[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"pack-detail-rag-pipelines-zh":3,"seo:pack:rag-pipelines:zh":87},{"code":4,"message":5,"data":6},200,"操作成功",{"pack":7},{"slug":8,"icon":9,"tone":10,"status":11,"status_label":12,"title":13,"description":14,"items":15,"install_cmd":86},"rag-pipelines","📚","#3B82F6","stable","稳定","RAG 流水线","Quivr \u002F RAGFlow \u002F GraphRAG + 生产环境最佳实践。少走第一版“取得很烂”的弯路。",[16,28,37,45,52,60,69,79],{"id":17,"uuid":18,"slug":19,"title":20,"description":21,"author_name":22,"view_count":23,"vote_count":24,"lang_type":25,"type":26,"type_label":27},322,"96223597-08c2-4e60-b84e-0c4779641933","quivr-opinionated-rag-framework-any-llm-96223597","Quivr — Opinionated RAG Framework for Any LLM","Quivr is an opinionated RAG framework supporting any LLM, multiple file types, and customizable retrieval. 39.1K+ stars. Apache 2.0.","Script Depot",310,0,"en","script","Script",{"id":29,"uuid":30,"slug":31,"title":32,"description":33,"author_name":22,"view_count":34,"vote_count":24,"lang_type":25,"type":35,"type_label":36},245,"7785d7a8-fc57-42ab-ba6b-4a970404fadc","ragflow-deep-document-understanding-rag-engine-7785d7a8","RAGFlow — Deep Document Understanding RAG Engine","Open-source RAG engine with deep document understanding. Parses complex PDFs, tables, images. Agent-powered Q&A with citations. Multi-model. 77K+ stars.",381,"skill","Skill",{"id":38,"uuid":39,"slug":40,"title":41,"description":42,"author_name":43,"view_count":44,"vote_count":24,"lang_type":25,"type":35,"type_label":36},418,"ac77668d-1767-4b86-ac8c-1c050166d21b","graphrag-knowledge-graph-rag-microsoft-ac77668d","GraphRAG — Knowledge Graph RAG by Microsoft","Build knowledge graphs from documents for smarter RAG. Local and global search over entity relationships. By Microsoft Research. 31K+ stars.","Microsoft AI",374,{"id":46,"uuid":47,"slug":48,"title":49,"description":50,"author_name":22,"view_count":51,"vote_count":24,"lang_type":25,"type":35,"type_label":36},242,"b0f93b10-3339-4ca0-ad20-d6335a3d7785","kotaemon-open-source-rag-document-chat-b0f93b10","Kotaemon — Open-Source RAG Document Chat","Clean, open-source RAG tool for chatting with your documents. Supports PDF, DOCX, web pages. Multi-model, citation, and multi-user. Self-hostable. 25K+ stars.",332,{"id":53,"uuid":54,"slug":55,"title":56,"description":57,"author_name":58,"view_count":59,"vote_count":24,"lang_type":25,"type":35,"type_label":36},1306,"e0e719be-37db-11f1-9bc6-00163e2b0d79","verba-golden-ragtriever-weaviate-e0e719be","Verba — The Golden RAGtriever by Weaviate","Verba is an open-source RAG (Retrieval-Augmented Generation) chatbot from the Weaviate team. Drop in PDFs, web pages, or notes; pick a model (OpenAI, Ollama, Anthropic); and get a polished chat UI with semantic search built in.","AI Open Source",354,{"id":61,"uuid":62,"slug":63,"title":64,"description":65,"author_name":66,"view_count":51,"vote_count":24,"lang_type":25,"type":67,"type_label":68},654,"7ded33e8-464c-4c8f-b3de-6dcf14c0eaf4","rag-best-practices-production-pipeline-guide-2026-7ded33e8","RAG