2a\u540d\u4e3a <code>search_knowledge_base<\/code> \u7684\u5de5\u5177\u3002LLM \u5982\u679c\u8ba4\u4e3a\u521d\u59cb\u8d44\u6599\u4e0d\u591f\uff0c\u53ef\u4ee5\u6362\u4e00\u79cd\u95ee\u6cd5\u8c03\u7528\u8fd9\u4e2a\u5de5\u5177\uff0c\u670d\u52a1\u7aef\u518d\u8865\u5145\u4e00\u6279\u7247\u6bb5\u3002<\/p>\n<p>\u56e0\u6b64\uff0c\u5f53\u524d Demo \u53ea\u80fd\u8ba9\u6a21\u578b\u51b3\u5b9a\u201c\u662f\u5426\u7528\u65b0\u7684\u67e5\u8be2\u518d\u6b21\u641c\u7d22\u77e5\u8bc6\u5e93\u201d\u3002\u5b83\u4e0d\u80fd\u5728 SQL\u3001grep\u3001\u6587\u4ef6\u8bfb\u53d6\u548c\u5411\u91cf\u641c\u7d22\u4e4b\u95f4\u81ea\u4e3b\u9009\u62e9\uff0c\u4e5f\u6ca1\u6709\u5b9e\u73b0\u4ee3\u7801\u8c03\u7528\u94fe\u5206\u6790\u3002\u9875\u9762\u4e0a\u7684\u201cAgent \u6b65\u9aa4\u6570\u201d\u548c\u201c\u68c0\u7d22\u6b21\u6570\u201d\u53ea\u662f\u7528\u6765\u663e\u793a\u8fd9\u4e00\u8f6e\u5b9e\u9645\u8c03\u7528\u4e86\u51e0\u6b21 LLM \u548c\u77e5\u8bc6\u5e93\u641c\u7d22\u3002<\/p>\n<h2>\u4ec0\u4e48\u65f6\u5019\u9009\u62e9\u54ea\u79cd\u68c0\u7d22\u65b9\u6848<\/h2>\n<p>\u53ef\u4ee5\u76f4\u63a5\u6839\u636e\u8d44\u6599\u7c7b\u578b\u548c\u95ee\u9898\u7279\u70b9\u9009\u62e9\uff1a<\/p>\n<table>\n<thead>\n<tr>\n<th>\u8981\u89e3\u51b3\u7684\u95ee\u9898<\/th>\n<th>\u4f18\u5148\u8003\u8651\u7684\u65b9\u5f0f<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>\u6309\u8ba2\u5355\u72b6\u6001\u3001\u65e5\u671f\u3001\u7528\u6237 ID \u7b49\u5b57\u6bb5\u7b5b\u9009\u6570\u636e<\/td>\n<td>SQL<\/td>\n<\/tr>\n<tr>\n<td>\u67e5\u627e\u9519\u8bef\u7801\u3001\u5408\u540c\u53f7\u3001\u51fd\u6570\u540d\u7b49\u51c6\u786e\u5b57\u7b26\u4e32<\/td>\n<td>BM25\u3001grep \u6216\u5176\u4ed6\u5173\u952e\u8bcd\u68c0\u7d22<\/td>\n<\/tr>\n<tr>\n<td>\u7528\u6237\u7528\u81ea\u7136\u8bed\u8a00\u63d0\u95ee\uff0c\u539f\u6587\u53ef\u80fd\u4f7f\u7528\u4e0d\u540c\u8bf4\u6cd5<\/td>\n<td>\u5411\u91cf\u68c0\u7d22<\/td>\n<\/tr>\n<tr>\n<td>\u65e2\u8981\u7406\u89e3\u81ea\u7136\u8bed\u8a00\uff0c\u53c8\u8981\u5339\u914d\u51c6\u786e\u7f16\u53f7<\/td>\n<td>\u5411\u91cf\u68c0\u7d22\u4e0e BM25 \u7ec4\u5408<\/td>\n<\/tr>\n<tr>\n<td>\u5fc5\u987b\u6839\u636e\u7b2c\u4e00\u6b21\u7ed3\u679c\u624d\u80fd\u51b3\u5b9a\u4e0b\u4e00\u6b65\u67e5\u4ec0\u4e48<\/td>\n<td>\u5728\u66f4\u5b8c\u6574\u7684\u7cfb\u7edf\u4e2d\u8003\u8651 Agent \u5de5\u5177\uff1b\u5f53\u524d Demo \u53ea\u652f\u6301\u518d\u6b21\u8c03\u7528\u540c\u4e00\u4e2a\u6df7\u5408\u68c0\u7d22\u5de5\u5177<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>\u5bf9\u5f53\u524d Demo \u8fd9\u7c7b\u5236\u5ea6\u95ee\u7b54\uff0c\u901a\u5e38\u5148\u53d6\u4e00\u6b21 Top-K \u5c31\u80fd\u751f\u6210\u7b54\u6848\uff0c\u56fa\u5b9a\u6df7\u5408\u68c0\u7d22\u5e94\u5f53\u662f\u4e3b\u8981\u6d41\u7a0b\u3002\u989d\u5916\u641c\u7d22\u53ea\u7528\u4e8e\u521d\u59cb\u7247\u6bb5\u786e\u5b9e\u7f3a\u5c11\u7b54\u6848\u7684\u60c5\u51b5\uff0c\u4f8b\u5982\u4e00\u4e2a\u95ee\u9898\u540c\u65f6\u8be2\u95ee\u62a5\u9500\u9650\u989d\u548c\u9519\u8bef\u7801\u542b\u4e49\uff0c\u800c\u521d\u59cb Top-K \u53ea\u8986\u76d6\u4e86\u5176\u4e2d\u4e00\u90e8\u5206\u3002\u6b64\u65f6\u6a21\u578b\u53ef\u4ee5\u628a\u7f3a\u5c11\u7684\u90e8\u5206\u6539\u5199\u6210\u65b0\u67e5\u8be2\uff0c\u518d\u8c03\u7528\u540c\u4e00\u4e2a\u6df7\u5408\u68c0\u7d22\u5de5\u5177\u3002<\/p>\n<p>\u5982\u679c\u5b9e\u9645\u4e1a\u52a1\u8fd8\u9700\u8981\u67e5\u8be2\u6570\u636e\u5e93\u3001\u641c\u7d22\u4ee3\u7801\u6216\u8bfb\u53d6\u6307\u5b9a\u6587\u4ef6\uff0c\u53ef\u4ee5\u5728\u66f4\u5b8c\u6574\u7684\u7cfb\u7edf\u4e2d\u628a SQL\u3001grep \u548c\u6587\u4ef6\u8bfb\u53d6\u5206\u522b\u505a\u6210\u5de5\u5177\uff0c\u8ba9\u6a21\u578b\u6839\u636e\u4e0a\u4e00\u6b65\u7ed3\u679c\u51b3\u5b9a\u4e0b\u4e00\u6b65\u64cd\u4f5c\u3002\u8fd9\u5df2\u7ecf\u8d85\u51fa\u5f53\u524d Demo \u7684\u80fd\u529b\u8303\u56f4\u3002\u591a\u6b65\u5de5\u5177\u8c03\u7528\u4f1a\u589e\u52a0\u54cd\u5e94\u65f6\u95f4\u548c\u8d39\u7528\uff0c\u56e0\u6b64\u53ea\u6709\u4e00\u6b21\u56fa\u5b9a\u68c0\u7d22\u65e0\u6cd5\u5b8c\u6210\u4efb\u52a1\u65f6\u624d\u9700\u8981\u91c7\u7528\u3002<\/p>\n<h3>\u4ee3\u7801 Agent \u4e3a\u4ec0\u4e48\u91c7\u7528\u4e0d\u540c\u7684\u641c\u7d22\u65b9\u6848<\/h3>\n<p>\u4e0b\u9762\u4e09\u4e2a\u4f8b\u5b50\u7684\u4f5c\u7528\u4e0d\u662f\u6bd4\u8f83\u4ea7\u54c1\u597d\u574f\uff0c\u800c\u662f\u8bf4\u660e\u4ee3\u7801\u641c\u7d22\u6ca1\u6709\u552f\u4e00\u65b9\u6848\u3002\u4ee3\u7801\u5e93\u662f\u5426\u7ecf\u5e38\u53d8\u5316\u3001\u89c4\u6a21\u6709\u591a\u5927\u3001\u80fd\u5426\u5efa\u7acb\u7d22\u5f15\uff0c\u4ee5\u53ca\u7528\u6237\u662f\u5728\u67e5\u51c6\u786e\u540d\u79f0\u8fd8\u662f\u81ea\u7136\u8bed\u8a00\u6982\u5ff5\uff0c\u90fd\u4f1a\u5f71\u54cd\u6280\u672f\u9009\u62e9\u3002<\/p>\n<table>\n<thead>\n<tr>\n<th>\u4ea7\u54c1<\/th>\n<th>\u516c\u5f00\u8d44\u6599\u63cf\u8ff0\u7684\u505a\u6cd5<\/th>\n<th>\u5bf9\u68c0\u7d22\u9009\u578b\u7684\u542f\u793a<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>Claude Code<\/td>\n<td>\u4e3b\u8981\u8ba9\u6a21\u578b\u8c03\u7528 Glob\u3001Grep\u3001Read \u7b49\u5de5\u5177\u76f4\u63a5\u67e5\u770b\u5f53\u524d\u4ee3\u7801\uff1b\u65e9\u671f\u5c1d\u8bd5\u8fc7\u8bed\u4e49\u7d22\u5f15\uff0c\u540e\u6765\u6ca1\u6709\u7ee7\u7eed\u91c7\u7528<\/td>\n<td>\u4ee3\u7801\u53d8\u5316\u9891\u7e41\u65f6\uff0c\u76f4\u63a5\u641c\u7d22\u5f53\u524d\u6587\u4ef6\u53ef\u4ee5\u7701\u53bb\u7d22\u5f15\u540c\u6b65\uff0c\u4f46\u6a21\u578b\u53ef\u80fd\u9700\u8981\u591a\u8f6e\u67e5\u627e<\/td>\n<\/tr>\n<tr>\n<td>Cursor<\/td>\n<td>\u540c\u65f6\u4f7f\u7528 grep \u548c\u8bed\u4e49\u641c\u7d22\uff0c\u5e76\u589e\u91cf\u66f4\u65b0\u4ee3\u7801\u7d22\u5f15<\/td>\n<td>\u7cbe\u786e\u540d\u79f0\u4ea4\u7ed9 grep\uff0c\u81ea\u7136\u8bed\u8a00\u95ee\u9898\u4ea4\u7ed9\u8bed\u4e49\u641c\u7d22\uff0c\u4e24\u79cd\u65b9\u5f0f\u53ef\u4ee5\u4e92\u8865<\/td>\n<\/tr>\n<tr>\n<td>GitHub Copilot<\/td>\n<td>\u5728\u652f\u6301\u7684\u4ed3\u5e93\u4e0a\u4e0b\u6587\u4e2d\u5efa\u7acb\u8bed\u4e49\u4ee3\u7801\u641c\u7d22\u7d22\u5f15\uff0c\u4e5f\u4f1a\u6309\u9700\u6267\u884c\u8bed\u4e49\u641c\u7d22<\/td>\n<td>\u662f\u5426\u9884\u5148\u5efa\u7d22\u5f15\u53d6\u51b3\u4e8e\u4ed3\u5e93\u73af\u5883\u3001\u5e73\u53f0\u80fd\u529b\u548c\u7ec4\u7ec7\u7b56\u7565<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>\u8fd9\u4e9b\u516c\u5f00\u8d44\u6599\u6765\u81ea\u4e0d\u540c\u65f6\u95f4\u3001\u4e0d\u540c\u4ea7\u54c1\u548c\u4e0d\u540c\u6d4b\u8bd5\u6761\u4ef6\uff0c\u4e0d\u80fd\u7528\u6765\u5f97\u51fa\u8c01\u7684\u641c\u7d22\u6548\u679c\u66f4\u597d\u3002\u5b83\u4eec\u771f\u6b63\u8bf4\u660e\u7684\u662f\uff1a\u9879\u76ee\u5e94\u8be5\u6839\u636e\u81ea\u5df1\u7684\u4ee3\u7801\u89c4\u6a21\u3001\u66f4\u65b0\u9891\u7387\u3001\u6570\u636e\u8fb9\u754c\u548c\u6d4b\u8bd5\u7ed3\u679c\u9009\u62e9\u76f4\u63a5\u641c\u7d22\u3001\u8bed\u4e49\u7d22\u5f15\u6216\u4e24\u8005\u7ec4\u5408\u3002<\/p>\n<h2>\u600e\u6837\u5224\u65ad RAG \u662f\u5426\u6709\u6548<\/h2>\n<p>\u9875\u9762\u80fd\u591f\u8fd4\u56de\u7b54\u6848\uff0c\u53ea\u80fd\u8bf4\u660e\u7a0b\u5e8f\u5df2\u7ecf\u8fd0\u884c\u8d77\u6765\uff0c\u4e0d\u80fd\u8bf4\u660e\u7b54\u6848\u53ef\u9760\u3002\u5373\u4f7f\u63a5\u53e3\u72b6\u6001\u662f 200\u3001\u76f8\u4f3c\u5ea6\u662f 0.86\u3001\u7b54\u6848\u540e\u9762\u5199\u7740 <code>[\u6765\u6e90 1]<\/code>\uff0c\u4ecd\u7136\u53ef\u80fd\u53d1\u751f\u4ee5\u4e0b\u95ee\u9898\uff1a<\/p>\n<ul>\n<li>\u627e\u5230\u7684\u662f\u4e0d\u76f8\u5173\u7684\u8d44\u6599\u3002<\/li>\n<li>\u627e\u5230\u4e86\u6b63\u786e\u8d44\u6599\uff0c\u4f46\u628a\u5173\u952e\u6761\u4ef6\u6f0f\u6389\u4e86\u3002<\/li>\n<li>\u8d44\u6599\u91cc\u6ca1\u6709\u7b54\u6848\uff0c\u6a21\u578b\u5374\u751f\u6210\u4e86\u8d44\u6599\u4e4b\u5916\u7684\u5185\u5bb9\u3002<\/li>\n<li>\u7b54\u6848\u5185\u5bb9\u53ef\u80fd\u6b63\u786e\uff0c\u4f46\u6807\u6ce8\u7684\u6765\u6e90\u4e0d\u80fd\u652f\u6301\u5b83\u3002<\/li>\n<\/ul>\n<p>\u8bc4\u4f30\u65f6\u9700\u8981\u5206\u5f00\u68c0\u67e5\u4e24\u4e2a\u73af\u8282\uff1a<\/p>\n<ol>\n<li><strong>\u68c0\u7d22<\/strong>\uff1a\u6b63\u786e\u7247\u6bb5\u662f\u5426\u8fdb\u5165 Top-K\uff0c\u6392\u540d\u662f\u5426\u8db3\u591f\u9760\u524d\u3002<\/li>\n<li><strong>\u751f\u6210<\/strong>\uff1aLLM \u662f\u5426\u6839\u636e\u8fd9\u4e9b\u7247\u6bb5\u56de\u7b54\uff0c\u662f\u5426\u9057\u6f0f\u6761\u4ef6\u3001\u6dfb\u52a0\u4e86\u8d44\u6599\u4e2d\u6ca1\u6709\u7684\u5185\u5bb9\uff0c\u5f15\u7528\u80fd\u5426\u652f\u6491\u7ed3\u8bba\u3002<\/li>\n<\/ol>\n<p>\u68c0\u7d22\u5931\u8d25\u548c\u751f\u6210\u5931\u8d25\u7684\u4fee\u590d\u4f4d\u7f6e\u4e0d\u540c\u3002\u6b63\u786e\u7247\u6bb5\u6ca1\u6709\u8fdb\u5165 