内容来源：https://blog.n8n.io/spotlight-you-com/

内容总结：

新闻摘要工具的痛点被一套新工作流破解：先过滤，再花钱

核心观点：新闻摘要的难点从来不是找新闻，而是判断什么值得读。

搜索和摘要的成本都很低，真正昂贵的是对每一条内容做出取舍判断——而这在过去意味着必须调用大模型。如果每天早上用通用大模型跑六十篇文章，花费不菲且延迟明显，最终却发现其中五十四篇是同一份新闻稿的转载。因此大多数摘要工具干脆放弃过滤：要么全部推送，要么用大模型摘要全部内容，本质上没有区别。结果是每天早上8点准时推送，无论是否真有大事发生，三周之内就没人再点开了。每天都“狼来了”的摘要，比没有摘要更糟。

新方案：先过滤，后写作

You.com 高级软件工程师 Tyler Eastman 介绍了一套基于 n8n 的工作流。其核心思路是：先让便宜的模型过滤，再让昂贵的模型只处理通过过滤的内容。

工作流使用 TypeSafe AI 的首个 System One 模型 Jev 对每篇文章打分，每篇成本约万分之四美分。在这个价位上，可以对每篇文章分别提出三个问题，而不是用一个模糊问题笼统判断。通过过滤的内容再交给 You.com Research，生成带引用的简短摘要。

整个早上的运行成本在1到4美分之间。更重要的是：如果某个话题当天没有实质性新闻，工作流对该话题什么都不推送——正是这一点让用户到第三个月仍愿意打开它。

工作流五步拆解

第一步：配置话题与关注点。 在 Config 节点中设置话题（模板默认三个：模型发布、智能体工具、AI监管）、团队描述、关注事项以及重要性阈值。团队描述看似简单，实则影响巨大——“我们为开发者提供文档抽取API”和“我们为银行运行机器学习基础设施”，对同一篇推理定价变动文章的评分会截然不同。

第二步：搜索当日新闻并去重。 You.com 节点按话题搜索过去一天的新闻，此前已出现过的文章会被跳过，避免同一新闻连续三天出现在摘要中。搜索范围宜宽，因为下游过滤足够便宜，返回过多结果不是问题。

第三步：向 Jev 提出三个问题。 这是整个工作流的核心节点。Jev 不是聊天模型，它接收一个状态和一组结构化问题，返回带校准置信度的结构化数值，响应时间远低于一秒，且不生成任何文本。三个问题分别是：是否报道了真实进展？是否与话题相关？对团队有多重要？三个问题独立并行评估，而非合并为一个笼统判断——这样当摘要开始推送错误内容时，你能知道是哪个问题出了偏差。六十篇文章的总过滤成本约为四分之一美分。

第四步：按事件而非文章分组。 一次模型发布到中午可能被六家媒体报道，如果列出六条就是注水。通过过滤的文章按话题分组，多来源报道同一事件合并为一条，来源多反而是重要性的信号。

第五步：每个话题写一份带引用的摘要。 此时才运行昂贵的操作。You.com Research 为每个话题生成带链接引用的简短摘要，推送到 Slack。引用是让摘要在工作场景中可用的关键——有链接的断言可以转发给产品负责人，没有链接的只是传闻。没有实质内容的话题则不推送，连“今日无新闻”都不发。

成本结构与核心优势

一个早晨的运行成本约为1到4美分，包括搜索、每篇文章一次 Jev 调用、以及每个有新闻的话题一份研究摘要。安静的日子成本更低，因为没有内容的话题不会触发研究步骤。

成本低的根本原因在于：昂贵步骤只会看到已经通过每篇万分之四美分过滤的内容。如果颠倒顺序，让重型模型处理全部内容并要求它保持选择性，就等于为六十篇文章的推理付费，只换来四篇文章的价值。

