<?xml version="1.0" encoding="utf-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
    <channel>
        <title>How we taught agents to use good retrieval - Hanna Lichtenberg, Mixedbread AI</title>
        <link>https://video.ut0pia.org/videos/watch/219dca16-d2aa-433e-bb15-277c1e25e2c0</link>
        <description>RAG is dead. Again. Vector search is useless. All you need is BM25. Not even BM25, all you need is grep. Or maybe even just cat+ls. If you care at all about agents, you probably read a variation of this as part of your daily routine. In a way, isn't it true that semantic search is full of failure cases? And yet, in all sorts of knowledge tasks, whether it be Deep Research, financial analysis or legal research, grep does not seem to cut it. The Oracle Gap, the performance difference between perfect retrieval and grep-based retrieval, is well into the double digits percents. In practice, this means that your agent fails to surface that one hidden clause in that 187 pages contract. Or it doesn't properly notice that Q4 results were amended. In the end, this means that a human has to re-do all of its work, erasing all benefits. But if keyword search does not work, and semantic search is dead, then what is the way out? We argue that the reason for the impressive performance on simple, lexical search methods is simply because models were never taught to use better tools. When using weak tools, they run into the limits of these tools. When provided with better search, they write queries for the weak tools they know, and semantic search fails as a result. Join us for this session to hear about how we are addressing this problem, co-designing agents with state-of-the-art retrieval tools to teach them that they have more than one tool in their best. Speakers: Hanna Lichtenberg (Mixedbread AI): Hanna is an AI Engineer at Mixedbread, working on agentic retrieval research and agent infrastructure. X/Twitter: https://x.com/hannaLicht LinkedIn: https://www.linkedin.com/in/hanna-lichtenberg-64778b221/ GitHub: https://github.com/HannaLicht</description>
        <lastBuildDate>Wed, 08 Jul 2026 12:08:30 GMT</lastBuildDate>
        <docs>https://validator.w3.org/feed/docs/rss2.html</docs>
        <generator>PeerTube - https://video.ut0pia.org</generator>
        <image>
            <title>How we taught agents to use good retrieval - Hanna Lichtenberg, Mixedbread AI</title>
            <url>https://video.ut0pia.org/lazy-static/avatars/0287a09a-aae7-4840-9843-b416426e7046.webp</url>
            <link>https://video.ut0pia.org/videos/watch/219dca16-d2aa-433e-bb15-277c1e25e2c0</link>
        </image>
        <copyright>All rights reserved, unless otherwise specified in the terms specified at https://video.ut0pia.org/about and potential licenses granted by each content's rightholder.</copyright>
        <atom:link href="https://video.ut0pia.org/feeds/video-comments.xml?videoId=219dca16-d2aa-433e-bb15-277c1e25e2c0" rel="self" type="application/rss+xml"/>
    </channel>
</rss>