Top-K \u65f6\uff0c\u5e94\u68c0\u67e5\u5206\u5757\u3001Embedding\u3001\u5173\u952e\u8bcd\u68c0\u7d22\u548c\u6392\u5e8f\uff1b\u6b63\u786e\u7247\u6bb5\u5df2\u7ecf\u4ea4\u7ed9 LLM \u800c\u7b54\u6848\u4ecd\u7136\u51fa\u9519\u65f6\uff0c\u518d\u68c0\u67e5\u63d0\u793a\u8bcd\u3001\u6a21\u578b\u548c\u5f15\u7528\u5904\u7406\u3002<\/p>\n<h3>\u51c6\u5907\u5e26\u6807\u51c6\u7b54\u6848\u7684\u6d4b\u8bd5\u9898<\/h3>\n<p>\u8bc4\u4f30\u96c6\uff08eval set\uff0c\u4e5f\u53eb golden set\uff09\u7531\u6d4b\u8bd5\u95ee\u9898\u3001\u6b63\u786e\u6765\u6e90\u3001\u7b54\u6848\u8981\u70b9\u548c\u53ef\u56de\u7b54\u6027\u7ec4\u6210\u3002\u6bcf\u9053\u9898\u90fd\u8981\u63d0\u524d\u5199\u660e\u5224\u5b9a\u6807\u51c6\uff0c\u5426\u5219\u4e0d\u540c\u7248\u672c\u4e4b\u95f4\u65e0\u6cd5\u6309\u540c\u4e00\u89c4\u5219\u6bd4\u8f83\u3002<\/p>\n<p>\u672c\u6587\u540e\u9762\u7684\u5b9e\u6d4b\u53ea\u4f7f\u7528\u9879\u76ee\u5185\u7f6e\u8d44\u6599\uff0c\u4e0d\u5305\u542b\u7528\u6237\u4e0a\u4f20\u7684\u6587\u6863\u3002\u8fd9\u4e9b\u8d44\u6599\u4e0d\u662f\u9879\u76ee\u76ee\u5f55\u91cc\u7684\u72ec\u7acb\u6587\u4ef6\uff0c\u800c\u662f\u5f00\u53d1\u8005\u76f4\u63a5\u5199\u5728 <code>packages\/server\/src\/seedDocuments.ts<\/code> \u4e2d\u7684 4 \u6bb5\u793a\u4f8b\u6587\u672c\u3002\u8fd0\u884c <code>npm run ingest<\/code> \u65f6\uff0c\u7a0b\u5e8f\u4f1a\u81ea\u52a8\u628a\u5b83\u4eec\u52a0\u5165\u7d22\u5f15\u3002\u661f\u6cb3\u79d1\u6280\u4e5f\u662f\u865a\u6784\u540d\u79f0\uff0c\u5b83\u7684\u5dee\u65c5\u5236\u5ea6\u53ea\u7528\u4e8e\u6f14\u793a\u95ee\u7b54\u3002<\/p>\n<p><img decoding=\"async\"   referrerpolicy=\"no-referrer\" src=\"https:\/\/segmentfault.com\/img\/remote\/1460000048055390\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" title=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" loading=\"lazy\"><\/p>\n<p>\u9875\u9762\u4e0a\u7684\u4e09\u9053\u53ef\u56de\u7b54\u793a\u4f8b\u9898\u90fd\u5728\u8be2\u95ee\u661f\u6cb3\u79d1\u6280\u7684\u62a5\u9500\u5236\u5ea6\uff0c\u6240\u4ee5\u4f7f\u7528\u9875\u9762\u65f6\u6700\u5bb9\u6613\u6ce8\u610f\u5230\u8fd9\u6bb5\u8d44\u6599\u3002\u53e6\u5916\u4e09\u6bb5\u4e5f\u7531\u7a0b\u5e8f\u81ea\u52a8\u52a0\u5165\uff0c\u53ea\u662f\u5206\u522b\u7528\u4e8e\u56de\u7b54 RAG \u539f\u7406\u3001\u68c0\u7d22\u65b9\u5f0f\u548c\u6a21\u578b\u914d\u7f6e\u65b9\u9762\u7684\u95ee\u9898\u3002\u56db\u6bb5\u6587\u672c\u7684\u7528\u9014\u5982\u4e0b\uff1a<\/p>\n<table>\n<thead>\n<tr>\n<th>\u4ee3\u7801\u4e2d\u9884\u7f6e\u7684\u793a\u4f8b\u6587\u672c<\/th>\n<th>\u4e3a\u4ec0\u4e48\u653e\u8fdb Demo<\/th>\n<th>\u5305\u542b\u7684\u5c0f\u6807\u9898<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>RAG \u57fa\u7840\u77e5\u8bc6\uff08\u5185\u7f6e\uff09<\/td>\n<td>\u6d4b\u8bd5\u80fd\u5426\u56de\u7b54\u201cRAG \u662f\u4ec0\u4e48\u201d\u7b49\u95ee\u9898<\/td>\n<td>\u5b9a\u4e49\u3001\u4e09\u4e2a\u9636\u6bb5\u3001\u6570\u636e\u66f4\u65b0\u3001\u5411\u91cf\u6570\u636e\u5e93<\/td>\n<\/tr>\n<tr>\n<td>\u6df7\u5408\u68c0\u7d22\u8bf4\u660e\uff08\u5185\u7f6e\uff09<\/td>\n<td>\u6d4b\u8bd5\u80fd\u5426\u56de\u7b54\u5411\u91cf\u68c0\u7d22\u548c BM25 \u65b9\u9762\u7684\u95ee\u9898<\/td>\n<td>\u5411\u91cf\u68c0\u7d22\u3001BM25\u3001\u52a0\u6743\u878d\u5408<\/td>\n<\/tr>\n<tr>\n<td>\u661f\u6cb3\u79d1\u6280\u5dee\u65c5\u4e0e\u4ea4\u901a\u62a5\u9500\u5236\u5ea6<\/td>\n<td>\u4e3a\u9875\u9762\u4e0a\u7684\u62a5\u9500\u793a\u4f8b\u95ee\u9898\u63d0\u4f9b\u7b54\u6848<\/td>\n<td>\u5e02\u5185\u4ea4\u901a\u8d39\u7528\u3001\u7968\u636e\u671f\u9650\u3001\u4f4f\u5bbf\u6807\u51c6\u3001\u62a5\u9500\u5165\u53e3<\/td>\n<\/tr>\n<tr>\n<td>Demo \u751f\u6210\u6a21\u578b\u914d\u7f6e\uff08\u5185\u7f6e\uff09<\/td>\n<td>\u8bf4\u660e Demo \u600e\u6837\u8c03\u7528\u751f\u6210\u6a21\u578b\u4ee5\u53ca\u600e\u6837\u4fdd\u5b58 API Key<\/td>\n<td>\u5de5\u4f5c\u65b9\u5f0f\u3001\u652f\u6301\u7684\u5382\u5546\u3001API Key<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>\u5206\u5757\u65f6\uff0c\u6bcf\u4e2a\u5c0f\u6807\u9898\u548c\u5b83\u4e0b\u9762\u7684\u6b63\u6587\u4f1a\u6210\u4e3a\u4e00\u4e2a\u53ef\u4ee5\u5355\u72ec\u641c\u7d22\u7684\u7247\u6bb5\uff0c\u4e5f\u5c31\u662f\u4e00\u4e2a chunk\u3002\u56db\u6bb5\u6587\u672c\u5206\u522b\u4ea7\u751f 4\u30013\u30014\u30013 \u4e2a chunk\uff0c\u52a0\u8d77\u6765\u662f <code>4 + 3 + 4 + 3 = 14<\/code>\u3002\u56e0\u6b64\uff0c\u201c14 \u4e2a\u5185\u7f6e chunk\u201d\u6307 4 \u6bb5\u793a\u4f8b\u6587\u672c\u88ab\u62c6\u6210\u7684 14 \u4e2a\u53ef\u641c\u7d22\u7247\u6bb5\u3002<\/p>\n<p><code>npm run ingest<\/code> \u53ea\u4f1a\u91cd\u65b0\u751f\u6210\u8fd9\u4e9b\u5185\u7f6e\u7247\u6bb5\uff0c\u4e0d\u4f1a\u5220\u9664\u7528\u6237\u5df2\u7ecf\u4e0a\u4f20\u7684\u6587\u6863\uff0c\u6240\u4ee5 14 \u662f\u5185\u7f6e\u7247\u6bb5\u7684\u6570\u91cf\uff0c\u4e0d\u4e00\u5b9a\u662f\u5f53\u524d\u7d22\u5f15\u7684\u603b\u6570\u3002\u9875\u9762\u9884\u8bbe\u4e86 4 \u4e2a\u793a\u4f8b\u95ee\u9898\uff0c\u524d\u4e09\u9053\u80fd\u4ece\u5dee\u65c5\u5236\u5ea6\u4e2d\u627e\u5230\u7b54\u6848\uff0c\u7b2c\u56db\u9053\u7528\u4e8e\u68c0\u67e5\u8d44\u6599\u4e0d\u8db3\u65f6\u80fd\u5426\u62d2\u7b54\uff1a<\/p>\n<table>\n<thead>\n<tr>\n<th>\u6d4b\u8bd5\u95ee\u9898<\/th>\n<th>\u77e5\u8bc6\u5e93\u80fd\u5426\u56de\u7b54<\/th>\n<th>\u5e94\u8be5\u627e\u5230\u54ea\u4efd\u8d44\u6599<\/th>\n<th>\u7b54\u6848\u5fc5\u987b\u5305\u542b\u4ec0\u4e48<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>\u6253\u8f66\u8d39\u8d85\u8fc7\u516b\u767e\u5757\u5e94\u8be5\u600e\u4e48\u529e\uff1f<\/td>\n<td>\u80fd<\/td>\n<td>\u5dee\u65c5\u5236\u5ea6 &gt; \u5e02\u5185\u4ea4\u901a\u8d39\u7528<\/td>\n<td>800 \u5143\u3001\u90e8\u95e8\u8d1f\u8d23\u4eba\u5ba1\u6279\u3001\u8bf4\u660e\u4e1a\u52a1\u539f\u56e0<\/td>\n<\/tr>\n<tr>\n<td><code>ERR_FIN_403<\/code> \u662f\u4ec0\u4e48\u610f\u601d\uff1f<\/td>\n<td>\u80fd<\/td>\n<td>\u5dee\u65c5\u5236\u5ea6 &gt; \u62a5\u9500\u5165\u53e3<\/td>\n<td>\u6ca1\u6709\u62a5\u9500\u6743\u9650\u3001\u8054\u7cfb\u90e8\u95e8\u7ba1\u7406\u5458<\/td>\n<\/tr>\n<tr>\n<td>\u4e0a\u6d77\u4f4f\u5bbf\u4e00\u665a\u6700\u591a\u62a5\u591a\u5c11\uff1f<\/td>\n<td>\u80fd<\/td>\n<td>\u5dee\u65c5\u5236\u5ea6 &gt; \u4f4f\u5bbf\u6807\u51c6<\/td>\n<td>600 \u5143\/\u665a\u3001\u8d85\u6807\u524d\u7f6e\u7533\u8bf7<\/td>\n<\/tr>\n<tr>\n<td>\u516c\u53f8\u7684\u5e74\u5047\u6709\u591a\u5c11\u5929\uff1f<\/td>\n<td>\u5426<\/td>\n<td>\u65e0\u5bf9\u5e94\u6765\u6e90<\/td>\n<td>\u8bf4\u660e\u8d44\u6599\u7f3a\u5c11\u5e74\u5047\u5929\u6570<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>\u524d\u4e09\u9053\u9898\u7528\u4e8e\u8ba1\u7b97\u540e\u6587\u7684\u68c0\u7d22\u547d\u4e2d\u6307\u6807\uff1b\u201c\u516c\u53f8\u7684\u5e74\u5047\u6709\u591a\u5c11\u5929\uff1f\u201d\u7528\u4e8e\u5355\u72ec\u68c0\u67e5\u62d2\u7b54\u3002<\/p>\n<p>\u6b63\u5f0f\u8bc4\u4f30\u8fd8\u5e94\u8986\u76d6\u9519\u522b\u5b57\u3001\u7b80\u79f0\u3001\u4e0d\u540c\u8868\u8ff0\u3001\u8fc7\u671f\u8d44\u6599\u3001\u76f8\u4e92\u51b2\u7a81\u7684\u8d44\u6599\u548c\u65e0\u6743\u67e5\u770b\u7684\u8d44\u6599\u3002\u4f8b\u5982\uff0c\u201c\u4e0a\u6d77\u9152\u5e97\u80fd\u62a5\u51e0\u94b1\u201d\u201c\u9b54\u90fd\u4f4f\u5bbf\u6807\u51c6\u201d\u548c\u201c\u4e0a\u6d77\u51fa\u5dee\u4f4f\u5e97\u4e0a\u9650\u201d\u53ef\u80fd\u90fd\u5728\u95ee\u540c\u4e00\u4ef6\u4e8b\u3002<\/p>\n<p>\u5f53\u524d Demo \u53ea\u6709 4 \u9053\u793a\u4f8b\u9898\uff0c\u6570\u91cf\u592a\u5c11\uff0c\u4e0d\u9002\u5408\u62c6\u6210\u4e24\u7ec4\u3002\u6b63\u5f0f\u9879\u76ee\u79ef\u7d2f\u4e86\u8db3\u591f\u591a\u7684\u6d4b\u8bd5\u9898\u540e\uff0c\u53ef\u4ee5\u8fd9\u6837\u505a\uff1a\u5047\u8bbe\u5171\u6709 100 \u9053\u9898\uff0c\u5148\u7528\u5176\u4e2d 80 \u9053\u53cd\u590d\u8c03\u6574\u5206\u5757\u65b9\u5f0f\u548c\u68c0\u7d22\u6743\u91cd\uff1b\u5269\u4e0b 20 \u9053\u5728\u8c03\u53c2\u671f\u95f4\u4e0d\u7528\uff0c\u7b49\u65b9\u6848\u786e\u5b9a\u540e\u518d\u505a\u6700\u7ec8\u9a8c\u6536\u3002<\/p>\n<p>\u5982\u679c\u5de5\u7a0b\u5e08\u770b\u7740\u5168\u90e8 100 \u9053\u9898\u7684\u6210\u7ee9\u53cd\u590d\u6539\u53c2\u6570\uff0c\u6700\u540e\u9009\u51fa\u7684\u53c2\u6570\u53ef\u80fd\u53ea\u9002\u5408\u8fd9\u4e9b\u5df2\u7ecf\u89c1\u8fc7\u7684\u9898\u3002\u6362\u6210\u65b0\u7684\u7528\u6237\u95ee\u6cd5\uff0c\u68c0\u7d22\u6548\u679c\u4ecd\u7136\u53ef\u80fd\u5f88\u5dee\u3002\u7559\u51fa 