相比简单摘要工具，这套方案带来三点不同：

• 安静日保持沉默，这是让人第三个月仍愿意打开的关键；
• 可调节的重要性阈值，而非每次输出偏差都要重新争论的提示词；
• 每条断言都有引用，摘要可直接转发。

后续扩展方向

• 在大新闻日启用更重的研究：Research 操作支持后台模式，返回任务ID后轮询获取结果，适合提高高内容量话题的研究强度。
• 增加第四个问题：Jev 并行评估问题，增加一个几乎不影响响应时间。例如增加紧迫度评分，将重要话题立即单独推送，而不必等到早上8点。
• 限定来源范围：搜索操作支持 include_domains、exclude_domains 和 boost_domains，每项最多500个值。
• 抓取全文：将通过的URL送入 Contents 操作，按 max_age 阈值抓取并归档页面。
• 更换推送目的地：Slack 只是模板默认选项，可替换为邮件、Notion 页面或积累每日要闻的可搜索表格。

结论

过滤一直是新闻摘要中昂贵的那一半，这正是大多数工具并未真正过滤的原因。改变的是“做一次判断”的价格。在万分之四美分的价位上，可以对信息洪流中的每篇文章提出三个细致问题，而整个过滤步骤的花费仍低于前沿模型一次摘要的成本。广泛搜索，严格过滤，只写剩下的内容。

中文翻译：

本篇“验证节点聚焦”客座文章由 You.com 高级软件工程师 Tyler Eastman 撰写。

新闻摘要最难的部分从来都不是找新闻。

搜索很便宜，摘要也很便宜。真正昂贵的是判断哪些内容值得一读，因为这需要对每一条内容逐一做出判断，而判断从来都意味着调用模型。每天早上让一个通用大语言模型跑六十篇文章，你要付出真金白银和实实在在的延迟，结果只是被告知其中五十四篇是同一份新闻稿的转载。

所以大多数摘要干脆放弃了筛选。它们把所有内容都发出去，或者把所有内容的 LLM 摘要都发出去——后者不过是前者多绕了几步而已。无论哪种方式，这东西每天早上 8 点准时出现在 Slack 里，不管有没有事情发生，不出三周人们就不再点开了。一份天天喊“狼来了”的摘要，比没有摘要还糟糕。

这个工作流先筛选。TypeSafe AI 的首个 System One 模型 Jev 会对搜索返回的每篇文章打分，每篇大约零点零零零四美分。在这个价格下，你可以针对每条结果分别问三个独立的问题，而不是对所有结果笼统地问一个模糊的问题。通过筛选的内容会交给 You.com Research，由它针对每个主题撰写一篇简短的带引用摘要。

一整个早晨的花费大约在一到四美分之间。而在安静的日子里，如果某个主题没有发生实质性的事情，工作流对该主题就什么都不发。最后这个行为，才是让人们到第三个月还在读它的原因。

摘要的工作原理

图 1：广搜索、严筛选，然后只对通过筛选的内容花钱。

一个定时触发器在每个工作日早上 8 点（按你实例的时区）触发。Config 节点保存你的主题、团队描述以及团队关注的内容，这样只需编辑一个节点，而不是四个。

对于每个主题，You.com 节点搜索过去一天的新闻。之前运行中已经见过的内容会被跳过，所以周二的摘要不会重复报道周一的发布。

每篇存活的文章都会送到 Jev，它在一次调用中回答三个问题：这篇文章是否报道了一个真实的新进展，它是否确实与主题相关，以及它对你们的团队有多重要。通过重要性阈值的文章继续往下走。其余的被丢弃，再也不会让你花一分钱。

剩下的内容按主题分组，这样六家媒体报道同一个模型发布就合并成一条，而不是六条。然后 You.com Research 为每个主题撰写一篇带链接引用的简短摘要，再由 Slack 发布出去。

顺序很重要。最昂贵的操作放在最后运行，处理的是一个已经被筛选器削减过的集合，而筛选器的成本只是其零头。

开始之前

• 一个 n8n 实例。n8n Cloud 和自托管都可以。
• 安装 You.com 节点。@youdotcom-oss/n8n-nodes-youdotcom 是一个经过验证的社区节点。添加一个节点，在搜索面板中输入 you.com，然后选择它。你也可以通过 设置 > 社区节点 > 安装，使用包名进行安装。
• 一个 You.com API 密钥。在此注册并在仪表盘中创建一个密钥。
• 一个 TypeSafe API 密钥。Jev 在撰写本文时处于早期访问阶段；申请访问后你会被从等待名单中移出。
• 一个 Slack 凭据以及你的机器人已加入的频道。

