[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"$f3zofbd8lecxy":3},{"code":4,"message":5,"data":6,"timestamp":64},10000,"success",{"full_name":7,"position":8,"description":9,"language":10,"stars":11,"forks":12,"stars_gained":13,"age_days":14,"hn_threads":15,"one_liner":42,"summary_md":43,"topic":44,"github_topics":47,"license":60,"homepage":61,"updated_at":62,"history":63},"VectifyAI\u002FPageIndex",8,"📑 PageIndex: Document Index for Vectorless, Reasoning-based RAG","Python",39614,3426,3867,571,[16,22,28,33,37],{"id":17,"title":18,"points":19,"comments":20,"posted_at":21},45036944,"Show HN: PageIndex – Vectorless RAG",192,128,"2025-08-27",{"id":23,"title":24,"points":25,"comments":26,"posted_at":27},43656476,"Show HN: RAG, No Vectors",11,1,"2025-04-11",{"id":29,"title":30,"points":25,"comments":31,"posted_at":32},45015281,"Show HN: Human-like RAG — no vectors",0,"2025-08-25",{"id":34,"title":35,"points":8,"comments":31,"posted_at":36},49918836,"Show HN: Vectorless, Reasoning-Based RAG","2026-10-01",{"id":38,"title":39,"points":40,"comments":31,"posted_at":41},45773923,"Show HN: Vision-Based, Vectorless RAG for Long Douments",6,"2025-10-31","面向长专业 PDF 的推理式 RAG，用 LLM 沿文档树检索，取代向量相似度。","## 解决什么问题\n向量 RAG 把文档切块、做 embedding、按语义相似度召回。在 10-K 财报、合同、监管文件、技术手册这类长而结构复杂的文档上，相似不等于相关：召回的是措辞相近却答非所问的段落，真正相关但换个说法的内容反而漏掉。PageIndex 换了个思路：先给文档建一棵层级树索引，再让 LLM 像人类专家那样顺着目录逐层推理，翻到该读的那一节。\n\n## 有什么不同\n和 Chroma、pgvector 这类向量库相比，差别不在调参，而在检索机制本身：不切块、不做 embedding，返回结果可以追溯到具体的文档节点，天然便于给出引用和审计。和把整份 PDF 直接塞进长上下文模型相比，PageIndex 只读推理路径经过的节点，查询成本不随文档长度线性上涨——README 给出的数据是原生 PDF 输入在 52 页时贵 2.1 倍、420 页时贵 16.6 倍，805 页则根本放不进上下文窗口。代价是延迟和模型依赖：检索本身是一步步的 LLM 推理，聊天模型弱会直接拖累效果，这点比向量方案更敏感。\n\n## 适合谁\n适合在密集、结构化的长文档上做问答、尽职调查或研究工具的团队：金融科技和法务科技的工程师、分析师，语料是几十到几千份长 PDF，而不是几百万条短网页。反过来，对高 QPS、低延迟的短文本语义搜索，对多语言或模糊召回有明显依赖的场景，或者不愿意为每次查询支付 LLM token 的团队，它并不合适。\n\n## 上手成本\nPython 包，MIT 协议，需要自备 LLM API key（OpenAI，或装可选依赖后用 Claude 订阅额度）。建索引是一次性成本，README 称约 $0.001\u002F页、耗时 13 秒到 4.5 分钟；但每次查询的开销取决于聊天模型在树上走多远，最好用自己的语料实测。项目很新（2025 年 4 月创建），README 明显偏向云端和 App，本地 SDK 细节、Agent 集成、多文档检索都在站外文档里，动手前先读文档。\n\n## 点评\nPageIndex 是一套真正不同的检索设计，不是又一层向量库包装；README 报告的 FinanceBench 成绩和 39.6k 星标说明关注度真实存在，HotRepo 记录的这个月它涨了约 3,867 星，月榜排到第 20。需要留意的是成熟度和成本结构：项目刚过一年，Hacker News 上的讨论不多（一条 192 分的 Show HN），而逐次查询的 LLM 推理和一次相似度查找完全不是一个计价方式。文档长、答案需要可追溯的话，值得认真做原型；想直接替换成熟的向量方案，还早了。",{"slug":45,"name":46},"ai-engineering","AI 工程",[48,49,50,51,52,53,54,55,56,57,58,59],"agentic-ai","agents","ai","ai-agents","context-engineering","information-retrieval","llm","rag","reasoning","retrieval","retrieval-augmented-generation","vector-database","MIT","https:\u002F\u002Fpageindex.ai","2026-10-24T16:35:23Z",[],1792868439]