<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: Jangsik Kim</title>
    <description>The latest articles on DEV Community by Jangsik Kim (@jangsik1986).</description>
    <link>https://dev.to/jangsik1986</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4258075%2F4df34cd5-ebae-41fb-9588-45a623da1423.jpg</url>
      <title>DEV Community: Jangsik Kim</title>
      <link>https://dev.to/jangsik1986</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/jangsik1986"/>
    <language>en</language>
    <item>
      <title>Filtering RSS &amp; Trend Topics in n8n Before Calling an LLM (Using Jev AI) — Looking for Better Patterns!</title>
      <dc:creator>Jangsik Kim</dc:creator>
      <pubDate>Sun, 22 Nov 2026 00:15:38 +0000</pubDate>
      <link>https://dev.to/jangsik1986/filtering-rss-trend-topics-in-n8n-before-calling-an-llm-using-jev-ai-looking-for-better-32gl</link>
      <guid>https://dev.to/jangsik1986/filtering-rss-trend-topics-in-n8n-before-calling-an-llm-using-jev-ai-looking-for-better-32gl</guid>
      <description>&lt;p&gt;Every blog automation starts the same way: a schedule fires, you pull in RSS feeds, Google Trends, news and YouTube titles, and then you hand the whole pile to an expensive LLM and ask "pick three good topics".&lt;/p&gt;

&lt;p&gt;That step is where the money goes. Most of the candidates are noise — yesterday's weather, a holiday that already passed, a brand announcement with no search intent behind it. You are paying frontier-model prices to say "no" 80% of the time.&lt;/p&gt;

&lt;p&gt;So I moved the "no" upstream.&lt;/p&gt;

&lt;h2&gt;
  
  
  The pipeline
&lt;/h2&gt;

&lt;p&gt;My n8n workflow runs a Korean-language blog with five channels: product comparisons, IT guides, personal finance, trend analysis, and single-product reviews. It looks like this:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;schedule → collect seeds (Google Trends / SNS / News RSS / YouTube)
         → dedupe + normalize
         → [Jev batch prepare] → [Jev call] → [Jev apply]   ← new
         → prompt build → LLM topic generation → validation → sheet
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The new part is a three-node detour with a single HTTP Request node inside it. Nothing else in the workflow changed.&lt;/p&gt;

&lt;h2&gt;
  
  
  What Jev is, in one paragraph
&lt;/h2&gt;

&lt;p&gt;Jev (&lt;code&gt;typesafe/jev-1.14&lt;/code&gt;, from TypeSafe AI) is not a chat model. It is a "System One" decision model: you send a state plus a set of questions typed as &lt;code&gt;choice&lt;/code&gt;, &lt;code&gt;score&lt;/code&gt;, or &lt;code&gt;noul&lt;/code&gt; (a yes/no probability), and it returns numbers you can drop straight into an &lt;code&gt;if&lt;/code&gt;. Measured in my own testing: 205–539 ms server-side, 0.27–0.92 s round trip. Output tokens are free; you pay per request.&lt;/p&gt;

&lt;p&gt;That last property is what makes it usable as a filter: one request can carry up to eight questions and still costs one credit.&lt;/p&gt;

&lt;h2&gt;
  
  
  Packing eight seeds into one call
&lt;/h2&gt;

&lt;p&gt;The prepare step groups seeds into batches of eight and builds the request body:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight javascript"&gt;&lt;code&gt;&lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="no"&gt;MODEL&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="s1"&gt;'typesafe/jev-1.14'&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="no"&gt;BATCH&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="mi"&gt;8&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;        &lt;span class="c1"&gt;// Jev's per-request question limit&lt;/span&gt;
&lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="no"&gt;MAX_SEEDS&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="mi"&gt;24&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;   &lt;span class="c1"&gt;// 24 seeds / 8 = 3 requests = 3 credits per day&lt;/span&gt;

&lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="n"&gt;state&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;batch&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;map&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;s&lt;/span&gt; &lt;span class="o"&gt;=&amp;gt;&lt;/span&gt; &lt;span class="sb"&gt;`${s.kind} · ${s.title}`&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="n"&gt;questions&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="n"&gt;batch&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;forEach&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;s&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;i&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;=&amp;gt;&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
  &lt;span class="n"&gt;questions&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s1"&gt;'s'&lt;/span&gt; &lt;span class="o"&gt;+&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;i&lt;/span&gt; &lt;span class="o"&gt;+&lt;/span&gt; &lt;span class="mi"&gt;1&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
    &lt;span class="kd"&gt;type&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s1"&gt;'noul'&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="n"&gt;instructions&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
      &lt;span class="n"&gt;candidate_title&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;s&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;title&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
      &lt;span class="n"&gt;channel_context&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="no"&gt;CHANNEL_CTX&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
      &lt;span class="n"&gt;question&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s1"&gt;'Is `candidate_title` a topic worth writing today for one of these channels?'&lt;/span&gt;
    &lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="n"&gt;criteria&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
      &lt;span class="kc"&gt;true&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;  &lt;span class="s1"&gt;'concrete, actually being searched today, has information or purchase intent'&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
      &lt;span class="kc"&gt;false&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s1"&gt;'vague, trivial, out of season, or no searchable or revenue demand'&lt;/span&gt;
    &lt;span class="p"&gt;}&lt;/span&gt;
  &lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="n"&gt;json&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="n"&gt;body&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="n"&gt;model&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="no"&gt;MODEL&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;state&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;questions&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The HTTP Request node is deliberately boring:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight http"&gt;&lt;code&gt;&lt;span class="no"&gt;POST&lt;/span&gt; &lt;span class="n"&gt;https&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="c1"&gt;//thejevai.com/v1/systemone&lt;/span&gt;
&lt;span class="nc"&gt;Authorization&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="nc"&gt;Bearer&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt; $&lt;span class="n"&gt;env&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="no"&gt;JEV_API_KEY&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;
&lt;span class="nc"&gt;Body&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="no"&gt;JSON&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;stringify&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;$&lt;span class="n"&gt;json&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;body&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Two settings matter more than the URL:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Never Error&lt;/strong&gt; on, &lt;code&gt;onError: continue&lt;/code&gt;, three retries. A filter that can stop the pipeline is worse than no filter at all.&lt;/li&gt;
&lt;li&gt;Seeds are picked round-robin across sources (&lt;code&gt;trend&lt;/code&gt; → &lt;code&gt;sns&lt;/code&gt; → &lt;code&gt;news&lt;/code&gt; → &lt;code&gt;yt&lt;/code&gt;), so one noisy feed cannot eat the whole batch.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Applying the answers
&lt;/h2&gt;

&lt;p&gt;The response comes back keyed by question id, with &lt;code&gt;answers.s1.noul&lt;/code&gt; holding a probability between 0 and 1. I keep everything at or above &lt;strong&gt;0.45&lt;/strong&gt; and re-attach each answer to its original seed.&lt;/p&gt;

&lt;p&gt;Two guards, both learned the hard way:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;A floor on kept seeds.&lt;/strong&gt; If the filter is too aggressive on a quiet day (&lt;code&gt;kept &amp;lt; 8&lt;/code&gt;), the highest-scoring rejected seeds are restored. The filter narrows the funnel; it must not empty it.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Fail open.&lt;/strong&gt; If a batch errors out, those seeds pass through untouched. A dead filter degrades to yesterday's behaviour instead of silently producing nothing.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Real numbers from one morning run: 24 seeds checked, 15 kept, 9 dropped, 3 credits, about 900 ms total, 0 failed batches.&lt;/p&gt;