自己动手试试

将免费模板导入你的 n8n 实例即可跟着操作。画布加载时所有节点都已连接并配置好。

获取工作流

第一步：设置你的主题和关注内容

打开 Config。这里有四项内容。

你的主题。模板自带的是模型发布、智能体工具和 AI 监管。三个是一个不错的起始数量，因为这样早晨的帖子足够短，人们才真的会读。

你的团队。用一两句话描述你们做什么。这是 Jev 用来评估重要性的依据，它的作用比看起来要大得多。“我们为文档提取提供面向开发者的 API”和“我们为一家银行运营机器学习基础设施”，面对同一篇关于推理定价变化的文章，评分会截然不同。

你关注的内容。那些会改变你决策的具体事项：竞争对手发布了新产品、价格变动、有合规截止日期的监管规定。

materialityThreshold（重要性阈值）。一篇文章需要达到多高的分数才能进入摘要。从模板的默认值开始，经过一周真实的早晨后再调整。

最后这一项才是你真正会去调的旋钮。太高了，安静的日子会安静得让人觉得出了问题；太低了，你又回到了没人读的摘要。它只是一个节点里的一个数字，这正是把判断标准放在 Config 里而不是埋在提示词里的意义所在。

第二步：搜索当天新闻并跳过已见过的内容

You.com 节点搜索每个主题过去一天的新闻。

之前运行中出现过的文章会在其他任何事情发生之前被跳过。没有这一步，一个在新闻周期中持续三天的故事会连续三天出现在摘要里，人们就会学会整条消息直接略过。去重在每日摘要中不是锦上添花，它是工具和累赘之间的区别。

图 2：单个主题返回的结果比你想读的要多。没关系，因为读它的成本几乎为零。

在这里要广搜索。下游的筛选足够便宜，返回太多结果不值得花精力去优化掉。

第三步：对每篇文章问 Jev 三个问题

这是整个工作流围绕构建的节点。

Jev 不是聊天模型。你给它一个状态和一组有类型的问题，它在远不到一秒内返回带有校准置信度的有类型值。它不生成文本，这意味着它也不可能生成错误形状的文本。在这里，状态就是文章，三个问题是：

这篇文章是否报道了一个真实的新进展？ 把产品发布和关于该发布的评论文章区分开来，把融资消息和提到该融资的消息汇总区分开来。仅这一个问题就能剔除掉一个粗糙摘要本来会发给你的大部分内容。

它是否与这个主题相关？ 新闻搜索会漂移。一个关于 AI 监管的查询会返回一篇关于数据中心分区争议的报道，它严格来说不算错，只是不是你想要的。

这对这个团队有多重要？ 根据 Config 中的团队描述和优先级进行评分。这就是产生 materialityThreshold 的那个问题。

图 3：三个问题，一次调用，一篇文章。每个问题都针对同一状态独立评估。

用三个独立的问题而不是一个合并的问题，这是刻意的，也遵循了 TypeSafe 自己关于组合原子问题的指导。每个问题只问一件范围明确的事，每个问题都被独立且并行地评估，而组合它们的逻辑存在于你的工作流中，你可以看到。如果改成要一个单一的“这篇该不该进摘要”的裁决，你就把三种不同的判断埋进了一个你无法单独调整的数字里。当摘要开始推送错误的内容时，你想知道是三个问题中的哪一个出了偏差。

每篇文章的成本约为 $0.00004。三个主题六十篇文章大约是四分之一美分。这个数字正是这个工作流能够真正做到筛选的全部原因，而不是把所有内容都摘要一遍然后管那叫筛选。

