[{"data":1,"prerenderedAt":272},["ShallowReactive",2],{"note:\u002Frag\u002F08_rag_评估调优":3},{"id":4,"title":5,"body":6,"date":250,"description":251,"draft":252,"extension":253,"featured":252,"kind":254,"lastmod":250,"meta":255,"navigation":259,"path":260,"planned":261,"section":262,"seo":263,"source":264,"stem":265,"tags":266,"toc":259,"weight":270,"__hash__":271},"notes\u002Frag\u002F08_RAG_评估调优.md","RAG 效果评估与调优",{"type":7,"value":8,"toc":239},"minimark",[9,14,66,70,106,110,114,118,164,168,183,187,204,207],[10,11,13],"h2",{"id":12},"_1-检索侧指标","1. 检索侧指标",[15,16,17,30],"table",{},[18,19,20],"thead",{},[21,22,23,27],"tr",{},[24,25,26],"th",{},"指标",[24,28,29],{},"含义",[31,32,33,42,50,58],"tbody",{},[21,34,35,39],{},[36,37,38],"td",{},"Recall@k",[36,40,41],{},"正确答案是否出现在前 k 个结果中",[21,43,44,47],{},[36,45,46],{},"Hit Rate",[36,48,49],{},"至少命中一个相关文档的查询占比",[21,51,52,55],{},[36,53,54],{},"MRR",[36,56,57],{},"第一个正确答案排在第几位的平均倒数",[21,59,60,63],{},[36,61,62],{},"NDCG",[36,64,65],{},"考虑排序位置的加权质量",[10,67,69],{"id":68},"_2-生成侧指标","2. 生成侧指标",[15,71,72,80],{},[18,73,74],{},[21,75,76,78],{},[24,77,26],{},[24,79,29],{},[31,81,82,90,98],{},[21,83,84,87],{},[36,85,86],{},"Faithfulness",[36,88,89],{},"答案是否忠于提供的资料，不编造",[21,91,92,95],{},[36,93,94],{},"Answer Relevancy",[36,96,97],{},"答案是否真的回答了问题",[21,99,100,103],{},[36,101,102],{},"Context Precision \u002F Recall",[36,104,105],{},"给模型的上下文是否精准、完整",[10,107,109],{"id":108},"_3-常用评估工具","3. 常用评估工具",[111,112,113],"p",{},"常用评估工具：RAGAS、TruLens、LangSmith。可以把一批“问题 + 标准答案”的测试集跑一遍，用指标驱动优化，而不是只靠人工抽查。",[10,115,117],{"id":116},"_4-常见调优点","4. 常见调优点",[119,120,121,129,135,146,152,158],"ol",{},[122,123,124,128],"li",{},[125,126,127],"strong",{},"切分策略","：父块 800 到 1200 字、子块 150 到 300 字、重叠 20 到 50 字是常用起点；表格、代码、合同条款可能需要专门的切分器。",[122,130,131,134],{},[125,132,133],{},"召回数量","：混合检索先召回 20 到 50 个候选，Reranker 精排后取 Top 3 到 5。",[122,136,137,140,141,145],{},[125,138,139],{},"融合权重","：",[142,143,144],"code",{},"WeightedRanker(0.7, 1.0)"," 表示更依赖语义检索；编号、型号类问题可以调高 Sparse 权重。",[122,147,148,151],{},[125,149,150],{},"元数据过滤","：按部门、文档类型、时间范围过滤，减少无关结果。",[122,153,154,157],{},[125,155,156],{},"查询改写","：多轮对话、指代模糊时，先改写查询再检索。",[122,159,160,163],{},[125,161,162],{},"提示词约束","：要求“资料中没有就说明不知道”，并在答案后附来源编号。",[10,165,167],{"id":166},"_5-进阶方向","5. 进阶方向",[169,170,171,174,177,180],"ul",{},[122,172,173],{},"Agentic RAG：让模型自主决定什么时候检索、检索几次、调用哪些工具",[122,175,176],{},"GraphRAG：用知识图谱组织实体关系，适合“关联性”问题",[122,178,179],{},"多路召回：FAQ 精确匹配 + 向量检索 + SQL\u002F全文检索并行",[122,181,182],{},"长文档：分层摘要、段落定位、表格结构理解",[10,184,186],{"id":185},"_6-常见误区","6. 常见误区",[119,188,189,192,195,198,201],{},[122,190,191],{},"只做 Dense 检索，不做关键词召回：编号、型号、专业名词很容易漏。",[122,193,194],{},"只做检索，不做 Reranker：粗排结果的前几名不一定最相关。",[122,196,197],{},"把整个文档塞进上下文：超出窗口或稀释重点，应切块和精排。",[122,199,200],{},"认为 RAG 不需要评估：切分、召回、排序、提示词每一步都可能引入问题。",[122,202,203],{},"更新知识只替换文件：必须重新切分、重新向量化，并更新索引。",[10,205,206],{"id":206},"参考文档",[169,208,209,218,225,232],{},[122,210,211],{},[212,213,217],"a",{"href":214,"rel":215},"https:\u002F\u002Fbge-model.com\u002Fbge\u002Fbge_m3.html",[216],"nofollow","BGE-M3 官方说明",[122,219,220],{},[212,221,224],{"href":222,"rel":223},"https:\u002F\u002Fmilvus.io\u002Fdocs\u002Fzh\u002Fhybrid_search_with_milvus.md",[216],"Milvus Hybrid Search with BGE-M3",[122,226,227],{},[212,228,231],{"href":229,"rel":230},"https:\u002F\u002Fhelp.aliyun.com\u002Fzh\u002Fmodel-studio\u002Fqwen-api-via-dashscope",[216],"DashScope 千问 API",[122,233,234],{},[212,235,238],{"href":236,"rel":237},"https:\u002F\u002Fhuggingface.co\u002FBAAI\u002Fbge-reranker-v2-m3",[216],"BAAI\u002Fbge-reranker-v2-m3",{"title":240,"searchDepth":241,"depth":241,"links":242},"",2,[243,244,245,246,247,248,249],{"id":12,"depth":241,"text":13},{"id":68,"depth":241,"text":69},{"id":108,"depth":241,"text":109},{"id":116,"depth":241,"text":117},{"id":166,"depth":241,"text":167},{"id":185,"depth":241,"text":186},{"id":206,"depth":241,"text":206},"2026-05-26","从检索侧到生成侧的评估指标与调优路径，用数据驱动 RAG 系统的持续改进。",false,"md","article",{"migration":256},{"generator":257,"sourceSha256":258},"nuxt-site\u002Fscripts\u002Fmigrate-content.mjs","4f51574bf3364a4ab3cce76adf7f2ad7ccb2fff57cc4974dfa0b3447e0d2ffc7",true,"\u002Frag\u002F08_rag_评估调优",null,"rag",{"title":5,"description":251},"content\u002Frag\u002F08_RAG_评估调优.md","rag\u002F08_RAG_评估调优",[267,268,269],"RAG","评估","LangSmith",25,"ZA6cd4eKPXc5aP8mZIIIU3AX9HIe1VUhBvES-erSWLk",1791279145580]