Best Practices — Production Pipeline Guide 2026","Comprehensive guide to building production RAG pipelines. Covers chunking strategies, embedding models, vector databases, retrieval techniques, evaluation, and common pitfalls with code examples.","Prompt Lab","prompt","Prompt",{"id":70,"uuid":71,"slug":72,"title":73,"description":74,"author_name":75,"view_count":76,"vote_count":24,"lang_type":25,"type":77,"type_label":78},635,"f73611a0-142f-4364-97dc-b57eb03473ad","tavily-search-api-built-ai-agents-rag-f73611a0","Tavily — Search API Built for AI Agents & RAG","Search API designed specifically for AI agents and RAG pipelines. Returns clean, LLM-ready results with content extraction, no HTML parsing needed. Official MCP server available. 5,000+ stars.","Tavily",337,"mcp","MCP",{"id":80,"uuid":81,"slug":82,"title":83,"description":84,"author_name":22,"view_count":85,"vote_count":24,"lang_type":25,"type":35,"type_label":36},205,"761bd107-7156-4c62-b268-62a3fb9818dc","haystack-ai-orchestration-search-rag-761bd107","Haystack — AI Orchestration for Search & RAG","Open-source AI orchestration framework by deepset. Build production RAG pipelines, semantic search, and agent workflows with modular components. 25K+ GitHub stars.",263,"tokrepo install pack\u002Frag-pipelines",{"pageType":88,"pageKey":8,"locale":89,"title":90,"metaDescription":91,"h1":13,"tldr":92,"bodyMarkdown":93,"faq":94,"schema":110,"internalLinks":120,"citations":133,"wordCount":146,"generatedAt":147},"pack","zh","RAG 流水线：Quivr \u002F RAGFlow \u002F GraphRAG 生产实战 · TokRepo","跳过第一版「检索很烂」的弯路。Quivr \u002F RAGFlow \u002F GraphRAG + 切片 \u002F 重排 \u002F 评测三件套，能扛上生产的 RAG 模式。TokRepo 一条命令装齐。","八个 RAG 资产 —— 开源引擎（Quivr \u002F RAGFlow \u002F GraphRAG）+ 切片 \u002F 重排 \u002F 评测模式，让你的 RAG 从 demo 到生产。","## 这个 pack 装了什么\n\n多数团队周末就能跑出一个 RAG demo，然后花六个月在「为什么答案微妙地不对」上挣扎。这个包收齐 **八个资产**，帮你越过那堵墙：三个生产级引擎、三个检索\u002F索引模式、两个评测工具。\n\n| # | 资产 | 层 | 为什么收 |\n|---|---|---|---|\n| 1 | Quivr | 全栈 RAG | \"第二大脑\" 参考实现，MIT 协议 |\n| 2 | RAGFlow | 全栈 RAG | 深度文档解析 —— 表格 \u002F 表单比 LangChain 强 |\n| 3 | GraphRAG | 检索 | 微软的知识图谱方案，支持多跳问题 |\n| 4 | 切片模式 | 索引 | 语义 vs 定长 vs 递归 —— 哪种场景赢 |\n| 5 | 混合检索 | 检索 | BM25 + 稠密向量 + 重排 |\n| 6 | Cross-encoder 重排 | 检索 | 你能加上的最大单项精度提升 |\n| 7 | RAG 评测套件 | 可观测 | 黄金集 + LLM-as-judge 夜间回归 |\n| 8 | 引用强制 | 护栏 | 检索分数低于阈值时拒答而非编造 |\n\n## 为什么要装\n\n光靠向量检索能拿到 demo 70% 的质量。剩下 30%（用户真会注意到的部分）来自*非向量*层：怎么切片、怎么重排、怎么判断检索失败该让 LLM 拒答而不是幻觉。\n\n我们做过的每次 RAG 审计都会撞到三个失败模式：\n\n1. **切片破坏了上下文**。单纯按 512 token 切会把表格切两半，标题孤零零落下来。RAGFlow 的版式感知解析器解决这个；纯 LangChain 流水线不行。\n2. **Top-k 检索返回近重复**。余弦相似度最爱把同一段的 5 个改写一起捞上来。加一层 cross-encoder 重排（BGE-reranker \u002F Cohere Rerank）能把多数语料的重复 payload 砍 60%+。\n3. **没有多跳推理**。单次向量查找答不了 \"对比 X 在 2022 \u002F 2023 \u002F 2024 年的变化\"。GraphRAG 在建索引时就构图，让基于遍历的回答成为可能。\n\n## 一条命令装齐\n\n```bash\n# 装整个 pack\ntokrepo install pack\u002Frag-pipelines\n\n# 或者只装你想先跑的引擎\ntokrepo install quivr\ntokrepo install ragflow\ntokrepo install graphrag\n```\n\nTokRepo CLI 在八个 AI 工具间统一了设置文件位置，引擎装好后直接接进现有的 Claude Code \u002F Cursor \u002F Codex CLI 项目。\n\n## 常见坑\n\n- **把 RAG 当成「全部 embed」**。