第四步：按故事分组，而不是按文章分组

一个模型发布到午饭前就会在六个地方产生报道。一份列出六条同一事件内容的摘要是注水，读者会注意到。

有实质内容的文章在撰写摘要之前按主题分组，这样多家媒体报道同一进展就变成一条内容，背后有多个来源。一个故事有更多来源是它重要的信号，而不是占据六行的理由。

第五步：为每个主题撰写一篇带引用的摘要

直到这时才运行任何昂贵的东西。You.com Research 接收某个主题存活的文章，撰写一篇带链接引用的简短摘要，结果发送到 Slack。

引用是让这在工作场景中可用的关键。一篇声称竞争对手改了价格的摘要是传言；一篇附链接声称此事的摘要是你可以转发给产品负责人的东西。

没有实质性内容的主题不发帖子。不是发一条说“今天没什么事”的帖子——那本身也是一种噪音。什么都不发。

图 4：四个主题有新闻。第五个没有，所以它不在这里。

成本是多少

一个早晨大约花费 $0.01 到 $0.04。这就是全部：搜索、每篇文章一次 Jev 调用，以及每个有新闻的主题一次研究摘要。安静的日子花费更少，因为研究步骤不会对没有内容的主题运行。

它便宜的结构性原因是，昂贵的那一步只会看到已经通过筛选器（每项成本四千分之一美分）的内容。把顺序反过来，让重型模型跑所有内容并要求它有选择性地输出，你就要为六十篇文章的推理付费，却只得到四篇文章的价值。

三个更简单的摘要做不到的东西：

• 安静的日子保持沉默，这正是让人们到第三个月还在点开它的原因。
• 一个你可以调整的重要性阈值，而不是每次输出漂移时你都要重新争论的提示词。
• 每条论断都有引用，所以摘要可以转发。

接下来可以做什么

在大日子里运行更重的研究。Research 操作有一个后台模式，返回一个任务 id 而不是保持连接打开，搭配一个 Get Research Task 操作，你可以轮询直到状态返回 completed。如果你要提高内容很多的主题的努力级别，这值得加上；frontier，即最高努力级别，只以这种方式运行。

加第四个问题。Jev 对同一状态并行评估问题，所以再加一个几乎不会影响响应时间。一个紧急度评分，把某个主题提升为独立的即时消息而不是等到早上 8 点，是一个自然的下一步。

限定来源范围。You.com 搜索操作接受 include_domains、exclude_domains 和 boost_domains，每个最多 500 个值。提升你信任的媒体而不排除其他所有，或者排除一个总是在你的主题下排名靠前的内容农场。

拉取全文。把存活的 URL 送入 Contents 操作，配合 max_age 阈值来抓取和归档页面本身。Contents 在一个节点中接受多个 URL。

更换目的地。模板里用 Slack 是因为早上 8 点人们已经在那里了。换成邮件、Notion 页面，或者一个累积每日重要内容可搜索记录的表格。

结论

筛选一直是新闻摘要中昂贵的那一半，这就是为什么大多数摘要并没有真正在筛选。改变的是判断的成本。在零点零零零四美分这个价位，你可以对信息洪流中的每篇文章认真问三个问题，而整个筛选步骤的花费仍然比一个前沿模型的一次摘要还少。广搜索、严筛选，然后只对剩下的内容动笔。

立即试用

导入工作流，把所有节点连同连接和配置一起放入你的 n8n 实例。

获取模板

关于 You.com 节点支持的完整操作集，包括 Answer、Finance Research 和 Contents，请参阅 You.com API 文档。关于 Jev 的原语以及如何在架构上使用置信度，请参阅 TypeSafe 文档。

