<?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[历史消息截断后，模型忘了用户已经给过的信息]]></title><description><![CDATA[<p dir="auto">我们多轮问答有个问题，用户第 2 轮给了订单号，第 9 轮模型又问一次订单号。查日志发现前面消息被截断了。</p>
]]></description><link>https://localaihub.com/topic/82/历史消息截断后-模型忘了用户已经给过的信息</link><generator>RSS for Node</generator><lastBuildDate>Wed, 03 Jun 2026 18:50:44 GMT</lastBuildDate><atom:link href="https://localaihub.com/topic/82.rss" rel="self" type="application/rss+xml"/><pubDate>Mon, 04 May 2026 22:33:00 GMT</pubDate><ttl>60</ttl><item><title><![CDATA[Reply to 历史消息截断后，模型忘了用户已经给过的信息 on Tue, 05 May 2026 18:19:00 GMT]]></title><description><![CDATA[<p dir="auto">记得加回放测试。截断 bug 靠人工聊天很难稳定复现。</p>
]]></description><link>https://localaihub.com/post/513</link><guid isPermaLink="true">https://localaihub.com/post/513</guid><dc:creator><![CDATA[何必呢]]></dc:creator><pubDate>Tue, 05 May 2026 18:19:00 GMT</pubDate></item><item><title><![CDATA[Reply to 历史消息截断后，模型忘了用户已经给过的信息 on Tue, 05 May 2026 18:08:00 GMT]]></title><description><![CDATA[<p dir="auto">我回去改：聊天历史截断前先抽关键事实，摘要结构化，订单号进入会话状态。</p>
]]></description><link>https://localaihub.com/post/512</link><guid isPermaLink="true">https://localaihub.com/post/512</guid><dc:creator><![CDATA[小谢]]></dc:creator><pubDate>Tue, 05 May 2026 18:08:00 GMT</pubDate></item><item><title><![CDATA[Reply to 历史消息截断后，模型忘了用户已经给过的信息 on Tue, 05 May 2026 15:29:00 GMT]]></title><description><![CDATA[<p dir="auto">前端也能帮一点，用户已填写信息放在表单态，不要都混在聊天里。</p>
]]></description><link>https://localaihub.com/post/511</link><guid isPermaLink="true">https://localaihub.com/post/511</guid><dc:creator><![CDATA[葡萄冰]]></dc:creator><pubDate>Tue, 05 May 2026 15:29:00 GMT</pubDate></item><item><title><![CDATA[Reply to 历史消息截断后，模型忘了用户已经给过的信息 on Tue, 05 May 2026 14:13:00 GMT]]></title><description><![CDATA[<p dir="auto">我们后来每轮都跑“状态抽取”，模型输出 JSON，后端校验字段，失败就不用。稳定很多。</p>
]]></description><link>https://localaihub.com/post/510</link><guid isPermaLink="true">https://localaihub.com/post/510</guid><dc:creator><![CDATA[别问问就是缓存]]></dc:creator><pubDate>Tue, 05 May 2026 14:13:00 GMT</pubDate></item><item><title><![CDATA[Reply to 历史消息截断后，模型忘了用户已经给过的信息 on Tue, 05 May 2026 12:04:00 GMT]]></title><description><![CDATA[<p dir="auto">可以让模型参与摘要，但最终要有程序规则兜底关键槽位。完全让模型决定，出错时很难复盘。</p>
]]></description><link>https://localaihub.com/post/509</link><guid isPermaLink="true">https://localaihub.com/post/509</guid><dc:creator><![CDATA[林小北]]></dc:creator><pubDate>Tue, 05 May 2026 12:04:00 GMT</pubDate></item><item><title><![CDATA[Reply to 历史消息截断后，模型忘了用户已经给过的信息 on Tue, 05 May 2026 10:09:00 GMT]]></title><description><![CDATA[<p dir="auto">历史截断前是不是要让模型自己决定保留什么？</p>
]]></description><link>https://localaihub.com/post/508</link><guid isPermaLink="true">https://localaihub.com/post/508</guid><dc:creator><![CDATA[小傅]]></dc:creator><pubDate>Tue, 05 May 2026 10:09:00 GMT</pubDate></item><item><title><![CDATA[Reply to 历史消息截断后，模型忘了用户已经给过的信息 on Tue, 05 May 2026 08:12:00 GMT]]></title><description><![CDATA[<p dir="auto">工具结果也分长期和短期。订单状态这种会变，别永久写死；用户已授权查询这种可以保留。</p>