20 \u9053\u9a8c\u6536\u9898\uff0c\u5c31\u662f\u4e3a\u4e86\u68c0\u67e5\u53c2\u6570\u80fd\u5426\u5904\u7406\u6ca1\u6709\u53c2\u4e0e\u8c03\u4f18\u7684\u95ee\u9898\u3002<\/p>\n<p>\u6d4b\u8bd5\u9898\u8fd8\u5e94\u5355\u72ec\u4fdd\u5b58\u5728\u6587\u4ef6\u4e2d\uff0c\u4f8b\u5982 <code>eval-v1.json<\/code>\u3002\u6587\u4ef6\u91cc\u8bb0\u5f55\u6bcf\u9053\u9898\u7684\u6b63\u786e\u6765\u6e90\u3001\u7b54\u6848\u8981\u70b9\u548c\u53ef\u56de\u7b54\u6027\uff1b\u589e\u52a0\u6216\u4fee\u6539\u9898\u76ee\u65f6\u518d\u751f\u6210 <code>eval-v2.json<\/code>\u3002\u8fd9\u6837\u53ef\u4ee5\u786e\u8ba4\u4e24\u6b21\u6d4b\u8bd5\u662f\u5426\u4f7f\u7528\u4e86\u540c\u4e00\u6279\u9898\u3002\u6b63\u786e\u6765\u6e90\u6700\u597d\u8bb0\u5f55\u4e3a\u201c\u6587\u6863\u540d\u79f0 + \u5c0f\u6807\u9898\u6216\u4e00\u5c0f\u6bb5\u539f\u6587\u201d\uff0c\u4e0d\u8981\u53ea\u8bb0\u5f55 chunk ID\uff0c\u56e0\u4e3a\u91cd\u65b0\u5206\u5757\u540e ID \u53ef\u80fd\u6539\u53d8\u3002<\/p>\n<h3>\u5148\u68c0\u67e5\u641c\u7d22\u6709\u6ca1\u6709\u627e\u5230\u6b63\u786e\u8d44\u6599<\/h3>\n<p>RAG \u56de\u7b54\u9519\u8bef\uff0c\u53ef\u80fd\u662f\u641c\u7d22\u6ca1\u6709\u627e\u5230\u6b63\u786e\u8d44\u6599\uff0c\u4e5f\u53ef\u80fd\u662f LLM \u6ca1\u6709\u6b63\u786e\u4f7f\u7528\u8d44\u6599\u3002\u4e3a\u4e86\u5148\u6392\u9664\u7b2c\u4e00\u79cd\u95ee\u9898\uff0c\u8fd9\u4e00\u6b65\u6682\u65f6\u4e0d\u770b LLM \u7684\u56de\u7b54\uff0c\u53ea\u68c0\u67e5\u641c\u7d22\u7ed3\u679c\u3002<\/p>\n<p>\u6d4b\u8bd5\u524d\u8981\u4e3a\u6bcf\u9053\u9898\u6307\u5b9a\u4e00\u4e2a\u201c\u5fc5\u987b\u627e\u5230\u7684\u7247\u6bb5\u201d\uff1a<\/p>\n<table>\n<thead>\n<tr>\n<th>\u95ee\u9898<\/th>\n<th>\u5fc5\u987b\u627e\u5230\u7684\u7247\u6bb5<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>\u6253\u8f66\u8d39\u8d85\u8fc7\u516b\u767e\u5757\u5e94\u8be5\u600e\u4e48\u529e\uff1f<\/td>\n<td>\u5e02\u5185\u4ea4\u901a\u8d39\u7528<\/td>\n<\/tr>\n<tr>\n<td><code>ERR_FIN_403<\/code> \u662f\u4ec0\u4e48\u610f\u601d\uff1f<\/td>\n<td>\u62a5\u9500\u5165\u53e3<\/td>\n<\/tr>\n<tr>\n<td>\u4e0a\u6d77\u4f4f\u5bbf\u4e00\u665a\u6700\u591a\u62a5\u591a\u5c11\uff1f<\/td>\n<td>\u4f4f\u5bbf\u6807\u51c6<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>\u6d4b\u8bd5\u65f6\u53ea\u4fdd\u7559 14 \u4e2a\u5185\u7f6e chunk\uff0c\u4e0d\u52a0\u5165\u7528\u6237\u4e0a\u4f20\u7684\u6587\u6863\uff0c\u5e76\u56fa\u5b9a <code>Top-K = 5<\/code>\u3001\u5411\u91cf\/BM25 \u6743\u91cd\u4e3a <code>0.7\/0.3<\/code>\u3002\u7a0b\u5e8f\u5206\u522b\u641c\u7d22\u8fd9\u4e09\u9053\u9898\uff0c\u7ed3\u679c\u5982\u4e0b\uff1a<\/p>\n<table>\n<thead>\n<tr>\n<th>\u95ee\u9898<\/th>\n<th>\u6392\u5728\u524d\u9762\u7684\u641c\u7d22\u7ed3\u679c<\/th>\n<th align=\"right\">\u6b63\u786e chunk \u7684\u540d\u6b21<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>\u6253\u8f66\u8d39\u8d85\u8fc7\u516b\u767e\u5757\u5e94\u8be5\u600e\u4e48\u529e\uff1f<\/td>\n<td>\u2460 \u4f4f\u5bbf\u6807\u51c6\uff1b\u2461 \u7968\u636e\u671f\u9650\uff1b\u2462 \u5e02\u5185\u4ea4\u901a\u8d39\u7528\uff08\u6b63\u786e\uff09<\/td>\n<td align=\"right\">3<\/td>\n<\/tr>\n<tr>\n<td><code>ERR_FIN_403<\/code> \u662f\u4ec0\u4e48\u610f\u601d\uff1f<\/td>\n<td>\u2460 \u62a5\u9500\u5165\u53e3\uff08\u6b63\u786e\uff09<\/td>\n<td align=\"right\">1<\/td>\n<\/tr>\n<tr>\n<td>\u4e0a\u6d77\u4f4f\u5bbf\u4e00\u665a\u6700\u591a\u62a5\u591a\u5c11\uff1f<\/td>\n<td>\u2460 \u4f4f\u5bbf\u6807\u51c6\uff08\u6b63\u786e\uff09<\/td>\n<td align=\"right\">1<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>\u201c\u6253\u8f66\u8d39\u201d\u8fd9\u9053\u9898\u7684\u6b63\u786e\u7247\u6bb5\u6392\u5728\u7b2c 3\uff1b<code>ERR_FIN_403<\/code> \u548c\u201c\u4e0a\u6d77\u4f4f\u5bbf\u201d\u4e24\u9053\u9898\u7684\u6b63\u786e\u7247\u6bb5\u90fd\u6392\u5728\u7b2c 1\u3002\u4e0b\u9762\u7528\u4e09\u4e2a\u6570\u5b57\u6982\u62ec\u8fd9\u4e2a\u7ed3\u679c\uff1a<\/p>\n<pre><code class=\"text\">Hit@1 = 2 \/ 3 = 0.667\nHit@3 = 3 \/ 3 = 1.000\nMRR   = (1\/3 + 1 + 1) \/ 3 = 0.778<\/code><\/pre>\n<p><code>Hit@1 = 2\/3<\/code> \u8868\u793a\u53ea\u67e5\u770b\u6bcf\u9053\u9898\u7684\u7b2c 1 \u6761\u7ed3\u679c\u65f6\uff0c\u4e09\u9053\u9898\u4e2d\u6709\u4e24\u9053\u80fd\u591f\u76f4\u63a5\u627e\u5230\u6b63\u786e\u7247\u6bb5\u3002<code>Hit@3 = 3\/3<\/code> \u8868\u793a\u67e5\u770b\u6bcf\u9053\u9898\u7684\u524d 3 \u6761\u65f6\uff0c\u4e09\u9053\u9898\u90fd\u80fd\u627e\u5230\u6b63\u786e\u7247\u6bb5\u3002<\/p>\n<p>MRR \u7528\u6765\u533a\u5206\u6b63\u786e\u7247\u6bb5\u6392\u5728\u7b2c 1 \u8fd8\u662f\u7b2c 3\u3002\u6392\u7b2c 1 \u8bb0 1 \u5206\uff0c\u6392\u7b2c 3 \u8bb0 <code>1\/3<\/code> \u5206\uff0c\u518d\u5bf9\u4e09\u9053\u9898\u53d6\u5e73\u5747\uff0c\u6240\u4ee5\u8fd9\u91cc\u7684 MRR \u662f <code>0.778<\/code>\u3002\u8fd9\u4e2a\u6570\u5b57\u8d8a\u63a5\u8fd1 1\uff0c\u8868\u793a\u6b63\u786e\u7247\u6bb5\u901a\u5e38\u51fa\u73b0\u5f97\u8d8a\u9760\u524d\u3002<\/p>\n<p>\u5f53\u524d Demo \u4f1a\u628a\u524d 5 \u6761\u7ed3\u679c\u4ea4\u7ed9 LLM\uff0c\u56e0\u6b64\u6b63\u786e\u7247\u6bb5\u5982\u679c\u6ca1\u6709\u8fdb\u5165\u524d 5 \u6761\uff0cLLM \u5c31\u65e0\u6cd5\u4f9d\u636e\u5b83\u56de\u7b54\uff0c\u8fd9\u662f\u6700\u57fa\u672c\u7684\u5e95\u7ebf\u3002\u8d22\u52a1\u5236\u5ea6\u6d89\u53ca\u91d1\u989d\u548c\u5ba1\u6279\u6761\u4ef6\uff0c\u8fd8\u53ef\u4ee5\u63d0\u51fa\u66f4\u4e25\u683c\u7684\u8981\u6c42\uff0c\u4f8b\u5982\u5173\u952e\u95ee\u9898\u7684\u6b63\u786e\u7247\u6bb5\u5fc5\u987b\u8fdb\u5165\u524d 3 \u6761\uff0c\u5e76\u4e14\u65b0\u7248\u672c\u4e0d\u80fd\u6bd4\u65e7\u7248\u672c\u9000\u6b65\u3002<\/p>\n<p>\u8fd9\u4e09\u9053\u9898\u53ea\u80fd\u7528\u6765\u6f14\u793a\u8ba1\u7b97\u65b9\u6cd5\uff0c\u6570\u91cf\u4e0d\u8db3\u4ee5\u5224\u65ad\u7cfb\u7edf\u80fd\u5426\u4e0a\u7ebf\u3002\u6b63\u5f0f\u6d4b\u8bd5\u8fd8\u8981\u589e\u52a0\u66f4\u591a\u771f\u5b9e\u95ee\u6cd5\uff0c\u5e76\u8bb0\u5f55\u8d44\u6599\u7248\u672c\u3001Embedding \u6a21\u578b\u7248\u672c\u3001\u5206\u5757\u65b9\u5f0f\u3001Top-K\u3001\u68c0\u7d22\u6743\u91cd\u548c\u4ee3\u7801\u7248\u672c\uff0c\u624d\u80fd\u6b63\u786e\u6bd4\u8f83\u4e24\u6b21\u7ed3\u679c\u3002<\/p>\n<h3>\u518d\u68c0\u67e5\u6700\u7ec8\u56de\u7b54<\/h3>\n<p>\u5373\u4f7f\u68c0\u7d22\u547d\u4e2d\u6b63\u786e\u7247\u6bb5\uff0cLLM \u4ecd\u7136\u53ef\u80fd\u56de\u7b54\u9519\u8bef\u3002\u56e0\u6b64\uff0c\u5bf9\u4e8e\u6bcf\u4e2a\u7b54\u6848\u90fd\u9700\u8981\u68c0\u67e5\u4ee5\u4e0b\u5185\u5bb9\uff1a<\/p>\n<table>\n<thead>\n<tr>\n<th>\u68c0\u67e5\u9879<\/th>\n<th>\u8981\u95ee\u7684\u95ee\u9898<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>\u7b54\u6848\u662f\u5426\u6b63\u786e<\/td>\n<td>\u91d1\u989d\u3001\u65e5\u671f\u3001\u9519\u8bef\u7801\u548c\u7ed3\u8bba\u662f\u5426\u7b26\u5408\u6807\u51c6\u7b54\u6848\uff1f<\/td>\n<\/tr>\n<tr>\n<td>\u662f\u5426\u6f0f\u6389\u91cd\u70b9<\/td>\n<td>\u6d4b\u8bd5\u9898\u8981\u6c42\u7684\u5173\u952e\u6761\u4ef6\u662f\u5426\u90fd\u8bf4\u5230\u4e86\uff1f<\/td>\n<\/tr>\n<tr>\n<td>\u662f\u5426\u5fe0\u4e8e\u8d44\u6599<\/td>\n<td>\u7b54\u6848\u4e2d\u7684\u6bcf\u4e2a\u4e8b\u5b9e\u80fd\u5426\u5728\u672c\u6b21\u627e\u5230\u7684\u8d44\u6599\u91cc\u627e\u5230\u4f9d\u636e\uff1f<\/td>\n<\/tr>\n<tr>\n<td>\u5f15\u7528\u662f\u5426\u6709\u6548<\/td>\n<td><code>[\u6765\u6e90 N]<\/code> \u662f\u5426\u5b58\u5728\uff0c\u800c\u4e14\u771f\u7684\u652f\u6301\u5b83\u524d\u9762\u7684\u7ed3\u8bba\uff1f<\/td>\n<\/tr>\n<tr>\n<td>\u62d2\u7b54\u662f\u5426\u6b63\u786e<\/td>\n<td>\u77e5\u8bc6\u5e93\u6ca1\u6709\u7b54\u6848\u65f6\uff0c\u6a21\u578b\u662f\u5426\u8bf4\u660e\u8d44\u6599\u4e0d\u8db3\uff1f<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>\u6709\u4e9b\u5de5\u5177\u628a\u201c\u662f\u5426\u5fe0\u4e8e\u8d44\u6599\u201d\u79f0\u4e3a faithfulness \u6216 groundedness\uff0c\u68c0\u67e5\u7684\u662f\u6a21\u578b\u6709\u6ca1\u6709\u8bf4\u51fa\u8d44\u6599\u4e2d\u4e0d\u5b58\u5728\u7684\u5185\u5bb9\u3002<\/p>\n<p>\u201c\u5fe0\u4e8e\u8d44\u6599\u201d\u4e0d\u7b49\u4e8e\u201c\u4e8b\u5b9e\u6b63\u786e\u201d\u3002\u5982\u679c\u77e5\u8bc6\u5e93\u91cc\u4fdd\u5b58\u7684\u662f\u8fc7\u671f\u5236\u5ea6\uff0c\u6a21\u578b\u5373\u4f7f\u9010\u5b57\u590d\u8ff0\uff0c\u7b54\u6848\u4ecd\u7136\u4f1a\u9519\u3002\u56e0\u6b64\uff0c\u6d4b\u8bd5\u4f7f\u7528\u7684\u6587\u6863\u8fd8\u8981\u7ecf\u8fc7\u6709\u6548\u6027\u548c\u7248\u672c\u68c0\u67e5\u3002<\/p>\n<p>\u5bb9\u6613\u5224\u65ad\u7684\u5185\u5bb9\u53ef\u4ee5\u76f4\u63a5\u7528\u4ee3\u7801\u68c0\u67e5\uff0c\u4f8b\u5982\u5f15\u7528\u7f16\u53f7\u662f\u5426\u5b58\u5728\u3001\u7b54\u6848\u662f\u5426\u4e3a\u7a7a\u3001\u662f\u5426\u5305\u542b\u6b63\u786e\u91d1\u989d\u3002\u6bd4\u8f83\u7075\u6d3b\u7684\u81ea\u7136\u8bed\u8a00\u7b54\u6848\u53ef\u4ee5\u7531\u4eba\u5de5\u62bd\u67e5\uff0c\u4e5f\u53ef\u4ee5\u8ba9\u53e6\u4e00\u4e2a