{"id":90460,"date":"2026-08-05T10:40:34","date_gmt":"2026-08-05T02:40:34","guid":{"rendered":"https:\/\/www.wsisp.com\/helps\/90460.html"},"modified":"2026-08-05T10:40:34","modified_gmt":"2026-08-05T02:40:34","slug":"10-%e4%bd%bf%e7%94%a8reranker%e6%8f%90%e5%8d%87rag%e5%9b%9e%e7%ad%94%e5%87%86%e7%a1%ae%e7%8e%87","status":"publish","type":"post","link":"https:\/\/www.wsisp.com\/helps\/90460.html","title":{"rendered":"10-\u4f7f\u7528Reranker\u63d0\u5347RAG\u56de\u7b54\u51c6\u786e\u7387"},"content":{"rendered":"

\u4f7f\u7528 Reranker（\u91cd\u6392\u6a21\u578b）\u63d0\u5347 RAG \u77e5\u8bc6\u5e93\u56de\u7b54\u51c6\u786e\u7387<\/h2>\n

\u7cfb\u5217：\u4ece\u96f6\u6784\u5efa\u4f01\u4e1a RAG \u77e5\u8bc6\u5e93（\u7b2c 10 \u7bc7）<\/p>\n

1. \u4e3a\u4ec0\u4e48\u53ec\u56de\u540e\u8fd8\u8981\u91cd\u6392<\/h3>\n

\u7b2c\u4e00\u9636\u6bb5\u68c0\u7d22\u8ffd\u6c42\u201c\u522b\u6f0f\u6389\u201d，\u901a\u5e38\u5feb\u901f\u53d6 20\u2014100 \u4e2a\u5019\u9009；\u751f\u6210\u6a21\u578b\u4e0a\u4e0b\u6587\u5374\u53ea\u9002\u5408\u5c11\u91cf\u9ad8\u8d28\u91cf\u8bc1\u636e\u3002<\/p>\n

\u5927\u8303\u56f4\u53ec\u56de（\u9ad8 Recall） \u2192 Reranker \u7cbe\u6392（\u9ad8 Precision）
\n\u2192 \u53bb\u91cd\u4e0e\u4e0a\u4e0b\u6587\u9884\u7b97 \u2192 \u751f\u6210<\/p>\n

Embedding \u5e38\u5c06\u95ee\u9898\u548c\u6587\u6863\u5206\u522b\u7f16\u7801，\u518d\u6bd4\u8f83\u5411\u91cf；Cross-Encoder \u7c7b Reranker \u540c\u65f6\u8bfb\u53d6\u201c\u95ee\u9898 + \u5019\u9009\u201d，\u80fd\u8fdb\u884c\u66f4\u7ec6\u81f4\u7684\u76f8\u5173\u6027\u5224\u65ad，\u4f46\u8ba1\u7b97\u66f4\u6162\u3002<\/p>\n

2. \u91cd\u6392\u63a5\u53e3\u5fc5\u987b\u4fdd\u7559\u5019\u9009\u8eab\u4efd<\/h3>\n

from<\/span> dataclasses import<\/span> dataclass
\nfrom<\/span> typing import<\/span> Protocol<\/p>\n

@dataclass<\/span>(<\/span>frozen=<\/span>True<\/span>)<\/span>
\nclass<\/span> Candidate<\/span>:<\/span>
\n chunk_id:<\/span> str<\/span>
\n document_id:<\/span> str<\/span>
\n tenant_id:<\/span> str<\/span>
\n text:<\/span> str<\/span>
\n retrieval_score:<\/span> float<\/span><\/p>\n

@dataclass<\/span>(<\/span>frozen=<\/span>True<\/span>)<\/span>
\nclass<\/span> RerankedCandidate<\/span>:<\/span>
\n candidate:<\/span> Candidate
\n rerank_score:<\/span> float<\/span>
\n original_rank:<\/span> int<\/span><\/p>\n

class<\/span> Reranker<\/span>(<\/span>Protocol)<\/span>:<\/span>
\n def<\/span> score<\/span>(<\/span>self,<\/span> query:<\/span> str<\/span>,<\/span> texts:<\/span> list<\/span>[<\/span>str<\/span>]<\/span>)<\/span> –<\/span>><\/span> list<\/span>[<\/span>float<\/span>]<\/span>:<\/span>
\n """\u8fd4\u56de\u4e0e texts \u4e00\u4e00\u5bf9\u5e94\u7684\u76f8\u5173\u6027\u5206\u6570\u3002"""<\/span><\/p>\n