英文来源：

This Verified Node Spotlight is a guest post written by Tyler Eastman, Senior Software Engineer at You.com.
The hard part of a news digest was never finding the news.
Searching is cheap and summarizing is cheap. Deciding what's worth reading is the expensive part, because it takes a judgment call on every single item, and a judgment call has always meant a model call. Run a general-purpose LLM across sixty articles every morning and you're paying real money and real latency to be told that fifty-four of them are reprints of the same press release.
So most digests give up on filtering. They send everything, or they send an LLM summary of everything, which is the same thing with extra steps. Either way the thing arrives in Slack every morning at 8am whether or not anything happened, and within three weeks people stop opening it. A digest that cries wolf daily is worse than no digest.
This workflow filters first. Jev, the first System One model from TypeSafe AI, scores every article that comes back from the search for about four hundredths of a cent each. At that price you can afford to ask three separate questions about every result instead of one vague question about all of them. What survives the filter goes to You.com Research, which writes a short cited brief per topic.
A whole morning costs somewhere between one and four cents. And on a quiet day, when nothing material happened in a topic, the workflow posts nothing at all for that topic. That last behavior is the one that keeps people reading it in month three.
How the brief works
Figure 1: Search wide, filter hard, then spend money only on what survived.
A Schedule Trigger fires every weekday at 8am in your instance's timezone. Config holds your topics, a description of your team, and what that team cares about, so there's one node to edit rather than four.
For each topic, the You.com node searches news from the past day. Anything seen in an earlier run is skipped, so Tuesday's brief doesn't re-report Monday's launch.
Every surviving article goes to Jev, which answers three questions about it in a single call: whether it reports a real development, whether it's actually about the topic, and how much it matters to your team. Articles that clear the materiality threshold continue. The rest are dropped and never cost you anything again.
What's left is grouped by topic, so six outlets covering one model release collapse into one item rather than six. Then You.com Research writes a short brief per topic with linked citations, and Slack posts it.
The order matters. The expensive operation runs last, on a set that's already been cut down by a filter costing a fraction of a cent.
Before you start

