<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0"><channel><title><![CDATA[长上下文不是资料越多越好]]></title><description><![CDATA[<p dir="auto">我看到很多模型都支持很长上下文，是不是可以把所有产品文档、历史聊天、用户资料一次塞进去，省掉 RAG？</p>
]]></description><link>https://localaihub.com/topic/80/长上下文不是资料越多越好</link><generator>RSS for Node</generator><lastBuildDate>Wed, 03 Jun 2026 18:50:44 GMT</lastBuildDate><atom:link href="https://localaihub.com/topic/80.rss" rel="self" type="application/rss+xml"/><pubDate>Mon, 04 May 2026 19:01:00 GMT</pubDate><ttl>60</ttl><item><title><![CDATA[Reply to 长上下文不是资料越多越好 on Tue, 05 May 2026 11:18:00 GMT]]></title><description><![CDATA[<p dir="auto">对，最好把“引用了哪段资料”显示给内部审核。长上下文错了也要能追。</p>
]]></description><link>https://localaihub.com/post/483</link><guid isPermaLink="true">https://localaihub.com/post/483</guid><dc:creator><![CDATA[小满]]></dc:creator><pubDate>Tue, 05 May 2026 11:18:00 GMT</pubDate></item><item><title><![CDATA[Reply to 长上下文不是资料越多越好 on Tue, 05 May 2026 09:49:00 GMT]]></title><description><![CDATA[<p dir="auto">明白，长上下文当能力上限，不当架构方案。我们先保留 RAG，只把单份大文档问答走长上下文。</p>
]]></description><link>https://localaihub.com/post/482</link><guid isPermaLink="true">https://localaihub.com/post/482</guid><dc:creator><![CDATA[半截薯条]]></dc:creator><pubDate>Tue, 05 May 2026 09:49:00 GMT</pubDate></item><item><title><![CDATA[Reply to 长上下文不是资料越多越好 on Tue, 05 May 2026 08:11:00 GMT]]></title><description><![CDATA[<p dir="auto">可以做两层：检索 Top 8，压缩成可引用事实，再让长上下文模型综合。这样不浪费窗口。</p>
]]></description><link>https://localaihub.com/post/481</link><guid isPermaLink="true">https://localaihub.com/post/481</guid><dc:creator><![CDATA[会飞的杯子]]></dc:creator><pubDate>Tue, 05 May 2026 08:11:00 GMT</pubDate></item><item><title><![CDATA[Reply to 长上下文不是资料越多越好 on Tue, 05 May 2026 06:48:00 GMT]]></title><description><![CDATA[<p dir="auto">系统提示和安全边界不能被业务上下文挤掉。预算应该先给指令、工具协议、用户当前问题，再给证据。</p>
]]></description><link>https://localaihub.com/post/480</link><guid isPermaLink="true">https://localaihub.com/post/480</guid><dc:creator><![CDATA[林小北]]></dc:creator><pubDate>Tue, 05 May 2026 06:48:00 GMT</pubDate></item><item><title><![CDATA[Reply to 长上下文不是资料越多越好 on Tue, 05 May 2026 06:32:00 GMT]]></title><description><![CDATA[<p dir="auto">我见过一个误区：把 RAG 检索到的 20 段再加全量聊天记录，最后 token 爆掉，只能截断系统提示，灾难。</p>
]]></description><link>https://localaihub.com/post/479</link><guid isPermaLink="true">https://localaihub.com/post/479</guid><dc:creator><![CDATA[郭同学]]></dc:creator><pubDate>Tue, 05 May 2026 06:32:00 GMT</pubDate></item><item><title><![CDATA[Reply to 长上下文不是资料越多越好 on Tue, 05 May 2026 05:19:00 GMT]]></title><description><![CDATA[<p dir="auto">单文档推理、跨章节对照、要保留原文语境时。多文档知识库、频繁更新、权限过滤，RAG 更稳。</p>
]]></description><link>https://localaihub.com/post/478</link><guid isPermaLink="true">https://localaihub.com/post/478</guid><dc:creator><![CDATA[zeroOne]]></dc:creator><pubDate>Tue, 05 May 2026 05:19:00 GMT</pubDate></item><item><title><![CDATA[Reply to 长上下文不是资料越多越好 on Tue, 05 May 2026 04:15:00 GMT]]></title><description><![CDATA[<p dir="auto">那长上下文什么时候比 RAG 好？</p>