3. \u5b89\u5168\u7684\u6279\u91cf\u91cd\u6392<\/h3>\n

def<\/span> rerank<\/span>(<\/span>
\n query:<\/span> str<\/span>,<\/span>
\n candidates:<\/span> list<\/span>[<\/span>Candidate]<\/span>,<\/span>
\n reranker:<\/span> Reranker,<\/span>
\n top_n:<\/span> int<\/span> =<\/span> 5<\/span>,<\/span>
\n)<\/span> –<\/span>><\/span> list<\/span>[<\/span>RerankedCandidate]<\/span>:<\/span>
\n if<\/span> not<\/span> query.<\/span>strip(<\/span>)<\/span>:<\/span>
\n raise<\/span> ValueError(<\/span>"\u67e5\u8be2\u4e0d\u80fd\u4e3a\u7a7a"<\/span>)<\/span>
\n if<\/span> not<\/span> 1<\/span> <=<\/span> top_n <=<\/span> 20<\/span>:<\/span>
\n raise<\/span> ValueError(<\/span>"top_n \u5fc5\u987b\u5728 1 \u5230 20 \u4e4b\u95f4"<\/span>)<\/span>
\n if<\/span> not<\/span> candidates:<\/span>
\n return<\/span> [<\/span>]<\/span><\/p>\n

scores =<\/span> reranker.<\/span>score(<\/span>query,<\/span> [<\/span>item.<\/span>text for<\/span> item in<\/span> candidates]<\/span>)<\/span>
\n if<\/span> len<\/span>(<\/span>scores)<\/span> !=<\/span> len<\/span>(<\/span>candidates)<\/span>:<\/span>
\n raise<\/span> RuntimeError(<\/span>"Reranker \u5206\u6570\u6570\u91cf\u4e0e\u5019\u9009\u6570\u91cf\u4e0d\u4e00\u81f4"<\/span>)<\/span>
\n if<\/span> any<\/span>(<\/span>not<\/span> isinstance<\/span>(<\/span>score,<\/span> (<\/span>int<\/span>,<\/span> float<\/span>)<\/span>)<\/span> for<\/span> score in<\/span> scores)<\/span>:<\/span>
\n raise<\/span> RuntimeError(<\/span>"Reranker \u8fd4\u56de\u4e86\u975e\u6570\u503c\u5206\u6570"<\/span>)<\/span><\/p>\n

ranked =<\/span> [<\/span>
\n RerankedCandidate(<\/span>candidate,<\/span> float<\/span>(<\/span>score)<\/span>,<\/span> original_rank)<\/span>
\n for<\/span> original_rank,<\/span> (<\/span>candidate,<\/span> score)<\/span> in<\/span> enumerate<\/span>(<\/span>
\n zip<\/span>(<\/span>candidates,<\/span> scores)<\/span>,<\/span>
\n start=<\/span>1<\/span>,<\/span>
\n )<\/span>
\n ]<\/span>
\n ranked.<\/span>sort(<\/span>
\n key=<\/span>lambda<\/span> item:<\/span> (<\/span>
\n –<\/span>item.<\/span>rerank_score,<\/span>
\n item.<\/span>original_rank,<\/span>
\n item.<\/span>candidate.<\/span>chunk_id,<\/span>
\n )<\/span>
\n )<\/span>
\n return<\/span> ranked[<\/span>:<\/span>top_n]<\/span><\/p>\n

\u539f\u59cb\u6392\u540d\u4f5c\u4e3a\u540c\u5206\u6392\u5e8f\u4f9d\u636e，\u8ba9\u7ed3\u679c\u7a33\u5b9a\u4e14\u53ef\u8bca\u65ad\u3002<\/p>\n

4. \u7528\u786e\u5b9a\u6027 Fake Reranker \u6d4b\u6d41\u7a0b<\/h3>\n

class<\/span> KeywordReranker<\/span>:<\/span>
\n """\u6559\u5b66\u66ff\u8eab，\u53ea\u9a8c\u8bc1\u63a7\u5236\u6d41，\u4e0d\u4ee3\u8868\u771f\u5b9e\u91cd\u6392\u6a21\u578b\u3002"""<\/span><\/p>\n