]]></description><link>https://localaihub.com/post/507</link><guid isPermaLink="true">https://localaihub.com/post/507</guid><dc:creator><![CDATA[陈一]]></dc:creator><pubDate>Tue, 05 May 2026 08:12:00 GMT</pubDate></item><item><title><![CDATA[Reply to 历史消息截断后，模型忘了用户已经给过的信息 on Tue, 05 May 2026 06:36:00 GMT]]></title><description><![CDATA[<p dir="auto">还有一种是工具结果丢了。模型查过订单，但下一轮工具返回不在上下文里，它就又查。</p>
]]></description><link>https://localaihub.com/post/506</link><guid isPermaLink="true">https://localaihub.com/post/506</guid><dc:creator><![CDATA[周五下午]]></dc:creator><pubDate>Tue, 05 May 2026 06:36:00 GMT</pubDate></item><item><title><![CDATA[Reply to 历史消息截断后，模型忘了用户已经给过的信息 on Tue, 05 May 2026 05:15:00 GMT]]></title><description><![CDATA[<p dir="auto">那不行。写成结构化更稳：<code>order_id=xxx</code>、<code>confirmed=true</code>、<code>source_turn=2</code>。模型看得懂，系统也能校验。</p>
]]></description><link>https://localaihub.com/post/505</link><guid isPermaLink="true">https://localaihub.com/post/505</guid><dc:creator><![CDATA[zeroOne]]></dc:creator><pubDate>Tue, 05 May 2026 05:15:00 GMT</pubDate></item><item><title><![CDATA[Reply to 历史消息截断后，模型忘了用户已经给过的信息 on Tue, 05 May 2026 04:54:00 GMT]]></title><description><![CDATA[<p dir="auto">我们有 summary，但模型还是忘，因为 summary 太像自然语言：“用户提到了一些订单信息”。</p>
]]></description><link>https://localaihub.com/post/504</link><guid isPermaLink="true">https://localaihub.com/post/504</guid><dc:creator><![CDATA[小谢]]></dc:creator><pubDate>Tue, 05 May 2026 04:54:00 GMT</pubDate></item><item><title><![CDATA[Reply to 历史消息截断后，模型忘了用户已经给过的信息 on Tue, 05 May 2026 02:34:00 GMT]]></title><description><![CDATA[<p dir="auto">摘要要有版本。用户后来改口，旧摘要必须覆盖，不然状态里同时有两个订单号。</p>
]]></description><link>https://localaihub.com/post/503</link><guid isPermaLink="true">https://localaihub.com/post/503</guid><dc:creator><![CDATA[青菜]]></dc:creator><pubDate>Tue, 05 May 2026 02:34:00 GMT</pubDate></item><item><title><![CDATA[Reply to 历史消息截断后，模型忘了用户已经给过的信息 on Tue, 05 May 2026 01:49:00 GMT]]></title><description><![CDATA[<p dir="auto">截断策略要认业务字段。订单号、姓名、城市、产品版本、已尝试步骤，都是“关键事实”，不能按时间删。</p>
]]></description><link>https://localaihub.com/post/502</link><guid isPermaLink="true">https://localaihub.com/post/502</guid><dc:creator><![CDATA[melo]]></dc:creator><pubDate>Tue, 05 May 2026 01:49:00 GMT</pubDate></item><item><title><![CDATA[Reply to 历史消息截断后，模型忘了用户已经给过的信息 on Tue, 05 May 2026 00:30:00 GMT]]></title><description><![CDATA[<p dir="auto">我之前偷懒全靠历史，窗口大时没问题，一上线用户连续追问就露馅。</p>
]]></description><link>https://localaihub.com/post/501</link><guid isPermaLink="true">https://localaihub.com/post/501</guid><dc:creator><![CDATA[何必呢]]></dc:creator><pubDate>Tue, 05 May 2026 00:30:00 GMT</pubDate></item><item><title><![CDATA[Reply to 历史消息截断后，模型忘了用户已经给过的信息 on Mon, 04 May 2026 23:41:00 GMT]]></title><description><![CDATA[<p dir="auto">这个是状态管理问题，不是单纯模型问题。订单号这种槽位应该进会话状态，不能只躺在聊天历史里。</p>
]]></description><link>https://localaihub.com/post/500</link><guid isPermaLink="true">https://localaihub.com/post/500</guid><dc:creator><![CDATA[小林]]></dc:creator><pubDate>Mon, 04 May 2026 23:41:00 GMT</pubDate></item></channel></rss>