最便宜的精度提升其实是*不*索引低信号页面。先审语料；去重、去导航壳、去过期版本。\n- **跳过重排步骤**。在 top-50 → top-5 之间加 cross-encoder 重排，多数 RAG benchmark 上能把答案正确率提 15-25 个点。\"为了省延迟\"跳过几乎总是错。\n- **没评测套件**。跑不出黄金集回归就不知道上次 prompt 改是变好还是变差。先建评测再扩语料。\n- **存切片不带父级上下文**。永远保留指向源文档和相邻切片的指针；让 LLM 在需要时自己拉更多上下文。\n- **没搞清规模就先选向量库**。1 亿向量以上 Pinecone 才合理；1000 万以下 Qdrant 或 Chroma 单 VM 更快、更便宜、更好调。\n\n## 这个 pack 一个不够时\n\n如果瓶颈在*入库质量*（PDF \u002F 扫描 \u002F 多栏排版），配文档 AI 流水线 pack —— Surya \u002F Docling \u002F MinerU 在切片前清理源文件。如果瓶颈在*评测*，叠 LLM 评测 & 护栏 pack：DeepEval \u002F Ragas \u002F Promptfoo 直接接这里的评测套件。\n\n存储层：这个 pack 引擎无关 —— 看向量数据库横评 pack 按延迟 \u002F 成本 \u002F 准度选 Chroma \u002F Weaviate \u002F Pinecone \u002F Qdrant \u002F txtai。",[95,98,101,104,107],{"q":96,"a":97},"这些 RAG 引擎免费吗？","Quivr \u002F RAGFlow \u002F GraphRAG 都是宽松开源协议（Apache 2.0 \u002F MIT）。自建。你可能加的付费组件只有 embedding API（OpenAI \u002F Cohere \u002F Voyage）和托管向量库 —— 如果你不想自己跑。笔记本规模 demo 零成本；千万文档生产部署主要花在 embedding 账单，不是引擎。",{"q":99,"a":100},"GraphRAG 跟普通 RAG 比怎么样？","普通 RAG 按向量相似度取 top-k 切片塞进 prompt —— 单跳问题（「X 是什么」）很好用。GraphRAG 在建索引时就构实体关系图，能答多跳问题（「X 的角色在这些文档里怎么演变」）。代价是建索引贵 5-10 倍、慢。分析型查询用 GraphRAG，事实查找用普通 RAG。",{"q":102,"a":103},"用 Cursor 或 Codex CLI 也能用吗？","能 —— 这些是服务端引擎，不是编辑器插件。你把 RAGFlow \u002F Quivr 跑成服务，任何能发 HTTP 的 AI 编码工具都能查。TokRepo 装的是 docker-compose 和配置文件丢进项目，同一套配置在 Claude Code \u002F Cursor \u002F Codex CLI \u002F Cline 都能用。检索 API 是一样的。",{"q":105,"a":106},"这个 pack 跟向量数据库横评 pack 啥区别？","向量数据库横评回答「embedding 存哪」—— Chroma \u002F Qdrant \u002F Pinecone \u002F Weaviate 等。RAG 流水线回答「怎么从那个存储里检索 + 重排做出正确答案」。两个 pack 各选一项。多数生产配置底下是 Qdrant 或 pgvector，上面是 RAGFlow 或自定义流水线。",{"q":108,"a":109},"怎么知道我的 RAG 真的在工作？","从真实用户查询里挑 50-200 对问答建黄金集，每晚跑。盯三个数：检索召回（对的切片是否进 top-k）、答案正确率（LLM-as-judge 对比标准答案）、引用忠实度（答案是否引用真检索到的切片）。没这三个数你在瞎飞。Pack 28（LLM 评测 & 护栏）出套件。",{"@context":111,"@type":112,"name":113,"description":114,"numberOfItems":115,"publisher":116},"https:\u002F\u002Fschema.org","CollectionPage","RAG Pipelines","Quivr, RAGFlow, GraphRAG and the production patterns that beat naive vector search.",8,{"@type":117,"name":118,"url":119},"Organization","TokRepo","https:\u002F\u002Ftokrepo.com",[121,125,129],{"url":122,"anchor":123,"reason":124},"\u002Fzh\u002Fpacks\u002Fvector-db-showdown","向量数据库横评","RAG 底下的存储层",{"url":126,"anchor":127,"reason":128},"\u002Fzh\u002Fpacks\u002Fdocument-ai-pipeline","文档 AI 流水线","RAG 上游的入库层",{"url":130,"anchor":131,"reason":132},"\u002Fzh\u002Fpacks\u002Fllm-eval-guardrails","LLM 评测 & 护栏","上线前给检索质量打分",[134,138,142],{"claim":135,"source_name":136,"source_url":137},"GraphRAG combines knowledge graphs with retrieval for multi-hop reasoning","Microsoft GraphRAG","https:\u002F\u002Fgithub.com\u002Fmicrosoft\u002Fgraphrag",{"claim":139,"source_name":140,"source_url":141},"RAGFlow open-source engine for deep document understanding RAG","infiniflow\u002Fragflow","https:\u002F\u002Fgithub.com\u002Finfiniflow\u002Fragflow",{"claim":143,"source_name":144,"source_url":145},"Quivr personal generative AI second brain with RAG","QuivrHQ\u002Fquivr","https:\u002F\u002Fgithub.com\u002FQuivrHQ\u002Fquivr",459,"2026-05-02T15:00:00Z"]