def<\/span> score<\/span>(<\/span>self,<\/span> query:<\/span> str<\/span>,<\/span> texts:<\/span> list<\/span>[<\/span>str<\/span>]<\/span>)<\/span> –<\/span>><\/span> list<\/span>[<\/span>float<\/span>]<\/span>:<\/span>
\n query_chars =<\/span> set<\/span>(<\/span>query)<\/span>
\n return<\/span> [<\/span>
\n float<\/span>(<\/span>len<\/span>(<\/span>query_chars &<\/span> set<\/span>(<\/span>text)<\/span>)<\/span>)<\/span> \/<\/span> max<\/span>(<\/span>len<\/span>(<\/span>query_chars)<\/span>,<\/span> 1<\/span>)<\/span>
\n for<\/span> text in<\/span> texts
\n ]<\/span><\/p>\n

def<\/span> test_relevant_candidate_moves_forward<\/span>(<\/span>)<\/span> –<\/span>><\/span> None<\/span>:<\/span>
\n candidates =<\/span> [<\/span>
\n Candidate(<\/span>"c1"<\/span>,<\/span> "d1"<\/span>,<\/span> "t1"<\/span>,<\/span> "\u9000\u6b3e\u653f\u7b56\u603b\u89c8"<\/span>,<\/span> 0.95<\/span>)<\/span>,<\/span>
\n Candidate(<\/span>"c2"<\/span>,<\/span> "d2"<\/span>,<\/span> "t1"<\/span>,<\/span> "\u9000\u6b3e\u7533\u8bf7\u9700\u8981\u4eba\u5de5\u5ba1\u6838"<\/span>,<\/span> 0.80<\/span>)<\/span>,<\/span>
\n ]<\/span>
\n result =<\/span> rerank(<\/span>
\n "\u9000\u6b3e\u7533\u8bf7\u5982\u4f55\u5ba1\u6838"<\/span>,<\/span>
\n candidates,<\/span>
\n KeywordReranker(<\/span>)<\/span>,<\/span>
\n top_n=<\/span>1<\/span>,<\/span>
\n )<\/span>
\n assert<\/span> result[<\/span>0<\/span>]<\/span>.<\/span>candidate.<\/span>chunk_id ==<\/span> "c2"<\/span>
\n assert<\/span> result[<\/span>0<\/span>]<\/span>.<\/span>original_rank ==<\/span> 2<\/span><\/p>\n

5. \u771f\u5b9e Cross-Encoder \u7684\u9002\u914d\u793a\u610f<\/h3>\n

class<\/span> SentenceTransformersReranker<\/span>:<\/span>
\n """\u793a\u610f\u9002\u914d\u5668；\u6a21\u578b\u540d\u79f0\u548c\u8bb8\u53ef\u8bc1\u9700\u6309\u5f53\u524d\u5b98\u65b9\u8d44\u6599\u9009\u62e9\u3002"""<\/span><\/p>\n

def<\/span> __init__<\/span>(<\/span>self,<\/span> model)<\/span> –<\/span>><\/span> None<\/span>:<\/span>
\n self.<\/span>model =<\/span> model<\/p>\n

def<\/span> score<\/span>(<\/span>self,<\/span> query:<\/span> str<\/span>,<\/span> texts:<\/span> list<\/span>[<\/span>str<\/span>]<\/span>)<\/span> –<\/span>><\/span> list<\/span>[<\/span>float<\/span>]<\/span>:<\/span>
\n pairs =<\/span> [<\/span>(<\/span>query,<\/span> text)<\/span> for<\/span> text in<\/span> texts]<\/span>
\n raw_scores =<\/span> self.<\/span>model.<\/span>predict(<\/span>pairs)<\/span>
\n return<\/span> [<\/span>float<\/span>(<\/span>value)<\/span> for<\/span> value in<\/span> raw_scores]<\/span><\/p>\n

# \u793a\u4f8b\u521d\u59cb\u5316\u65b9\u5f0f\u9700\u8981\u5b89\u88c5\u5e76\u9501\u5b9a sentence-transformers：<\/span>
\n# from sentence_transformers import CrossEncoder<\/span>
\n# model = CrossEncoder("\u7ecf\u8fc7\u56e2\u961f\u8bc4\u6d4b\u5e76\u6279\u51c6\u7684\u6a21\u578b\u8def\u5f84")<\/span>
\n# reranker = SentenceTransformersReranker(model)<\/span><\/p>\n

\u672c\u6587\u6ca1\u6709\u4e0b\u8f7d\u6216\u8fd0\u884c\u5177\u4f53\u91cd\u6392\u6a21\u578b，\u4e0d\u58f0\u79f0\u4efb\u4f55\u6a21\u578b\u5728\u4f01\u4e1a\u6570\u636e\u4e0a\u7684\u51c6\u786e\u7387\u3002<\/p>\n