LLM \u6309\u56fa\u5b9a\u89c4\u5219\u534f\u52a9\u8bc4\u5206\u3002\u4e0d\u8fc7\uff0cLLM \u8bc4\u5ba1\u4e5f\u4f1a\u51fa\u9519\uff0c\u4e0d\u80fd\u628a\u5b83\u5f53\u4f5c\u7edd\u5bf9\u6b63\u786e\u7684\u88c1\u5224\uff1b\u91cd\u8981\u95ee\u9898\u4ecd\u8981\u4fdd\u7559\u4eba\u5de5\u590d\u6838\u3002<\/p>\n<h3>\u4e3a\u6574\u6761\u94fe\u8def\u8bbe\u5b9a\u901a\u8fc7\u6761\u4ef6<\/h3>\n<p>\u5bf9\u6700\u7ec8\u7528\u6237\u6765\u8bf4\uff0c\u4e00\u9053\u9898\u53ea\u6709\u540c\u65f6\u6ee1\u8db3\u4e0b\u9762\u51e0\u9879\u624d\u7b97\u771f\u6b63\u6210\u529f\uff1a<\/p>\n<ol>\n<li>\u641c\u7d22\u6a21\u5757\u627e\u5230\u4e86\u6b63\u786e\u8d44\u6599\u3002<\/li>\n<li>\u7b54\u6848\u5305\u542b\u5fc5\u987b\u8bf4\u660e\u7684\u8981\u70b9\u3002<\/li>\n<li>\u7b54\u6848\u6ca1\u6709\u7f16\u9020\u8d44\u6599\u4e4b\u5916\u7684\u4e8b\u5b9e\u3002<\/li>\n<li>\u5f15\u7528\u786e\u5b9e\u80fd\u591f\u652f\u6301\u7b54\u6848\u3002<\/li>\n<li>\u54cd\u5e94\u65f6\u95f4\u3001\u8d39\u7528\u548c\u8d44\u6599\u6743\u9650\u7b26\u5408\u8981\u6c42\u3002<\/li>\n<\/ol>\n<p>\u9ad8\u98ce\u9669\u95ee\u9898\u8fd8\u5e94\u8be5\u8bbe\u7f6e\u201c\u4e00\u7968\u5426\u51b3\u201d\u3002\u4f8b\u5982\u91d1\u989d\u3001\u65e5\u671f\u3001\u8eab\u4efd\u6216\u6743\u9650\u53ea\u8981\u6709\u4e00\u9879\u8bf4\u9519\uff0c\u6574\u9053\u9898\u5c31\u7b97\u5931\u8d25\uff0c\u4e0d\u80fd\u56e0\u4e3a\u8bed\u8a00\u6d41\u7545\u6216\u5176\u4ed6\u5206\u6570\u8f83\u9ad8\u5c31\u7b97\u901a\u8fc7\u3002<\/p>\n<p>\u4fee\u6539\u7cfb\u7edf\u540e\uff0c\u5e94\u4f7f\u7528\u540c\u4e00\u5957\u6d4b\u8bd5\u9898\u4e0e\u65e7\u7248\u672c\u6bd4\u8f83\u3002\u4f8b\u5982\u5206\u522b\u6d4b\u8bd5\u7eaf\u5411\u91cf\u3001\u7eaf BM25 \u548c\u6df7\u5408\u68c0\u7d22\uff0c\u6216\u8005\u53ea\u6539\u53d8 Top-K\uff0c\u518d\u89c2\u5bdf\u7ed3\u679c\u3002\u4e00\u6b21\u53ea\u6539\u53d8\u4e00\u4e2a\u4e3b\u8981\u6761\u4ef6\uff0c\u624d\u77e5\u9053\u63d0\u5347\u6216\u9000\u6b65\u662f\u7531\u4ec0\u4e48\u9020\u6210\u7684\u3002<\/p>\n<p>RAG \u6ca1\u6709\u9002\u7528\u4e8e\u6240\u6709\u516c\u53f8\u7684\u7edf\u4e00\u53ca\u683c\u5206\u6570\u3002\u4e00\u4e2a\u5185\u90e8\u95f2\u804a\u52a9\u624b\u548c\u4e00\u4e2a\u8d22\u52a1\u5236\u5ea6\u52a9\u624b\uff0c\u72af\u9519\u7684\u540e\u679c\u5b8c\u5168\u4e0d\u540c\u3002\u9879\u76ee\u5e94\u6839\u636e\u81ea\u5df1\u7684\u98ce\u9669\u51b3\u5b9a\u53d1\u5e03\u6761\u4ef6\uff0c\u4f8b\u5982\uff1a\u5173\u952e\u95ee\u9898\u5fc5\u987b\u5168\u90e8\u7b54\u5bf9\uff1b\u65e0\u6cd5\u56de\u7b54\u7684\u95ee\u9898\u4e0d\u80fd\u7f16\u9020\uff1b\u666e\u901a\u95ee\u9898\u7684\u547d\u4e2d\u7387\u4e0d\u80fd\u4f4e\u4e8e\u65e7\u7248\u672c\uff1b\u54cd\u5e94\u65f6\u95f4\u4e0d\u80fd\u8d85\u8fc7\u4ea7\u54c1\u9884\u7b97\u3002\u8fd9\u4e9b\u662f\u9879\u76ee\u81ea\u5df1\u7684\u9a8c\u6536\u89c4\u5219\uff0c\u4e0d\u662f\u884c\u4e1a\u7edf\u4e00\u6807\u51c6\u3002<\/p>\n<p>\u5f53\u6d4b\u8bd5\u5931\u8d25\u65f6\uff0c\u53ef\u4ee5\u6309\u4e0b\u9762\u7684\u987a\u5e8f\u6392\u67e5\uff1a<\/p>\n<table>\n<thead>\n<tr>\n<th>\u770b\u5230\u7684\u73b0\u8c61<\/th>\n<th>\u5148\u68c0\u67e5\u4ec0\u4e48<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>\u539f\u6587\u6ca1\u6709\u51fa\u73b0\u5728\u4efb\u4f55\u53ef\u641c\u7d22\u7247\u6bb5\u4e2d<\/td>\n<td>\u6587\u6863\u662f\u5426\u89e3\u6790\u6210\u529f\u3001\u5206\u5757\u662f\u5426\u5408\u7406\u3001\u7d22\u5f15\u662f\u5426\u66f4\u65b0<\/td>\n<\/tr>\n<tr>\n<td>\u6b63\u786e\u7247\u6bb5\u5b58\u5728\uff0c\u5374\u6ca1\u6709\u6392\u8fdb\u524d\u51e0\u540d<\/td>\n<td>\u5411\u91cf\u6a21\u578b\u3001\u5173\u952e\u8bcd\u641c\u7d22\u3001\u67e5\u8be2\u6539\u5199\u548c\u6392\u5e8f\u6743\u91cd<\/td>\n<\/tr>\n<tr>\n<td>\u6b63\u786e\u7247\u6bb5\u5df2\u7ecf\u4ea4\u7ed9 LLM\uff0c\u7b54\u6848\u4ecd\u7136\u9519\u8bef<\/td>\n<td>\u63d0\u793a\u8bcd\u3001\u6a21\u578b\u3001\u8d44\u6599\u51b2\u7a81\u548c\u62d2\u7b54\u89c4\u5219<\/td>\n<\/tr>\n<tr>\n<td>\u7b54\u6848\u6b63\u786e\uff0c\u4f46\u5f15\u7528\u5bf9\u4e0d\u4e0a<\/td>\n<td>\u6765\u6e90\u7f16\u53f7\u6620\u5c04\u548c\u5f15\u7528\u68c0\u67e5<\/td>\n<\/tr>\n<tr>\n<td>\u6d4b\u8bd5\u73af\u5883\u6b63\u5e38\uff0c\u7ebf\u4e0a\u5076\u5c14\u5931\u8d25<\/td>\n<td>\u6587\u6863\u6216\u6a21\u578b\u7248\u672c\u3001\u771f\u5b9e\u7528\u6237\u95ee\u6cd5\u3001\u5916\u90e8\u670d\u52a1\u9519\u8bef\u548c\u54cd\u5e94\u5ef6\u8fdf<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h2>\u4e0a\u7ebf\u4ee5\u540e\u600e\u6837\u76d1\u63a7 RAG<\/h2>\n<p>\u79bb\u7ebf\u8bc4\u4f30\u5728\u53d1\u5e03\u524d\u6216\u7248\u672c\u6bd4\u8f83\u65f6\u8fd0\u884c\uff0c\u7528\u5e26\u6807\u51c6\u7b54\u6848\u7684\u6d4b\u8bd5\u9898\u68c0\u67e5\u5df2\u77e5\u95ee\u9898\u662f\u5426\u9000\u6b65\u3002\u7ebf\u4e0a\u76d1\u63a7\u8bb0\u5f55\u7cfb\u7edf\u5904\u7406\u7528\u6237\u8bf7\u6c42\u65f6\u7684\u8017\u65f6\u3001\u9519\u8bef\u3001\u68c0\u7d22\u7ed3\u679c\u548c\u53cd\u9988\uff0c\u7528\u6765\u53d1\u73b0\u6d4b\u8bd5\u96c6\u5c1a\u672a\u8986\u76d6\u7684\u95ee\u9898\u3002<\/p>\n<p>\u7ebf\u4e0a\u8bf7\u6c42\u901a\u5e38\u6ca1\u6709\u53c2\u8003\u7b54\u6848\uff0c\u6240\u4ee5\u770b\u677f\u65e0\u6cd5\u81ea\u52a8\u5224\u65ad\u6bcf\u4e2a\u56de\u7b54\u662f\u5426\u7b26\u5408\u4e8b\u5b9e\uff1b\u79bb\u7ebf\u6d4b\u8bd5\u96c6\u53c8\u53ea\u8986\u76d6\u5df2\u7ecf\u6536\u5f55\u7684\u95ee\u6cd5\u3002\u4e24\u8005\u9700\u8981\u8fd9\u6837\u914d\u5408\uff1a\u53d1\u5e03\u524d\u7528\u79bb\u7ebf\u8bc4\u4f30\u62e6\u622a\u5df2\u77e5\u56de\u5f52\uff0c\u4e0a\u7ebf\u540e\u4ece\u5f02\u5e38\u8bf7\u6c42\u548c\u7528\u6237\u53cd\u9988\u4e2d\u62bd\u6837\uff0c\u7531\u4eba\u5de5\u786e\u8ba4\u95ee\u9898\uff0c\u518d\u628a\u8131\u654f\u540e\u7684\u5931\u8d25\u6837\u672c\u52a0\u5165\u8bc4\u4f30\u96c6\u3002<\/p>\n<p>\u6bcf\u4e2a\u7ebf\u4e0a\u8bf7\u6c42\u90fd\u5e94\u4fdd\u5b58\u4e00\u6761 trace\uff08\u8c03\u7528\u94fe\u8bb0\u5f55\uff09\u3002\u5b83\u7528\u540c\u4e00\u4e2a\u8bf7\u6c42 ID \u4e32\u8d77 Embedding\u3001\u68c0\u7d22\u3001\u91cd\u6392\u3001\u5de5\u5177\u8c03\u7528\u548c LLM \u751f\u6210\u7b49\u6b65\u9aa4\uff1b\u5176\u4e2d\u6bcf\u4e2a\u6b65\u9aa4\u79f0\u4e3a\u4e00\u4e2a span\uff08\u6b65\u9aa4\u8bb0\u5f55\uff09\u3002<\/p>\n<h3>\u4e00\u6761 trace \u8981\u8bb0\u5f55\u4ec0\u4e48<\/h3>\n<pre style=\"display:none;\"><code class=\"mermaid\">flowchart LR\n    Q[\u7ebf\u4e0a\u8bf7\u6c42] --&gt; T[\u4e00\u6761 RAG trace]\n    T --&gt; E[Embedding]\n    T --&gt; R1[\u521d\u59cb\u68c0\u7d22]\n    T --&gt; R2[\u53ef\u9009\u7684\u518d\u6b21\u68c0\u7d22]\n    T --&gt; L[LLM \u751f\u6210]\n    E --&gt; O[Trace \u5b58\u50a8]\n    R1 --&gt; O\n    R2 --&gt; O\n    L --&gt; O\n    O --&gt; D[\u6307\u6807\u3001\u770b\u677f\u4e0e\u544a\u8b66]\n    O --&gt; S[\u62bd\u6837\u81ea\u52a8\u8bc4\u5ba1\u4e0e\u4eba\u5de5\u590d\u6838]\n    S --&gt; F[\u5931\u8d25\u6837\u672c\u5e93]\n    F --&gt; G[\u79bb\u7ebf\u56de\u5f52\u8bc4\u4f30\u96c6]<\/code><\/pre>\n<p>\u4e0b\u9762\u8fd9\u4e9b\u5b57\u6bb5\u5206\u522b\u7528\u4e8e\u590d\u73b0\u8bf7\u6c42\u3001\u5b9a\u4f4d\u6545\u969c\u548c\u7edf\u8ba1\u8d8b\u52bf\uff1a<\/p>\n<table>\n<thead>\n<tr>\n<th>\u8bb0\u5f55\u5bf9\u8c61<\/th>\n<th>\u81f3\u5c11\u4fdd\u5b58\u4ec0\u4e48<\/th>\n<th>\u7528\u4e8e\u56de\u7b54\u4ec0\u4e48\u95ee\u9898<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>\u8bf7\u6c42\u4e0e\u7248\u672c<\/td>\n<td><code>requestId<\/code>\u3001\u65f6\u95f4\u3001\u73af\u5883\u3001\u5e94\u7528\u7248\u672c\u3001\u4f1a\u8bdd ID\uff1b\u7528\u6237\u6216\u4f7f\u7528\u65b9\u91c7\u7528\u4e0d\u53ef\u9006\u6807\u8bc6<\/td>\n<td>\u54ea\u4e2a\u7248\u672c\u3001\u54ea\u7c7b\u7528\u6237\u9047\u5230\u4e86\u95ee\u9898\uff1f<\/td>\n<\/tr>\n<tr>\n<td>\u77e5\u8bc6\u5e93<\/td>\n<td>\u8bed\u6599\u5feb\u7167\u3001\u6587\u6863\u66f4\u65b0\u65f6\u95f4\u3001Embedding \u6a21\u578b\u4e0e