]]></description><link>https://localaihub.com/post/477</link><guid isPermaLink="true">https://localaihub.com/post/477</guid><dc:creator><![CDATA[小周]]></dc:creator><pubDate>Tue, 05 May 2026 04:15:00 GMT</pubDate></item><item><title><![CDATA[Reply to 长上下文不是资料越多越好 on Tue, 05 May 2026 01:56:00 GMT]]></title><description><![CDATA[<p dir="auto">做历史摘要时要保留状态，不是保留所有原话。比如“当前诉求=开票，已确认抬头，未确认税号”。</p>
]]></description><link>https://localaihub.com/post/476</link><guid isPermaLink="true">https://localaihub.com/post/476</guid><dc:creator><![CDATA[index_0]]></dc:creator><pubDate>Tue, 05 May 2026 01:56:00 GMT</pubDate></item><item><title><![CDATA[Reply to 长上下文不是资料越多越好 on Mon, 04 May 2026 23:40:00 GMT]]></title><description><![CDATA[<p dir="auto">我更怕历史聊天。用户前面说“不要发票”，后面又说“要发票”，全塞进去模型可能两边都引用。</p>
]]></description><link>https://localaihub.com/post/475</link><guid isPermaLink="true">https://localaihub.com/post/475</guid><dc:creator><![CDATA[阿宁]]></dc:creator><pubDate>Mon, 04 May 2026 23:40:00 GMT</pubDate></item><item><title><![CDATA[Reply to 长上下文不是资料越多越好 on Mon, 04 May 2026 22:21:00 GMT]]></title><description><![CDATA[<p dir="auto">是优势，但优势是能处理更长输入，不是免疫信息噪声。长材料里有过期政策，模型也会认真引用。</p>
]]></description><link>https://localaihub.com/post/474</link><guid isPermaLink="true">https://localaihub.com/post/474</guid><dc:creator><![CDATA[leaf_1997]]></dc:creator><pubDate>Mon, 04 May 2026 22:21:00 GMT</pubDate></item><item><title><![CDATA[Reply to 长上下文不是资料越多越好 on Mon, 04 May 2026 21:45:00 GMT]]></title><description><![CDATA[<p dir="auto">但 Kimi 这类长上下文不是优势吗？</p>
]]></description><link>https://localaihub.com/post/473</link><guid isPermaLink="true">https://localaihub.com/post/473</guid><dc:creator><![CDATA[小傅]]></dc:creator><pubDate>Mon, 04 May 2026 21:45:00 GMT</pubDate></item><item><title><![CDATA[Reply to 长上下文不是资料越多越好 on Mon, 04 May 2026 21:02:00 GMT]]></title><description><![CDATA[<p dir="auto">长上下文适合“少量长材料”的精读，比如合同、会议纪要、单份手册。知识库长期问答还是要检索和结构化。</p>
]]></description><link>https://localaihub.com/post/472</link><guid isPermaLink="true">https://localaihub.com/post/472</guid><dc:creator><![CDATA[nora]]></dc:creator><pubDate>Mon, 04 May 2026 21:02:00 GMT</pubDate></item><item><title><![CDATA[Reply to 长上下文不是资料越多越好 on Mon, 04 May 2026 20:24:00 GMT]]></title><description><![CDATA[<p dir="auto">我们试过把 180 页 PDF 直接塞给长上下文模型，问第一章还行，问中间表格就开始编。</p>
]]></description><link>https://localaihub.com/post/471</link><guid isPermaLink="true">https://localaihub.com/post/471</guid><dc:creator><![CDATA[小满]]></dc:creator><pubDate>Mon, 04 May 2026 20:24:00 GMT</pubDate></item><item><title><![CDATA[Reply to 长上下文不是资料越多越好 on Mon, 04 May 2026 19:14:00 GMT]]></title><description><![CDATA[<p dir="auto">理论上能塞，不等于应该塞。长上下文贵、慢，而且模型不一定从后半段稳定拿到关键句。</p>
]]></description><link>https://localaihub.com/post/470</link><guid isPermaLink="true">https://localaihub.com/post/470</guid><dc:creator><![CDATA[不想写周报]]></dc:creator><pubDate>Mon, 04 May 2026 19:14:00 GMT</pubDate></item></channel></rss>