6. \u91cd\u6392\u4e4b\u540e\u8fd8\u8981\u5904\u7406\u4e0a\u4e0b\u6587<\/h3>\n

def<\/span> select_context<\/span>(<\/span>
\n ranked:<\/span> list<\/span>[<\/span>RerankedCandidate]<\/span>,<\/span>
\n max_chars:<\/span> int<\/span>,<\/span>
\n min_score:<\/span> float<\/span>,<\/span>
\n)<\/span> –<\/span>><\/span> list<\/span>[<\/span>Candidate]<\/span>:<\/span>
\n selected:<\/span> list<\/span>[<\/span>Candidate]<\/span> =<\/span> [<\/span>]<\/span>
\n used =<\/span> 0<\/span>
\n seen_documents:<\/span> set<\/span>[<\/span>str<\/span>]<\/span> =<\/span> set<\/span>(<\/span>)<\/span><\/p>\n

for<\/span> item in<\/span> ranked:<\/span>
\n if<\/span> item.<\/span>rerank_score <<\/span> min_score:<\/span>
\n continue<\/span>
\n size =<\/span> len<\/span>(<\/span>item.<\/span>candidate.<\/span>text)<\/span>
\n if<\/span> used +<\/span> size ><\/span> max_chars:<\/span>
\n continue<\/span>
\n # \u793a\u4f8b\u9650\u5236\u6bcf\u4efd\u6587\u6863\u4e00\u4e2a Chunk，\u907f\u514d\u8fd1\u91cd\u590d\u7ed3\u679c\u5360\u6ee1\u4e0a\u4e0b\u6587<\/span>
\n if<\/span> item.<\/span>candidate.<\/span>document_id in<\/span> seen_documents:<\/span>
\n continue<\/span>
\n selected.<\/span>append(<\/span>item.<\/span>candidate)<\/span>
\n seen_documents.<\/span>add(<\/span>item.<\/span>candidate.<\/span>document_id)<\/span>
\n used +=<\/span> size
\n return<\/span> selected<\/p>\n

\u771f\u5b9e\u9884\u7b97\u5e94\u4f7f\u7528\u76ee\u6807\u751f\u6210\u6a21\u578b Tokenizer，\u800c\u4e0d\u662f\u5b57\u7b26\u6570\u3002<\/p>\n

7. \u53ef\u590d\u9a8c\u7684\u5931\u8d25\u6d4b\u8bd5<\/h3>\n

class<\/span> BrokenReranker<\/span>:<\/span>
\n def<\/span> score<\/span>(<\/span>self,<\/span> query:<\/span> str<\/span>,<\/span> texts:<\/span> list<\/span>[<\/span>str<\/span>]<\/span>)<\/span> –<\/span>><\/span> list<\/span>[<\/span>float<\/span>]<\/span>:<\/span>
\n return<\/span> [<\/span>1.0<\/span>]<\/span> # \u6545\u610f\u8fd4\u56de\u9519\u8bef\u6570\u91cf<\/span><\/p>\n

def<\/span> test_broken_reranker_is_rejected<\/span>(<\/span>)<\/span> –<\/span>><\/span> None<\/span>:<\/span>
\n candidates =<\/span> [<\/span>
\n Candidate(<\/span>"c1"<\/span>,<\/span> "d1"<\/span>,<\/span> "t1"<\/span>,<\/span> "A"<\/span>,<\/span> 0.9<\/span>)<\/span>,<\/span>
\n Candidate(<\/span>"c2"<\/span>,<\/span> "d2"<\/span>,<\/span> "t1"<\/span>,<\/span> "B"<\/span>,<\/span> 0.8<\/span>)<\/span>,<\/span>
\n ]<\/span>
\n try<\/span>:<\/span>
\n rerank(<\/span>"\u95ee\u9898"<\/span>,<\/span> candidates,<\/span> BrokenReranker(<\/span>)<\/span>)<\/span>
\n except<\/span> RuntimeError:<\/span>
\n pass<\/span>
\n else<\/span>:<\/span>
\n raise<\/span> AssertionError(<\/span>"\u9519\u8bef\u7684\u5206\u6570\u6570\u91cf\u5fc5\u987b\u88ab\u62d2\u7edd"<\/span>)<\/span><\/p>\n

8. \u5982\u4f55\u8bc1\u660e\u91cd\u6392\u6709\u6548<\/h3>\n

\u56fa\u5b9a\u7b2c\u4e00\u9636\u6bb5\u5019\u9009，\u6bd4\u8f83\u91cd\u6392\u524d\u540e：<\/p>\n