revision\u3001\u5206\u5757\u548c\u7d22\u5f15\u7248\u672c<\/td>\n<td>\u5f53\u65f6\u68c0\u7d22\u7684\u662f\u54ea\u4e00\u7248\u8d44\u6599\uff0c\u80fd\u5426\u590d\u73b0\u6392\u540d\uff1f<\/td>\n<\/tr>\n<tr>\n<td>\u68c0\u7d22\u8f93\u5165<\/td>\n<td>\u539f\u59cb\u67e5\u8be2\u6216\u8131\u654f\u6458\u8981\u3001\u6539\u5199\u540e\u7684\u67e5\u8be2\u3001Top-K\u3001\u4e24\u8def\u6743\u91cd\u3001\u8fc7\u6ee4\u6761\u4ef6<\/td>\n<td>\u67e5\u8be2\u662f\u5426\u6539\u5199\u9519\u8bef\uff0c\u6743\u9650\u6216\u8fc7\u6ee4\u6761\u4ef6\u662f\u5426\u6392\u9664\u4e86\u6b63\u786e\u8d44\u6599\uff1f<\/td>\n<\/tr>\n<tr>\n<td>\u68c0\u7d22\u8f93\u51fa<\/td>\n<td>chunk\/source ID\u3001\u6392\u540d\u3001\u539f\u59cb\u5206\u3001\u5f52\u4e00\u5316\u5206\u3001\u6700\u7ec8\u5206\u3001\u7ed3\u679c\u6570\u91cf\u548c\u6bcf\u6b21\u68c0\u7d22\u8017\u65f6<\/td>\n<td>\u6b63\u786e\u7247\u6bb5\u5728\u54ea\u4e00\u6b65\u6389\u51fa Top-K\uff0c\u5ef6\u8fdf\u662f\u5426\u6765\u81ea\u68c0\u7d22\uff1f<\/td>\n<\/tr>\n<tr>\n<td>\u751f\u6210\u8fc7\u7a0b<\/td>\n<td>prompt \u7248\u672c\u3001\u6a21\u578b\u7248\u672c\u3001\u5b9e\u9645\u9001\u5165\u7684\u6765\u6e90 ID\u3001\u5de5\u5177\u8c03\u7528\u53ca\u6b65\u9aa4\u6570<\/td>\n<td>LLM \u63a5\u6536\u4e86\u54ea\u4e9b\u8d44\u6599\uff0c\u662f\u5426\u53d1\u751f\u91cd\u590d\u641c\u7d22\u6216\u591a\u4f59\u8c03\u7528\uff1f<\/td>\n<\/tr>\n<tr>\n<td>\u8fd0\u884c\u7ed3\u679c<\/td>\n<td>\u603b\u8017\u65f6\u3001\u5206\u9636\u6bb5\u8017\u65f6\u3001token\u3001\u4f30\u7b97\u6210\u672c\u3001\u9519\u8bef\u7c7b\u578b\u3001\u56de\u7b54\u6216\u62d2\u7b54\u7c7b\u522b\u3001\u5f15\u7528 ID<\/td>\n<td>\u8bf7\u6c42\u4e3a\u4f55\u5931\u8d25\u6216\u53d8\u6162\uff0c\u6210\u672c\u7531\u54ea\u4e00\u6b65\u4ea7\u751f\uff1f<\/td>\n<\/tr>\n<tr>\n<td>\u8d28\u91cf\u6807\u7b7e<\/td>\n<td>\u7528\u6237\u53cd\u9988\u3001\u6539\u5199\u540e\u91cd\u95ee\u3001\u8f6c\u4eba\u5de5\u3001\u81ea\u52a8\u8bc4\u5ba1\u7ed3\u679c\u548c\u4eba\u5de5\u590d\u6838\u6807\u7b7e<\/td>\n<td>\u54ea\u4e9b\u8bf7\u6c42\u9700\u8981\u62bd\u6837\u590d\u6838\uff0c\u54ea\u4e9b\u5931\u8d25\u5e94\u52a0\u5165\u8bc4\u4f30\u96c6\uff1f<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>\u95ee\u9898\u3001\u6587\u6863\u7247\u6bb5\u548c\u56de\u7b54\u53ef\u80fd\u5305\u542b\u79c1\u6709\u6570\u636e\u3002\u751f\u4ea7\u73af\u5883\u4e0d\u5e94\u9ed8\u8ba4\u628a\u6240\u6709\u539f\u6587\u3001\u5b8c\u6574 prompt\u3001\u5411\u91cf\u6216\u6a21\u578b\u9690\u85cf\u63a8\u7406\u5199\u8fdb\u65e5\u5fd7\uff1b\u5e94\u6309\u6570\u636e\u5206\u7c7b\u51b3\u5b9a\u662f\u5426\u8131\u654f\u3001\u622a\u65ad\u3001\u52a0\u5bc6\u6216\u53ea\u4fdd\u5b58 ID\uff0c\u5e76\u8bbe\u7f6e\u8bbf\u95ee\u63a7\u5236\u548c\u4fdd\u7559\u671f\u9650\u3002\u8c03\u8bd5\u6240\u9700\u7684\u53ef\u89c1\u6027\u4e0e\u6570\u636e\u6700\u5c0f\u5316\u9700\u8981\u4e00\u8d77\u8bbe\u8ba1\u3002<\/p>\n<h3>\u770b\u677f\u548c\u544a\u8b66\u5e94\u8be5\u770b\u4ec0\u4e48<\/h3>\n<p>\u524d\u9762\u7684 Hit@K\u3001MRR \u7b49\u6307\u6807\u9700\u8981\u9884\u5148\u77e5\u9053\u6b63\u786e chunk\uff0c\u9002\u5408\u79bb\u7ebf\u6d4b\u8bd5\u3002\u7ebf\u4e0a\u8bf7\u6c42\u901a\u5e38\u6ca1\u6709\u6807\u51c6\u7b54\u6848\uff0c\u56e0\u6b64\u770b\u677f\u4e0d\u80fd\u76f4\u63a5\u663e\u793a\u201c\u56de\u7b54\u6b63\u786e\u7387\u201d\u3002\u770b\u677f\u5148\u8d1f\u8d23\u53d1\u73b0\u7cfb\u7edf\u5728\u54ea\u4e2a\u73af\u8282\u51fa\u73b0\u5f02\u5e38\uff0c\u518d\u4ece\u5f02\u5e38\u8bf7\u6c42\u4e2d\u62bd\u53d6 trace \u505a\u4eba\u5de5\u68c0\u67e5\u3002<\/p>\n<p>\u53ef\u4ee5\u6309\u7167\u4e00\u6b21\u8bf7\u6c42\u7ecf\u8fc7\u7684\u987a\u5e8f\u5b89\u6392\u770b\u677f\uff1a<\/p>\n<table>\n<thead>\n<tr>\n<th>\u60f3\u786e\u8ba4\u4ec0\u4e48<\/th>\n<th>\u770b\u677f\u8bb0\u5f55\u4ec0\u4e48<\/th>\n<th>\u51fa\u73b0\u5f02\u5e38\u65f6\u5148\u68c0\u67e5\u4ec0\u4e48<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>\u8d44\u6599\u662f\u5426\u5df2\u7ecf\u8fdb\u5165\u7d22\u5f15<\/td>\n<td>\u89e3\u6790\u5931\u8d25\u7684\u6587\u6863\u6570\u3001\u5df2\u7d22\u5f15\u6587\u6863\u6570\u3001chunk \u603b\u6570\u3001\u6700\u8fd1\u4e00\u6b21\u7d22\u5f15\u65f6\u95f4<\/td>\n<td>\u6587\u6863\u4e0a\u4f20\u6216\u89e3\u6790\u662f\u5426\u5931\u8d25\uff0c\u7d22\u5f15\u4efb\u52a1\u662f\u5426\u4e2d\u65ad<\/td>\n<\/tr>\n<tr>\n<td>\u641c\u7d22\u670d\u52a1\u662f\u5426\u6b63\u5e38<\/td>\n<td>\u68c0\u7d22\u62a5\u9519\u6bd4\u4f8b\u3001\u6ca1\u6709\u8fd4\u56de\u4efb\u4f55 chunk \u7684\u8bf7\u6c42\u6bd4\u4f8b\u3001\u68c0\u7d22\u8017\u65f6\u3001Agent \u662f\u5426\u53cd\u590d\u641c\u7d22<\/td>\n<td>\u68c0\u7d22\u670d\u52a1\u662f\u5426\u8d85\u65f6\uff0c\u8fc7\u6ee4\u6761\u4ef6\u662f\u5426\u628a\u8d44\u6599\u5168\u90e8\u6392\u9664<\/td>\n<\/tr>\n<tr>\n<td>LLM \u662f\u5426\u6b63\u5e38\u751f\u6210<\/td>\n<td>\u6a21\u578b\u62a5\u9519\u548c\u8d85\u65f6\u3001\u5f15\u7528\u7f16\u53f7\u662f\u5426\u5b58\u5728\u3001\u7cfb\u7edf\u62d2\u7b54\u7684\u8bf7\u6c42\u6bd4\u4f8b\u3001\u6bcf\u6b21\u8bf7\u6c42\u6d88\u8017\u7684 token \u548c\u8d39\u7528<\/td>\n<td>\u6a21\u578b\u670d\u52a1\u662f\u5426\u5f02\u5e38\uff0c\u5f15\u7528\u7f16\u53f7\u662f\u5426\u6620\u5c04\u9519\u8bef\uff0c\u63d0\u793a\u8bcd\u662f\u5426\u5bfc\u81f4\u8fc7\u5ea6\u62d2\u7b54<\/td>\n<\/tr>\n<tr>\n<td>\u7528\u6237\u548c\u6743\u9650\u662f\u5426\u51fa\u73b0\u95ee\u9898<\/td>\n<td>\u70b9\u8d5e\/\u70b9\u8e29\u3001\u7528\u6237\u662f\u5426\u9a6c\u4e0a\u6362\u4e00\u79cd\u95ee\u6cd5\u91cd\u95ee\u3001\u8f6c\u4eba\u5de5\u6b21\u6570\u3001\u662f\u5426\u68c0\u7d22\u5230\u8d8a\u6743\u8d44\u6599<\/td>\n<td>\u56de\u7b54\u662f\u5426\u96be\u4ee5\u4f7f\u7528\uff0c\u6743\u9650\u8fc7\u6ee4\u662f\u5426\u5931\u6548<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>\u68c0\u7d22\u8017\u65f6\u4e0d\u80fd\u53ea\u770b\u5e73\u5747\u503c\uff0c\u56e0\u4e3a\u5c11\u91cf\u7279\u522b\u6162\u7684\u8bf7\u6c42\u4f1a\u88ab\u5e73\u5747\u503c\u63a9\u76d6\u3002\u5e38\u7528\u7684 P95 \u53ef\u4ee5\u8fd9\u6837\u7406\u89e3\uff1a\u628a 100 \u6b21\u8bf7\u6c42\u6309\u8017\u65f6\u4ece\u5feb\u5230\u6162\u6392\u5217\uff0c\u7b2c 95 \u4e2a\u8bf7\u6c42\u7684\u8017\u65f6\u5c31\u662f P95\u3002\u5982\u679c\u68c0\u7d22 P95 \u662f 800 \u6beb\u79d2\uff0c\u5927\u7ea6\u6709 95 \u6b21\u8bf7\u6c42\u80fd\u5728 800 \u6beb\u79d2\u5185\u5b8c\u6210\uff0c\u5269\u4e0b 5 \u6b21\u66f4\u6162\u3002\u4ea7\u54c1\u53ef\u4ee5\u636e\u6b64\u89c4\u5b9a\u201c\u68c0\u7d22 P95 \u4e0d\u5f97\u8d85\u8fc7 800 \u6beb\u79d2\u201d\u4e4b\u7c7b\u7684\u76ee\u6807\u3002<\/p>\n<p>\u9664\u4e86\u57fa\u672c\u7684\u62a5\u9519\u548c\u5ef6\u8fdf\u4e4b\u5916\uff0c\u770b\u677f\u8fd8\u53ef\u4ee5\u8bb0\u5f55 Top-1 \u5411\u91cf\u76f8\u4f3c\u5ea6\u548c\u62d2\u7b54\u7387\uff0c\u4f46\u8fd9\u4e24\u4e2a\u6570\u5b57\u53ea\u80fd\u5e2e\u52a9\u7b5b\u9009\u95ee\u9898\uff0c\u4e0d\u80fd\u76f4\u63a5\u5224\u65ad\u56de\u7b54\u662f\u5426\u6b63\u786e\u3002<\/p>\n<p>\u8fd9\u91cc\u7684 Top-1 \u5411\u91cf\u76f8\u4f3c\u5ea6\uff0c\u662f\u5411\u91cf\u68c0\u7d22\u7b2c\u4e00\u540d\u7684\u4f59\u5f26\u76f8\u4f3c\u5ea6\uff0c\u4e0d\u662f\u7b54\u6848\u6b63\u786e\u7387\u3002\u53ea\u6709\u4f7f\u7528\u76f8\u540c\u7684 Embedding \u6a21\u578b\u3001\u76f8\u8fd1\u7684\u8d44\u6599\u548c\u76f8\u8fd1\u7684\u95ee\u9898\u65f6\uff0c\u524d\u540e\u5206\u6570\u624d\u9002\u5408\u6bd4\u8f83\u3002\u66f4\u6362 Embedding \u6a21\u578b\u4ee5\u540e\uff0c\u8ba1\u7b97\u5206\u6570\u7684\u6807\u51c6\u4e5f\u53d8\u4e86\uff0c\u6b64\u65f6\u4e0d\u80fd\u628a\u65b0\u65e7\u5206\u6570\u76f4\u63a5\u5bf9\u6bd4\u3002<\/p>\n<p>\u62d2\u7b54\u7387\u662f\u7cfb\u7edf\u56de\u7b54\u201c\u73b0\u6709\u8d44\u6599\u4e0d\u8db3\u201d\u7684\u8bf7\u6c42\u6240\u5360\u6bd4\u4f8b\u3002\u62d2\u7b54\u7387\u4e0a\u5347\u53ef\u80fd\u6709\u4e24\u79cd\u539f\u56e0\uff1a\u68c0\u7d22\u51fa\u73b0\u95ee\u9898\uff0c\u5bfc\u81f4\u7cfb\u7edf\u6ca1\u6709\u627e\u5230\u539f\u672c\u5b58\u5728\u7684\u8d44\u6599\uff1b\u6216\u8005\u7528\u6237\u5f00\u59cb\u8be2\u95ee\u77e5\u8bc6\u5e93\u6ca1\u6709\u6536\u5f55\u7684\u65b0\u95ee\u9898\u3002\u5de5\u7a0b\u5e08\u9700\u8981\u62bd\u67e5\u5bf9\u5e94\u7684 trace\uff0c\u624d\u80fd\u533a\u5206\u8fd9\u4e24\u79cd\u60c5\u51b5\u3002\u70b9\u8d5e\u3001\u70b9\u8e29\u548c\u201c\u6ca1\u6709\u7ee7\u7eed\u8ffd\u95ee\u201d\u4e5f\u4e00\u6837\uff0c\u53ea\u80fd\u8bf4\u660e\u7528\u6237\u7684\u884c\u4e3a\uff0c\u4e0d\u80fd\u8bc1\u660e\u7b54\u6848\u7b26\u5408\u4e8b\u5b9e\u3002<\/p>\n<p>\u8bbe\u7f6e\u544a\u8b66\u65f6\uff0c\u4e0d\u8981\u628a\u6240\u6709\u8bf7\u6c42\u6df7\u5728\u4e00\u8d77\u53d6\u5e73\u5747\u3002\u770b\u677f\u81f3\u5c11\u8981\u80fd\u6309\u5e94\u7528\u7248\u672c\u3001\u7d22\u5f15\u7248\u672c\u3001\u4f7f\u7528\u65b9\u3001\u8bed\u8a00\u548c\u95ee\u9898\u7c7b\u578b\u7b5b\u9009\u3002\u4f8b\u5982\uff0c\u7d22\u5f15\u66f4\u65b0\u540e\u53ea\u6709\u67d0\u5bb6\u516c\u53f8\u7684\u4e2d\u6587\u8d22\u52a1\u95ee\u9898\u51fa\u9519\uff0c\u5168\u7ad9\u5e73\u5747\u503c\u53ef\u80fd\u53d8\u5316\u5f88\u5c0f\uff0c\u5206\u7ec4\u540e\u624d\u80fd\u770b\u5230\u95ee\u9898\u96c6\u4e2d\u5728\u54ea\u91cc\u3002<\/p>\n<p>\u544a\u8b66\u6761\u4ef6\u5e94\u5199\u6e05\u695a\u53d1\u751f\u4e86\u4ec0\u4e48\uff0c\u4ee5\u53ca\u9700\u8981\u68c0\u67e5\u54ea\u4e2a\u73af\u8282\u3002\u4f8b\u5982\uff1a<\/p>\n<ul>\n<li>\u66f4\u65b0\u7d22\u5f15\u540e\uff0c\u56fa\u5b9a\u6d4b\u8bd5\u9898\u7684 