<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
  <channel>
    <title>Eeshan Gupta — Writing</title>
    <link>https://eeshangupta.com/writing/</link>
    <description>Notes on RAG evaluation and retrieval experiments.</description>
    <language>en</language>
    <atom:link href="https://eeshangupta.com/feed.xml" rel="self" type="application/rss+xml"/>
    <lastBuildDate>Tue, 10 Nov 2026 04:30:00 +0000</lastBuildDate>
    <item>
      <title>A small local judge passed its pilot. Fresh data killed it.</title>
      <link>https://eeshangupta.com/writing/small-judge-fresh-data/</link>
      <guid isPermaLink="false">https://eeshangpt.github.io/writing/small-judge-fresh-data/</guid>
      <pubDate>Tue, 10 Nov 2026 04:30:00 +0000</pubDate>
      <description>A small local model judged whether RAG answers stick to their source. It ranked six generators correctly in a pilot (Kendall tau-b 0.733) and failed on fresh data (0.048). The cascade idea survived on one dataset.</description>
    </item>
    <item>
      <title>I said GraphRAG beats vector RAG. Then I measured it.</title>
      <link>https://eeshangupta.com/writing/graphrag-vs-tuned-hybrid/</link>
      <guid isPermaLink="false">https://eeshangpt.github.io/writing/graphrag-vs-tuned-hybrid/</guid>
      <pubDate>Sat, 24 Oct 2026 04:30:00 +0000</pubDate>
      <description>A small pre-registered experiment: when a graph retriever and a tuned hybrid retriever get the same context budget, how much of the graph advantage is left? +19.0 F1 points on 2WikiMultiHopQA, +3.5 on MuSiQue, with the limits stated.</description>
    </item>
  </channel>
</rss>