&lt;p&gt;The individual judgments looked about right:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Candidate (translated)&lt;/th&gt;
&lt;th&gt;noul&lt;/th&gt;
&lt;th&gt;Verdict&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;2026 budget wireless keyboard recommendations&lt;/td&gt;
&lt;td&gt;0.86&lt;/td&gt;
&lt;td&gt;keep&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Olive Young global sunscreen ranking (K-beauty)&lt;/td&gt;
&lt;td&gt;0.78&lt;/td&gt;
&lt;td&gt;keep&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Youth savings account guide (Korean gov. scheme)&lt;/td&gt;
&lt;td&gt;0.72&lt;/td&gt;
&lt;td&gt;keep&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;iPhone 18 rumors&lt;/td&gt;
&lt;td&gt;0.63&lt;/td&gt;
&lt;td&gt;keep&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Samsung 3nm node status&lt;/td&gt;
&lt;td&gt;0.46&lt;/td&gt;
&lt;td&gt;borderline keep&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Chuseok holiday highway traffic&lt;/td&gt;
&lt;td&gt;0.32&lt;/td&gt;
&lt;td&gt;drop&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Yesterday's Seoul weather&lt;/td&gt;
&lt;td&gt;0.11&lt;/td&gt;
&lt;td&gt;drop&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;The two seasonal rejects are the ones I care about most. A plain embedding-similarity filter would happily keep "Chuseok holiday traffic" because it &lt;em&gt;looks&lt;/em&gt; like the travel and finance topics that do well. "Is this worth writing today, in this country" is not a similarity question.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why not embeddings or a small local model?
&lt;/h2&gt;

&lt;p&gt;Honest answer: I have not measured them head to head yet, and that is part of why I am posting this.&lt;/p&gt;

&lt;p&gt;My reasoning so far:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Embeddings&lt;/strong&gt; need an anchor and a threshold. Great at dedupe and near-duplicate detection, which I still do separately. Weak at "is this worth writing today", because seasonality and intent are not distances.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;A local classifier&lt;/strong&gt; needs labelled data I do not have. At 24 seeds a day, I would be labelling them myself for weeks.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;A frontier LLM&lt;/strong&gt; does all of it well and costs roughly two orders of magnitude more per item at this volume.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The decision model sits in between: no labels, no vectors, a plain-language rubric I can edit without retraining, and a sub-second answer. It also decides &lt;em&gt;whether&lt;/em&gt; to spend the big-model call, which is where the saving actually comes from.&lt;/p&gt;

&lt;h2&gt;
  
  
  The cost shape
&lt;/h2&gt;

&lt;p&gt;Because billing is per request and not per token, batching is the whole game:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;8 seeds per request x 3 requests a day = 3 credits a day, about 90 a month&lt;/li&gt;
&lt;li&gt;the free signup credit was 10 (expiring in three days), so this needs a paid tier to become a daily habit ($10 for 100k credits in my case)&lt;/li&gt;
&lt;li&gt;dropping &lt;code&gt;MAX_SEEDS&lt;/code&gt; to 16 halves the daily cost&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  What I would like to hear from you
&lt;/h2&gt;

&lt;p&gt;If you have built this layer before, I would love your take on:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Threshold tuning.&lt;/strong&gt; 0.45 came from about 30 samples. Do you drift it per source, or per weekday, or keep one number?&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Batch size.&lt;/strong&gt; Is 8 the sweet spot for you, or does judgment quality degrade before the limit? I have not run smaller batches against the same seeds yet.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Alternatives.&lt;/strong&gt; Is there a local classifier or an embedding-plus-reranker setup that beats a decision model on 20-30 short texts a day, on a cheap box?&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Failure policy.&lt;/strong&gt; Fail open (mine) or fail closed? I chose open so the blog never misses a day, but that means a broken filter can silently ship bad topics.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;I am cross-posting this on Medium and Quora as well, and I will fold the best answers back into the workflow. If you are running something similar, I would rather learn from your scars than collect my own.&lt;/p&gt;

</description>
      <category>n8n</category>
      <category>ai</category>
      <category>automation</category>
      <category>webdev</category>
    </item>
  </channel>
</rss>