Hit@5 \u4f4e\u4e8e\u4e0a\u4e00\u7248\u672c\uff0c\u5148\u505c\u6b62\u53d1\u5e03\u5e76\u68c0\u67e5\u65b0\u7d22\u5f15\u3002<\/li>\n<li>\u4e0d\u5b58\u5728\u7684\u6765\u6e90\u7f16\u53f7\u7a81\u7136\u589e\u591a\uff0c\u5148\u68c0\u67e5\u6765\u6e90\u7f16\u53f7\u7684\u751f\u6210\u548c\u6620\u5c04\u3002<\/li>\n<li>\u68c0\u7d22 P95 \u8fde\u7eed\u8d85\u8fc7\u4ea7\u54c1\u89c4\u5b9a\u7684\u65f6\u95f4\uff0c\u5148\u68c0\u67e5\u68c0\u7d22\u670d\u52a1\u4ee5\u53ca\u4fdd\u5b58\u7d22\u5f15\u7684\u6570\u636e\u5e93\u6216\u6587\u4ef6\u3002<\/li>\n<li>\u6587\u6863\u6570\u6216 chunk \u6570\u5728\u6ca1\u6709\u66f4\u65b0\u8d44\u6599\u65f6\u7a81\u7136\u51cf\u5c11\uff0c\u5148\u68c0\u67e5\u7d22\u5f15\u4efb\u52a1\u662f\u5426\u5931\u8d25\u3002<\/li>\n<\/ul>\n<p>\u53d1\u5e03\u65b0\u7248\u672c\u65f6\uff0c\u53ef\u4ee5\u5206\u4e09\u6b65\u8fdb\u884c\uff1a<\/p>\n<ol>\n<li>\u5148\u8fd0\u884c\u5e26\u6807\u51c6\u7b54\u6848\u7684\u79bb\u7ebf\u6d4b\u8bd5\uff0c\u672a\u8fbe\u5230\u8981\u6c42\u5c31\u505c\u6b62\u53d1\u5e03\u3002<\/li>\n<li>\u98ce\u9669\u8f83\u9ad8\u7684\u6539\u52a8\u5148\u4f7f\u7528\u5f71\u5b50\u6d41\u91cf\uff1a\u65b0\u7248\u672c\u5904\u7406\u5c11\u91cf\u7ebf\u4e0a\u95ee\u9898\uff0c\u4f46\u5b83\u751f\u6210\u7684\u7b54\u6848\u4e0d\u5c55\u793a\u7ed9\u7528\u6237\u3002\u5de5\u7a0b\u5e08\u53ea\u6bd4\u8f83\u65b0\u65e7\u7248\u672c\u7684\u8017\u65f6\u3001\u62a5\u9519\u548c\u68c0\u7d22\u7ed3\u679c\u3002<\/li>\n<li>\u5f71\u5b50\u6d4b\u8bd5\u6ca1\u6709\u53d1\u73b0\u660e\u663e\u9519\u8bef\u540e\uff0c\u518d\u8ba9\u4e00\u5c0f\u90e8\u5206\u7528\u6237\u6536\u5230\u65b0\u7248\u672c\u7b54\u6848\uff0c\u7ee7\u7eed\u89c2\u5bdf\u53cd\u9988\u548c\u5f02\u5e38\u8bf7\u6c42\u3002<\/li>\n<\/ol>\n<p>\u65b0\u7d22\u5f15\u3001prompt \u548c\u751f\u6210\u6a21\u578b\u6700\u597d\u5206\u5f00\u53d1\u5e03\u3002\u4e00\u6b21\u53ea\u6539\u4e00\u9879\uff0c\u51fa\u73b0\u95ee\u9898\u65f6\u624d\u80fd\u5224\u65ad\u539f\u56e0\u3002<\/p>\n<h3>\u73b0\u6210\u5de5\u5177\u5206\u522b\u8d1f\u8d23\u4ec0\u4e48<\/h3>\n<p>\u6307\u6807\u8ba1\u7b97\u3001trace \u7684\u5b57\u6bb5\u7ea6\u5b9a\u548c\u53ef\u89c6\u5316\u5e73\u53f0\u5c5e\u4e8e\u4e0d\u540c\u5c42\u6b21\uff1a<\/p>\n<table>\n<thead>\n<tr>\n<th>\u7c7b\u578b<\/th>\n<th>\u4ee3\u8868\u65b9\u6848<\/th>\n<th>\u4f5c\u7528\u4e0e\u8fb9\u754c<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>\u8bc4\u4f30\u6307\u6807\/\u5e93<\/td>\n<td>RAGAS\u3001\u81ea\u5b9a\u4e49\u89c4\u5219\u548c\u4fe1\u606f\u68c0\u7d22\u6307\u6807<\/td>\n<td>\u8ba1\u7b97\u5fe0\u5b9e\u5ea6\u3001\u4e0a\u4e0b\u6587\u8d28\u91cf\u548c\u68c0\u7d22\u6307\u6807\uff1b\u90e8\u5206\u6307\u6807\u4e0d\u8981\u6c42\u53c2\u8003\u7b54\u6848\uff0c\u4f46\u4ecd\u4f9d\u8d56\u8bc4\u5206\u6a21\u578b\u3001\u9879\u76ee\u89c4\u5219\u6216\u4eba\u5de5\u6807\u7b7e\uff0c\u4e0d\u80fd\u5355\u72ec\u8bc1\u660e\u4e8b\u5b9e\u6b63\u786e<\/td>\n<\/tr>\n<tr>\n<td>\u901a\u7528\u9065\u6d4b<\/td>\n<td>OpenTelemetry + \u6307\u6807\/\u65e5\u5fd7\/trace \u540e\u7aef<\/td>\n<td>\u91c7\u96c6\u8017\u65f6\u3001\u9519\u8bef\u548c\u8c03\u7528\u94fe\uff1bRAG \u4e13\u7528\u5b57\u6bb5\u4ecd\u8981\u6309\u5b9e\u9645\u7ba1\u9053\u8865\u5145<\/td>\n<\/tr>\n<tr>\n<td>AI trace \u5b57\u6bb5\u7ea6\u5b9a<\/td>\n<td>OpenInference<\/td>\n<td>\u5728 OpenTelemetry \u4e4b\u4e0a\u89c4\u5b9a\u68c0\u7d22\u3001\u91cd\u6392\u3001Embedding \u548c LLM \u7b49\u6b65\u9aa4\u600e\u6837\u8bb0\u5f55\uff0c\u5e76\u7edf\u4e00\u6587\u6863 ID\u3001\u5206\u6570\u7b49\u5b57\u6bb5<\/td>\n<\/tr>\n<tr>\n<td>LLM \u5e94\u7528\u5e73\u53f0<\/td>\n<td>LangSmith<\/td>\n<td>\u63d0\u4f9b trace\u3001\u6570\u636e\u96c6\u3001\u79bb\u7ebf\/\u5728\u7ebf\u8bc4\u4f30\u3001\u53cd\u9988\u3001\u770b\u677f\u548c\u544a\u8b66<\/td>\n<\/tr>\n<tr>\n<td>\u5f00\u6e90\u53ef\u81ea\u6258\u7ba1\u5e73\u53f0<\/td>\n<td>Arize Phoenix<\/td>\n<td>\u901a\u8fc7 OpenTelemetry\/OpenInference \u63a5\u6536 trace\uff0c\u5e76\u63d0\u4f9b\u8bc4\u4f30\u3001\u6570\u636e\u96c6\u548c\u7248\u672c\u5b9e\u9a8c<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>\u4f8b\u5982\uff0cLangSmith \u6216 Phoenix \u53ef\u4ee5\u628a\u4e00\u6b21\u8bf7\u6c42\u7684\u67e5\u8be2\u3001\u68c0\u7d22\u7ed3\u679c\u3001prompt \u548c\u6a21\u578b\u8f93\u51fa\u653e\u5728\u540c\u4e00\u4e2a\u9875\u9762\u4e2d\u3002\u8c03\u7528\u8bb0\u5f55\u53ea\u80fd\u8bf4\u660e\u7cfb\u7edf\u5f53\u65f6\u4f7f\u7528\u4e86\u54ea\u4efd\u8d44\u6599\uff0c\u4e0d\u80fd\u5224\u65ad\u8fd9\u4efd\u8d44\u6599\u73b0\u5728\u662f\u5426\u4ecd\u7136\u6709\u6548\u3002\u5982\u679c\u77e5\u8bc6\u5e93\u6ca1\u6709\u8bb0\u5f55\u201c\u661f\u6cb3\u79d1\u6280\u5dee\u65c5\u5236\u5ea6\u201d\u7684\u751f\u6548\u65e5\u671f\u3001\u5931\u6548\u65e5\u671f\u6216\u5f53\u524d\u7248\u672c\uff0c\u5e73\u53f0\u5c31\u65e0\u6cd5\u81ea\u52a8\u5224\u65ad\u5b83\u662f\u5426\u5df2\u7ecf\u8fc7\u671f\u3002\u5224\u65ad\u56de\u7b54\u662f\u5426\u6b63\u786e\uff0c\u4ecd\u8981\u4f9d\u8d56\u7ecf\u8fc7\u786e\u8ba4\u7684\u6709\u6548\u6587\u6863\u3001\u53c2\u8003\u7b54\u6848\u3001\u4ee3\u7801\u89c4\u5219\u6216\u4eba\u5de5\u590d\u6838\u3002<\/p>\n<p>\u9009\u578b\u65f6\u5148\u786e\u8ba4\u79c1\u6709\u95ee\u9898\u548c\u6587\u6863\u7247\u6bb5\u80fd\u5426\u53d1\u9001\u5230\u7b2c\u4e09\u65b9\u3001\u662f\u5426\u9700\u8981\u81ea\u6258\u7ba1\u3001TypeScript SDK \u80fd\u5426\u8986\u76d6\u5f53\u524d\u8c03\u7528\u94fe\u3001\u91c7\u6837\u6210\u672c\u5982\u4f55\u63a7\u5236\uff0c\u4ee5\u53ca\u539f\u59cb trace \u80fd\u5426\u5bfc\u51fa\u3002\u65e0\u8bba\u4f7f\u7528\u54ea\u79cd\u5e73\u53f0\uff0c\u4e1a\u52a1\u5b57\u6bb5\u90fd\u5e94\u4fdd\u6301\u7a33\u5b9a\u5e76\u8bb0\u5f55\u8bed\u4e49\u7ea6\u5b9a\u7248\u672c\uff0c\u907f\u514d\u5e73\u53f0\u5347\u7ea7\u540e\u65e0\u6cd5\u6bd4\u8f83\u5386\u53f2\u6570\u636e\u3002<\/p>\n<h2>\u6210\u672c\u4e0e\u8fb9\u754c<\/h2>\n<p>RAG \u7684\u6210\u672c\u4e0d\u53ea\u6709 LLM \u5904\u7406\u6587\u672c\u6240\u6d88\u8017\u7684 token\u3002\u5efa\u7acb\u7d22\u5f15\u9700\u8981\u89e3\u6790\u6587\u6863\u5e76\u751f\u6210\u5411\u91cf\uff0c\u7528\u6237\u67e5\u8be2\u65f6\u8fd8\u8981\u5b8c\u6210\u68c0\u7d22\u548c\u6392\u5e8f\u3002\u5982\u679c\u5141\u8bb8\u6a21\u578b\u5faa\u73af\u8c03\u7528\u5de5\u5177\uff0c\u540c\u4e00\u4e2a\u95ee\u9898\u53ef\u80fd\u89e6\u53d1\u591a\u6b21\u68c0\u7d22\u548c\u6a21\u578b\u8c03\u7528\u3002<\/p>\n<p>Top-K \u592a\u5c0f\u5bb9\u6613\u6f0f\u6389\u6b63\u786e\u7247\u6bb5\uff0c\u592a\u5927\u5219\u4f1a\u628a\u66f4\u591a\u65e0\u5173\u5185\u5bb9\u4ea4\u7ed9\u6a21\u578b\u3002\u6587\u672c\u7247\u6bb5\u592a\u5c0f\uff0c\u6761\u4ef6\u548c\u7ed3\u8bba\u53ef\u80fd\u88ab\u62c6\u5f00\uff1b\u7247\u6bb5\u592a\u5927\uff0c\u53c8\u4e0d\u5bb9\u6613\u51c6\u786e\u5b9a\u4f4d\u7b54\u6848\u3002\u6bcf\u589e\u52a0\u4e00\u4e2a\u5411\u91cf\u6570\u636e\u5e93\u3001Embedding \u6a21\u578b\u6216\u91cd\u6392\u6a21\u578b\uff0c\u90fd\u5e94\u8be5\u7528\u6d4b\u8bd5\u7ed3\u679c\u8bc1\u660e\u5b83\u89e3\u51b3\u4e86\u5177\u4f53\u95ee\u9898\u3002<\/p>\n<p>\u8fd9\u4e2a Demo \u6ca1\u6709\u52a0\u5165 OCR\u3001\u72ec\u7acb\u91cd\u6392\u6a21\u578b\u3001\u8fd1\u4f3c\u6700\u8fd1\u90bb\u7d22\u5f15\uff08Approximate Nearest Neighbor\uff0cANN\uff09\u3001\u6743\u9650\u7cfb\u7edf\u548c\u591a\u79df\u6237\u80fd\u529b\uff0c\u4e5f\u8fd8\u6ca1\u6709\u8bc4\u4f30\u8fd0\u884c\u5668\u3001\u6301\u4e45\u5316 trace\u3001\u8d28\u91cf\u770b\u677f\u548c\u544a\u8b66\u3002\u8fd9\u4e2a\u9879\u76ee\u4e3b\u8981\u7528\u4e8e\u7406\u89e3 RAG \u6d41\u7a0b\u548c\u5236\u4f5c\u5c0f\u89c4\u6a21\u539f\u578b\uff1b\u8fdb\u5165\u751f\u4ea7\u73af\u5883\u524d\uff0c\u8fd8\u8981\u6839\u636e\u5b9e\u9645\u6570\u636e\u91cf\u3001\u8d28\u91cf\u76ee\u6807\u548c\u6743\u9650\u8981\u6c42\u8865\u9f50\u76f8\u5e94\u80fd\u529b\u3002<\/p>\n<h2>\u603b\u7ed3<\/h2>\n<p>RAG \u7531\u68c0\u7d22\u548c\u751f\u6210\u4e24\u4e2a\u73af\u8282\u7ec4\u6210\uff0c\u5411\u91cf\u6570\u636e\u5e93\u53ea\u662f\u68c0\u7d22\u5c42\u7684\u4e00\u79cd\u5e38\u89c1\u5b9e\u73b0\u3002<\/p>\n<p>Embedding \u8d1f\u8d23\u628a\u6587\u672c\u53d8\u6210\u53ef\u6bd4\u8f83\u7684\u5411\u91cf\uff1bChunking \u51b3\u5b9a\u53ef\u68c0\u7d22\u7247\u6bb5\u7684\u9897\u7c92\u5ea6\uff1b\u5411\u91cf\u641c\u7d22\u64c5\u957f\u8bed\u4e49\uff0cBM25 \u64c5\u957f\u7cbe\u786e\u8bcd\u9879\uff1b\u878d\u5408\u4e0e Top-K \u51b3\u5b9a\u54ea\u4e9b\u8d44\u6599\u8fdb\u5165\u4e0a\u4e0b\u6587\uff1bLLM \u5219\u8d1f\u8d23\u5728\u660e\u786e\u8fb9\u754c\u5185\u751f\u6210\u7b54\u6848\u548c\u5f15\u7528\u3002<\/p>\n<p>\u53ef\u9760\u7684 RAG \u7cfb\u7edf\u9700\u8981\u7ed9\u51fa\u7b54\u6848\uff0c\u4e5f\u8981\u4fdd\u7559\u68c0\u7d22\u7ed3\u679c\u3001\u6392\u5e8f\u5206\u548c\u5f15\u7528\u4f4d\u7f6e\uff0c\u65b9\u4fbf\u7528\u6237\u6838\u5bf9\u3002\u8d44\u6599\u4e0d\u8db3\u65f6\uff0c\u7cfb\u7edf\u5e94\u8be5\u660e\u786e\u62d2\u7edd\u56de\u7b54\u3002\u672c\u6587\u7684\u5b9e\u6d4b\u4e2d\uff0c\u9ed8\u8ba4\u6df7\u5408\u6743\u91cd\u66fe\u628a\u9519\u8bef\u7247\u6bb5\u6392\u5230\u7b2c\u4e00\uff0c\u8bf4\u660e\u68c0\u7d22\u8d28\u91cf\u9700\u8981\u5355\u72ec\u6d4b\u91cf\uff0c\u53c2\u6570\u4e5f\u8981\u6839\u636e\u8bc4\u4f30\u7ed3\u679c\u8c03\u6574\u3002\u4e0a\u7ebf\u540e\u8fd8\u8981\u7ed3\u5408 trace\u3001\u770b\u677f\u3001\u544a\u8b66\u548c\u62bd\u6837\u590d\u6838\u7ee7\u7eed\u53d1\u73b0\u65b0\u95ee\u9898\uff1b\u4e00\u6b21\u56de\u7b54\u6d41\u7545\u3001\u5f15\u7528\u683c\u5f0f\u6b63\u786e\uff0c\u4ecd\u4e0d\u8db3\u4ee5\u8bc1\u660e\u7cfb\u7edf\u53ef\u9760\u3002<\/p>\n<p>\u5728\u66f4\u590d\u6742\u7684\u7cfb\u7edf\u4e2d\uff0c\u5982\u679c\u56fa\u5b9a\u7684\u4e00\u6b21\u68c0\u7d22\u4e0d\u8db3\u4ee5\u5b8c\u6210\u4efb\u52a1\uff0c\u53ef\u4ee5\u628a BM25\u3001\u5411\u91cf\u641c\u7d22\u3001SQL\u3001grep \u548c\u6587\u4ef6\u8bfb\u53d6\u5206\u522b\u505a\u6210 Agent \u5de5\u5177\u3002\u5f53\u524d Demo \u6ca1\u6709\u5b9e\u73b0\u8fd9\u5957\u591a\u5de5\u5177\u67b6\u6784\uff0c\u53ea\u63d0\u4f9b\u4e86\u4e00\u4e2a\u53ef\u518d\u6b21\u8c03\u7528\u7684\u6df7\u5408\u68c0\u7d22\u5de5\u5177\u3002\u65e0\u8bba\u91c7\u7528\u54ea\u79cd\u65b9\u5f0f\uff0c\u6700\u7ec8\u90fd\u8981\u68c0\u67e5\u7cfb\u7edf\u80fd\u5426\u5728\u53ef\u63a5\u53d7\u7684\u6210\u672c\u548c\u6743\u9650\u8fb9\u754c\u5185\u627e\u5230\u6b63\u786e\u8d44\u6599\uff0c\u5e76\u5fe0\u5b9e\u5730\u4f7f\u7528\u8fd9\u4e9b\u8d44\u6599\u3002<\/p>\n<h3>\u6211\u7684\u5176\u4ed6 AI Agent \u6587\u7ae0<\/h3>\n<ul>\n<li><a href=\"https:\/\/link.segmentfault.com\/?enc=ywdyVtk1zGsrDVeZ14psPA%3D%3D.7N4Pw%2Fp3aCSM8VSBdYBFrJxmtxxQVDSQ4Hftt071E6hNbgmwAaj6JaCvNyzIYACIyeF%2Bow7AIz%2Fro7wTt%2BKLvR0bRzhlRIBkcuYDBWucvXg%3D\" rel=\"nofollow noopener noreferrer\" target=\"_blank\">\u6398\u91d1\u5c0f\u518c&#8211;\u4ece\u96f6\u6253\u9020\u4e00\u4e2a AI Agent CLI<\/a><\/li>\n<li><a href=\"https:\/\/link.segmentfault.com\/?enc=3RbVxjTryMsYRDpmDY7vUg%3D%3D.AvXeFlIXxQG2hgnthAK%2F1SmpP5yCEGuY8nrDzixNEToQuOgpldVXCglJ6H3Z96%2FjGulMuI1DPzfgd3Pllg8CwBAkxZK4z8wWXzhsIB2Hn5bH%2F%2BNmx8pa0eh20A0%2Barp7\" rel=\"nofollow noopener noreferrer\" target=\"_blank\">AI \u662f\u600e\u4e48\u64cd\u4f5c\u6d4f\u89c8\u5668\u7684\u2014\u2014browser use \u5b9e\u73b0\u539f\u7406<\/a><\/li>\n<li><a href=\"https:\/\/link.segmentfault.com\/?enc=uRSGuRGLV9mssf7N0w4ltg%3D%3D.JpIEg44ioKbtdhtFPO5C9GiwQP6ic4nGuXV14SqGtAhFPKQyXPWEc10nCJeMyGEVY5WBsdjTAqHKV3GX4VtRqjh6qWp37evA8A2IFTtWWlp%2BfByja4j7N7N3w8iZ%2FobOH1ep5QJA0khgeb9ZrmdZa6Thge5fnSe9MNrRpTufrb%2Brk%2FwQMj8VP53ltriYO1fyCYOFb9VQhfgLFpr8sWtVfQ%3D%3D\" rel=\"nofollow noopener noreferrer\" target=\"_blank\">\u5de5\u5177\u585e\u6ee1\u4e0a\u4e0b\u6587\u7a97\u53e3\u600e\u4e48\u529e\uff1f\u6df1\u5ea6\u62c6\u89e3 AI Agent Tool Search \u6309\u9700\u52a0\u8f7d\u5b9e\u73b0\u539f\u7406<\/a><\/li>\n<li><a href=\"https:\/\/link.segmentfault.com\/?enc=qclzUJ0UOyZyOiI0E9U8yw%3D%3D.Nn5XIK9lPm1xCAtWkfkyXyWuV9GDNI5L8tIcTygdfx5qxMd53cNSdsL2G9N3Mgro6CNmcpL9w54Elafpo7%2BJKFhh7M6eeEnBxJn3w9Stw%2FP%2BmT41%2FaiCTbMDEWORlPJm\" rel=\"nofollow noopener noreferrer\" target=\"_blank\">\u8ba9 AI Agent \u7cfb\u7edf\u81ea\u5df1\u53d1\u73b0 bug\u3001\u81ea\u5df1\u63d0\u4fee\u590d PR\uff1a\u81ea\u6211\u8fdb\u5316\u7684 Harness<\/a><\/li>\n<li><a href=\"https:\/\/link.segmentfault.com\/?enc=a7oRta%2Fd1K44AS%2BiSYLxYw%3D%3D.5TLuHTPNI6UANLL%2Fn1z%2FmpAs5jzR9tgwyUrRtwRUkK3EZCj8pi6%2BiWhvwbUMdgAo%2BxTbzRCIfLRvmq5VDU%2FaXfITToRBiMkcOfU%2BVNFIkASsiu9MvsutzMZRFjEEZjSd\" rel=\"nofollow noopener noreferrer\" target=\"_blank\">AI Agent \u5de5\u7a0b\u5e08\u5165\u95e8\u6307\u5357<\/a><\/li>\n<li>\n<p><a href=\"https:\/\/link.segmentfault.com\/?enc=dzefoxxm3SrRJCdNoQ%2FrLA%3D%3D.ImjeNT7v22yBOmsjmB6KdpQHl5E6onqTIIrYQSSyhZEml3hnVqh0QOEMOIzLj1Q%2BSB8Ght4RuMSMoXtx8DJr5wmaRs0dllRvW%2BY1n%2Bo%2B448Twi3UXsB7QiYy%2FQ5GepwU\" rel=\"nofollow noopener noreferrer\" target=\"_blank\">\u5982\u4f55\u4ece\u96f6\u5f00\u59cb\u5b9e\u73b0\u4e00\u4e2a AI Agent CLI<\/a><\/p>\n<h2>\u53c2\u8003\u8d44\u6599<\/h2>\n<\/li>\n<li>Lewis et al., 2020, <a href=\"https:\/\/link.segmentfault.com\/?enc=qNpdYlg4y29g9fJSvXJIdQ%3D%3D.O7Vcba9IwTyE2IOjowPoJHhDAG%2BeXZGRPlp8BEvREpDRep9AAEL2P0PhjJ4Clfzq\" rel=\"nofollow noopener noreferrer\" target=\"_blank\">Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks<\/a><\/li>\n<li>Microsoft Learn, <a href=\"https:\/\/link.segmentfault.com\/?enc=SrcL4IBZJwlClKoaAerBiA%3D%3D.UFC0tI%2BJfKVri0Sr2FhEuLlimdXIczW%2BSok9k1e9x0kZMjz8DLuHNV9r15WUmqnyLWAfjde4aohcEgmsQ%2BwB9Gg1pbgw%2BWg6wFKYPk0nabEg6DuhRhy%2BHU2o8%2FkkpUM5ARPLjJiezE833X62J14Ehe%2B1de9Hnk1zz3oRogSiQIY%3D\" rel=\"nofollow noopener noreferrer\" target=\"_blank\">Retrieval-Augmented Generation with Azure Files<\/a><\/li>\n<li>Azure Databricks, <a href=\"https:\/\/link.segmentfault.com\/?enc=DM6oj5PLUtLc1TAo9ILE%2FQ%3D%3D.v3AW0dbL1wRJ24vKkH9sXhdweojGqR2XAoVxoT%2FyuXTRtWOPyPGM6rzLmI5Ltqz8z3wtjVCJ%2Bx3F0OOMumQBYH0y0n7GL%2FutJNCfJ6S3zykC4L%2FOBtDBDbizYLLUpVKI01obfAjn049NLyH3qJIAiw%3D%3D\" rel=\"nofollow noopener noreferrer\" target=\"_blank\">Build an unstructured data pipeline for RAG<\/a><\/li>\n<li>LangChain, <a href=\"https:\/\/link.segmentfault.com\/?enc=7C3%2Blw8DU8hBGtbo9IzodQ%3D%3D.MAitPceYtyt4dc3saYOJPbyCjooCUvOKg0SaEEqTxPxcgnmI%2FTzEtKZqG4cR0kLyhf5peFrorQg5ILwVFt6dag886GfiiqSaCsDG9KIkSfg%3D\" rel=\"nofollow noopener noreferrer\" target=\"_blank\">Text splitter integrations<\/a><\/li>\n<li>Hugging Face, <a href=\"https:\/\/link.segmentfault.com\/?enc=KjNavZP5Ti9QTim3Aut%2Bow%3D%3D.0lEWDBSimVD1tpmQ08to%2FiUUNH12tAqeBzI1OjQjjF%2BsrEaGvGnXLmofAFvgfgIdKm%2FDOpxQRQ0B4BAY5ula6A%3D%3D\" rel=\"nofollow noopener noreferrer\" target=\"_blank\">Transformers.js 3.8.1 documentation<\/a><\/li>\n<li>Hugging Face, <a href=\"https:\/\/link.segmentfault.com\/?enc=9uFo5TpP0ORxib2xXVNs6w%3D%3D.N0Tb0BZOL963XEXnj7tnqLNY%2BD9xs6JaMixgDz5hDYRzfwoA7YiXqgirm6Ke%2FCqXaDCWtQt%2B0jXdsOLveCj3BA%3D%3D\" rel=\"nofollow noopener noreferrer\" target=\"_blank\"><code>intfloat\/multilingual-e5-small<\/code> model card<\/a><\/li>\n<li>Ollama, <a href=\"https:\/\/link.segmentfault.com\/?enc=IbRJdJRdLSltN3FWsS%2Bfow%3D%3D.EcSx62xT5m8BRakeAPOxiJTa8D5b%2FxlCKs%2B%2FFy%2FAAIJjUFxz7rGX9PEnQ3GZQtXz\" rel=\"nofollow noopener noreferrer\" target=\"_blank\">Embedding models<\/a><\/li>\n<li>Robertson &#038; Zaragoza, 2009, <a href=\"https:\/\/link.segmentfault.com\/?enc=dMMolpjr1732YbhhOoz%2F1w%3D%3D.eIhs3KC%2FwYOtDDMShcJTDAV%2BNZRc%2FdaIvMOd7lnhb7Wm0jWQgjX7uQEWN4QrAXgVSh4fOM9vEBYf04713HV2mA%3D%3D\" rel=\"nofollow noopener noreferrer\" target=\"_blank\">The Probabilistic Relevance Framework: BM25 and Beyond<\/a><\/li>\n<li>SQLite, <a href=\"https:\/\/link.segmentfault.com\/?enc=%2BlQ2pe3iTPYg0hYrPwUkkQ%3D%3D.NBkLxzK2Etvw40VMRNG9gY0RUZYPD3ssPjF7z8BckIGIF2HXz5W%2F1Lb1exwthd1M\" rel=\"nofollow noopener noreferrer\" target=\"_blank\">FTS5 Extension<\/a><\/li>\n<li>Elasticsearch, <a href=\"https:\/\/link.segmentfault.com\/?enc=XJnW14H6UkIYvtVvVGYz8w%3D%3D.ze8rbnf3mmg7cCHfYaiMId6lVP54YvFXqWkw7hPmnn9w1a1SDpkQ5l%2B%2B0JnYdocqkmrouMg3UsQOhfW8LizFc%2FDftKxWpfv%2BbMLZkFGszeE%3D\" rel=\"nofollow noopener noreferrer\" target=\"_blank\">Similarity settings<\/a><\/li>\n<li>PostgreSQL, <a href=\"https:\/\/link.segmentfault.com\/?enc=tRbkfpqaGTwNJEkpjgB87w%3D%3D.o%2B76FnKO9OkWSk%2FxlXIfAzvxyzaN0AJZVhgJfFTjzb53xh2QyaMD1XSXAReX3YQtPD0%2BeBv6CosK63p5KnSeiQ%3D%3D\" rel=\"nofollow noopener noreferrer\" target=\"_blank\">Full Text Search<\/a><\/li>\n<li>Cormack, Clarke &#038; B\u00fcttcher, 2009, <a href=\"https:\/\/link.segmentfault.com\/?enc=%2FxC3SvLfuPJ%2BNamxaWvzAg%3D%3D.L3lGbkgRQQM8jEYzilc9n3fcVPDEMmOqgtlud3mP6i8AbbLF8WwtbgJR2XLdv2fyZkZVPet4q%2FlDok1cTtPo0A%3D%3D\" rel=\"nofollow noopener noreferrer\" target=\"_blank\">Reciprocal Rank Fusion Outperforms Condorcet and Individual Rank Learning Methods<\/a><\/li>\n<li>sqlite-vec, <a href=\"https:\/\/link.segmentfault.com\/?enc=AYMGxBJx9F2doMhwDUiqXg%3D%3D.pBuCdy0AMr5Ya%2Fh%2F2y8zgCgSdmSZBvrDAQ%2Fwg0Zre4Ol%2FTIG5JrF8vVxtMlBWkIM\" rel=\"nofollow noopener noreferrer\" target=\"_blank\">SQLite vector search extension<\/a><\/li>\n<li>pgvector, <a href=\"https:\/\/link.segmentfault.com\/?enc=%2Fv4HlRBjNAo2viIc%2FbHLog%3D%3D.15%2FQZBtbvAbdU80HDGc1ybl3PvmFs8xZrlqWgn%2FOsgn0FVaNQjcGhaCDhvAJI3V4\" rel=\"nofollow noopener noreferrer\" target=\"_blank\">Open-source vector similarity search for PostgreSQL<\/a><\/li>\n<li>Boris Cherny, 2025-02-24, <a href=\"https:\/\/link.segmentfault.com\/?enc=L3MLF1ApT47Gg2QmRRL8YQ%3D%3D.jlFfS5iJmAvqJqYDhX1QKZY8lKG%2BfkzVzQy8E%2B9bNyy3DIF7Hnu1rUKtB9pW%2Fzl0him48NlAAx3%2BCtvd%2B5amCg%3D%3D\" rel=\"nofollow noopener noreferrer\" target=\"_blank\">Hacker News comment on Claude Code and Agentic Search<\/a><\/li>\n<li>Latent Space, 2025-05, <a href=\"https:\/\/link.segmentfault.com\/?enc=SMJijC4KsZIbPgSRZmojdQ%3D%3D.15VpAjG8KIUJSZrqhRP5P2JC9CxJmCBs8WqDeFdZ%2B%2FKezA1d5Lq5meBvANQjwYPv\" rel=\"nofollow noopener noreferrer\" target=\"_blank\">Claude Code: Anthropic&#8217;s Agent in Your Terminal<\/a><\/li>\n<li>Anthropic, <a href=\"https:\/\/link.segmentfault.com\/?enc=NHvrhhyoPDvcK5VSoBq4Zw%3D%3D.UOB0NjsqfcwqxGH2gpMLreS6QIPI7l7BFw0x1%2FArrNblDj%2Bfd4J18chKh1rQ4lWd\" rel=\"nofollow noopener noreferrer\" target=\"_blank\">Claude Code tools reference<\/a><\/li>\n<li>Cursor, 2025-11-06, <a href=\"https:\/\/link.segmentfault.com\/?enc=8tW1slMGfai24YgY5kEg%2Fw%3D%3D.2rTTAhglpEpZmpXRrDIMeIUqP6e3tOzJMYT81Y%2FU6NQ9%2B7v4IaIEgqP2QQYQKLtj\" rel=\"nofollow noopener noreferrer\" target=\"_blank\">Improving agent with semantic search<\/a><\/li>\n<li>Cursor, 2026-01-27, <a href=\"https:\/\/link.segmentfault.com\/?enc=VfYJ0EvflDjMaj62niO9uQ%3D%3D.rWxY5bAf0LK92xELZ8QrpldRMVMD%2FvGNqDcyukCL%2BuNTCpnN%2Fy3BBa%2F3LVs1hF1l1fB%2BzBPYb0JW%2FnoMVOpbmw%3D%3D\" rel=\"nofollow noopener noreferrer\" target=\"_blank\">Securely indexing large codebases<\/a><\/li>\n<li>GitHub Docs, <a href=\"https:\/\/link.segmentfault.com\/?enc=E3Iq1AXW4%2FGdfrhCAyvMGw%3D%3D.fUsuqTHboxJ%2F0QF8UEa3RVFTEC8Ojqia8uJA56TLEwsIi%2FqBHT%2BBLRfMh%2BUcTIkIqS7QYdg8%2BzLR1%2BZSvqk06ZMFRofjldl2K%2B9sK7cfzT0%3D\" rel=\"nofollow noopener noreferrer\" target=\"_blank\">Indexing repositories for GitHub Copilot<\/a><\/li>\n<li>Es et al., EACL 2024, <a href=\"https:\/\/link.segmentfault.com\/?enc=XdFabE5J7b%2FXOT9f78BMlw%3D%3D.mBGmlG0RuzVHytL5asXKYdqqb02gozWb%2F6EaoaCLnMtWCaSWz91bYX0sONWclCmN\" rel=\"nofollow noopener noreferrer\" target=\"_blank\">RAGAS: Automated Evaluation of Retrieval Augmented Generation<\/a><\/li>\n<li>RAGAS, <a href=\"https:\/\/link.segmentfault.com\/?enc=jTSjUwvUNhwI6QTU2H1aTg%3D%3D.SRp%2BUfCsc5mIdlyh5XNwz5DxOaKf2h0boBCAfDoLNcktUN7BdmDsUIrwPOhonQIWCzMcTn3RTgSlFxHP6rl3abo%2FV%2B55vrUjnK2AhcDzi8o8wnKr0uwucTt8pXh1uG%2FR\" rel=\"nofollow noopener noreferrer\" target=\"_blank\">Faithfulness<\/a>\u3001<a href=\"https:\/\/link.segmentfault.com\/?enc=5sANPR%2BuK9rF4uNSN6hHCQ%3D%3D.D57frXiE1He44bNQB7H6BXDbay%2FYCvF26W%2BCCeNqct1uE3%2Bo7k208uf3c4UC78pDCxGVYTtBZiWCxD%2BHgaUdaD2xescaEAJZODIFPLAHvbpf8p9AWt2f1kY3fRFUMVNr\" rel=\"nofollow noopener noreferrer\" target=\"_blank\">Context Precision<\/a> \u4e0e <a href=\"https:\/\/link.segmentfault.com\/?enc=Ev4dhS8bJR0pf%2BOBWNQU%2Bg%3D%3D.xVzByDwl4ksdR8EY4%2B3cfJHwXNkZY%2Fa2yUQfd8jKu49M%2BjwXhcScgOBpsCBe5l9%2BJxKffhyikpLHOk7y7VYDqpWw3pzkO48pneWtfSUsNtF6qgl%2FOcruIcA%2BHGi6FQMI\" rel=\"nofollow noopener noreferrer\" target=\"_blank\">Context Recall<\/a><\/li>\n<li>Manning, Raghavan &#038; Sch\u00fctze, <a href=\"https:\/\/link.segmentfault.com\/?enc=7OWMo89%2FjKlCz55MUYQdhA%3D%3D.nxBoS9JexIUc6QHD%2BlwQBQvzTu%2B3uqLLl5%2BJILw8PuNXFHHq5A227vHCKTOVEITvZD0ZA%2BKrwqlpcqYs9P8jPzL%2FmbDS1BrxFtZftqklbhYetttw0Cv%2FTJ9jqMPbwx5S\" rel=\"nofollow noopener noreferrer\" target=\"_blank\">Introduction to Information Retrieval: Evaluation<\/a><\/li>\n<li>Zheng et al., NeurIPS 2023, <a href=\"https:\/\/link.segmentfault.com\/?enc=InIJO%2FPn4eJC%2FkMQETsejA%3D%3D.LNFIOfXK7axk4Wez7Yrl0gQOQSzBMozHPnRory6Z2RnnXRZCbvhDo4wZYo8bYRtrXahI%2B3sbJbnWhYFAgMh%2F4pUMBcy40ON4cI0xz3jFQeIiDqndEjS5s1IOFT1tHTX%2BnP6f0GNSR20ENMc1il8fKQJifUFR7Y2nhR%2F0KfyKTX39MjWwmT6OUw5d6SMLR1VE\" rel=\"nofollow noopener noreferrer\" target=\"_blank\">Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena<\/a><\/li>\n<li>OpenInference, <a href=\"https:\/\/link.segmentfault.com\/?enc=eu33BgrZ1zl2Iq2jfmMuYg%3D%3D.OxJoLjAJVo5855lAdYYBQSTxts3AE6J5AdAh9PsZu27mpdbodnc1%2FIw%2BaZ%2FRjajB\" rel=\"nofollow noopener noreferrer\" target=\"_blank\">Specification<\/a> \u4e0e <a href=\"https:\/\/link.segmentfault.com\/?enc=yhe4U%2B5yLK8IGoYO6J643A%3D%3D.AyXaEobGFg3zmhAoPzVGBSIOIkxBJNM35Add0v5aEo52L%2B10gCDY4fdbsBM130YanEQQsR9MlOKCNK9DoT78el5b6Vqr7ZURXjnm8Q9%2FLQE%3D\" rel=\"nofollow noopener noreferrer\" target=\"_blank\">Semantic Conventions<\/a><\/li>\n<li>LangSmith, <a href=\"https:\/\/link.segmentfault.com\/?enc=yxFuwf2we8gIXMvE1dtYzw%3D%3D.X0A1razVy8q8841fgo3EMFoHIpJ%2BhXWmPKdFYY4A97sWrYM8IhD%2F7wkBTlMqtFf6%2FrBEoO3%2FlBOYwP%2FlqAkJyQ%3D%3D\" rel=\"nofollow noopener noreferrer\" target=\"_blank\">Evaluation concepts<\/a>\u3001<a href=\"https:\/\/link.segmentfault.com\/?enc=kHDwbe6Dt6SasM%2B1f34PqA%3D%3D.nMK%2Bc7INPKbqitYo%2FquoCPSfTWXE%2FgOMxDB2M6FQXvDpaBTyFnlHmf3G9xXoWYA819Ur5Mrf2Ff0Ejf4%2BqjP9w%3D%3D\" rel=\"nofollow noopener noreferrer\" target=\"_blank\">RAG observability tutorial<\/a> \u4e0e <a href=\"https:\/\/link.segmentfault.com\/?enc=8ifjA49i5dbBAlgUXtIhbQ%3D%3D.lYgUH%2FpKaPDLVi7quUHVujy59m4%2FzYoOQRx7YJkYB%2FLkBYpo1%2FFNVe4fLjfVh%2FEu\" rel=\"nofollow noopener noreferrer\" target=\"_blank\">Alerts<\/a><\/li>\n<li>Arize Phoenix, <a href=\"https:\/\/link.segmentfault.com\/?enc=OOBxaRWfZQOXJgi7yGmDgw%3D%3D.MsibRrinBVaY2pK44kGOCqnTMCpjHFM9BTiAW3hYTrKk8JJcOFbUYZJC4ocZpAg1\" rel=\"nofollow noopener noreferrer\" target=\"_blank\">Tracing<\/a> \u4e0e <a href=\"https:\/\/link.segmentfault.com\/?enc=3o3LTq3qanijW6UGmT%2FMJQ%3D%3D.GsUvkUY5om%2BycZVFTAURMYqEKdf%2Fyz9JY7k0k3PcKWLRrchZ6qqAlcNQpDq0PkQCDcx5C%2FtiXnTef%2F0QiyOEzw%3D%3D\" rel=\"nofollow noopener noreferrer\" target=\"_blank\">Evaluate RAG<\/a><\/li>\n<\/ul>\n<p><a href=\"https:\/\/segmentfault.com\/a\/1190000048055387\">https:\/\/segmentfault.com\/a\/1190000048055387<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>RAG \u662f\u4ec0\u4e48 RAG\uff08Retrieval-Augmented Generation\uff0c\u68c0\u7d22\u589e\u5f3a\u751f\u6210\uff09\u662f\u4e00\u79cd\u5148\u68c0 [&hellip;]<\/p>\n","protected":false},"author":7,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_jetpack_memberships_contains_paid_content":false,"footnotes":""},"categories":[4],"tags":[],"class_list":["post-10021418","post","type-post","status-publish","format-standard","hentry","category-share"],"jetpack_sharing_enabled":true,"jetpack_featured_media_url":"","_links":{"self":[{"href":"https:\/\/itmax.cn\/index.php?rest_route=\/wp\/v2\/posts\/10021418","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/itmax.cn\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/itmax.cn\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/itmax.cn\/index.php?rest_route=\/wp\/v2\/users\/7"}],"replies":[{"embeddable":true,"href":"https:\/\/itmax.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=10021418"}],"version-history":[{"count":0,"href":"https:\/\/itmax.cn\/index.php?rest_route=\/wp\/v2\/posts\/10021418\/revisions"}],"wp:attachment":[{"href":"https:\/\/itmax.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=10021418"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/itmax.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=10021418"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/itmax.cn\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=10021418"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}