• An n8n instance. n8n Cloud and self-hosted both work.
• The You.com node installed. @youdotcom-oss/n8n-nodes-youdotcom is a verified community node. Add a node, type you.com in the search panel, and select it. You can also install it from Settings > Community Nodes > Install using the package name.
• A You.com API key. Sign up here and create a key from the dashboard.
• A TypeSafe API key. Jev is in early access at the time of writing; request access and you'll be brought off the waitlist.
• A Slack credential and a channel your bot has joined.
  Try it yourself
  Import the free template to your n8n instance to follow along. The canvas loads with every node wired and configured.
  Get the workflowStep 1: Set your topics and what you care about
  Open Config. Four things live here.
  Your topics. The template ships with model releases, agent tooling, and AI regulation. Three is a good starting number because it keeps the morning post short enough to actually read.
  Your team. A sentence or two describing what you build. This is what Jev scores materiality against, and it's doing more work than it looks like. "We ship a developer-facing API for document extraction" and "we run ML infrastructure for a bank" will diverge sharply on the same article about an inference pricing change.
  What you care about. The specific things that would change a decision for you: a competitor shipping, a price change, a regulation with a compliance date.
  materialityThreshold
  . How high an article has to score to make the brief. Start where the template has it and adjust after a week of real mornings.
  That last one is the dial you'll actually touch. Too high and quiet days go silent in a way that feels broken; too low and you're back to a digest nobody reads. It's one number in one node, which is the point of keeping the judgment criteria in config instead of buried in a prompt.
  Step 2: Search the day's news and skip what you've already seen
  The You.com node searches each topic over the past day.
  Articles that appeared in an earlier run are skipped before anything else happens. Without this, a story that stays in the news cycle for three days shows up in three consecutive briefs, and people learn to skim past the whole message. Deduplication isn't a nice-to-have in a daily digest, it's the difference between a tool and a nuisance.
  Figure 2: A single topic returns more than you'd want to read. That's fine, because reading it costs almost nothing.
  Search wide here. The filter downstream is cheap enough that returning too many results is not a problem worth optimizing away.
  Step 3: Ask Jev three questions about every article
  This is the node the rest of the workflow is built around.
  Jev isn't a chat model. You hand it a state and a set of typed questions, and it returns typed values with calibrated confidence, in well under a second. It generates no text, which means it also can't generate the wrong shape of text. Here the state is the article, and the three questions are:
  Does this report a real development? Separates a launch from a think piece about the launch, a funding round from a roundup mentioning the funding round. This question alone removes most of what a naive digest would have sent you.
  Is it about this topic? News search drifts. A query about AI regulation will return a story about a data center zoning dispute, and it isn't wrong exactly, it's just not what you asked for.
  How much does this matter to this team? Scored against the team description and priorities from Config. This is the one that produces materialityThreshold
  .
  Figure 3: Three questions, one call, one article. Every question is evaluated independently against the same state.
  Three separate questions rather than one combined one is deliberate, and it follows TypeSafe's own guidance on composing atomic questions. Each asks one well-scoped thing, each is evaluated independently and in parallel, and the logic that combines them lives in your workflow where you can see it. Ask instead for a single "should this be in the brief" verdict and you've buried three different judgments inside one number you can't tune separately. When the brief starts surfacing the wrong things, you want to know which of the three questions drifted.
  The cost is about $0.00004 per article. Sixty articles across three topics is roughly a quarter of a cent. That number is the entire reason this workflow can filter at all, rather than summarizing everything and calling it filtering.
  Step 4: Group by story, not by article
  One model release produces coverage in six places by lunchtime. A digest that lists six items for one event is padding, and readers notice.
  Material articles are grouped by topic before the brief is written, so multiple outlets covering the same development become a single item with multiple sources behind it. More sources on one story is a signal the story matters, not a reason to take up six lines.
  Step 5: Write one cited brief per topic
  Only now does anything expensive run. You.com Research takes the surviving articles for a topic and writes a short brief with linked citations, and the result goes to Slack.
  Citations are the part that makes this usable in a work context. A brief that asserts a competitor changed their pricing is a rumor; a brief that asserts it with a link is something you can forward to your head of product.
  A topic with nothing material gets no post. Not a post saying there was nothing, which is its own kind of noise. Nothing.
  Figure 4: Four topics had news. The fifth didn't, so it isn't here.
  What it costs
  A morning runs about $0.01 to $0.04. That's the whole thing: searches, one Jev call per article, and a research brief per topic that had news. Quiet days cost less, because the research step doesn't run for topics with nothing in them.
  The structural reason it's cheap is that the expensive step only ever sees material that already passed a filter costing four thousandth of a cent per item. Invert the order, running the heavy model across everything and asking it to be selective, and you pay for sixty articles' worth of inference to receive four articles' worth of value.
  Three things you get that a simpler digest doesn't:
• Silence on quiet days, which is what keeps people opening it in month three.
• A materiality threshold you can tune, rather than a prompt you have to re-litigate every time the output drifts.
• Citations on every claim, so the brief is forwardable.
  Where to go from here
  Run heavier research on big days. The Research operation has a background mode that returns a task id instead of holding the connection open, paired with a Get Research Task operation you poll until the status comes back completed. Worth adding if you raise the effort level on topics with a lot of material; frontier, the highest effort level, runs this way only.
  Add a fourth question. Jev evaluates questions in parallel against the same state, so another one barely moves the response time. An urgency score that promotes a topic to its own immediate message instead of waiting for 8am is a natural next one.
  Scope the sources. The You.com search operation takes include_domains, exclude_domains, and boost_domains, up to 500 values each. Boost the outlets you trust without excluding everything else, or exclude a content farm that keeps ranking for your topics.
  Pull the full articles. Feed the surviving URLs into the Contents operation with a max_age threshold to fetch and archive the pages themselves. Contents takes multiple URLs in one node.
  Change the destination. Slack is in the template because that's where people already are at 8am. Swap in email, a Notion page, or a sheet that accumulates a searchable record of what mattered each day.
  Conclusion
  Filtering has always been the expensive half of a news digest, which is why most of them don't really filter. What changed is the price of a judgment call. At four hundredths of a cent you can ask three careful questions about every article in the firehose and still spend less on the entire filtering step than on one summary from a frontier model. Search wide, filter hard, then write about only what's left.
  Try it now
  Import the workflow to drop every node, wired and configured, into your n8n instance.
  Get the templateFor the full set of operations the You.com node supports, including Answer, Finance Research, and Contents, see the You.com API documentation. For Jev's primitives and how confidence is meant to be used architecturally, see the TypeSafe docs.