<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:content="http://purl.org/rss/1.0/modules/content/"><channel><title>Retrieval on Local First AI</title><link>https://localfirstai.eu/tags/retrieval/</link><description>Recent content in Retrieval on Local First AI</description><generator>Hugo</generator><language>en-us</language><lastBuildDate>Tue, 22 Sep 2026 00:00:00 +0000</lastBuildDate><atom:link href="https://localfirstai.eu/tags/retrieval/index.xml" rel="self" type="application/rss+xml"/><item><title>We Trained It Three Times. Then We Stopped.</title><link>https://localfirstai.eu/posts/2026-09-22-we-trained-it-three-times-then-stopped/</link><pubDate>Tue, 22 Sep 2026 00:00:00 +0000</pubDate><guid>https://localfirstai.eu/posts/2026-09-22-we-trained-it-three-times-then-stopped/</guid><description>We wanted a small model on a phone that answers Spanish property questions the way a professional would: facts from our study notes via retrieval, the manner trained into the weights. Three LoRA runs later the manner was installed and the model was worse every time. The build that won has no trained weights at all. Here&amp;#39;s the path, the numbers, and the hypothesis we&amp;#39;d like someone to break.</description></item></channel></rss>