{"id":5654,"date":"2024-04-11T09:58:54","date_gmt":"2024-04-11T09:58:54","guid":{"rendered":"https:\/\/laizee.ai\/?p=5654"},"modified":"2024-04-17T06:42:54","modified_gmt":"2024-04-17T06:42:54","slug":"was-ist-rag-retrieval-augmented-generation","status":"publish","type":"post","link":"https:\/\/laizee.ai\/blog\/was-ist-rag-retrieval-augmented-generation","title":{"rendered":"Was ist Retrieval Augmented Generation (RAG) und wie funktioniert es?"},"content":{"rendered":"<style>.kadence-column5654_7545ce-13 > .kt-inside-inner-col,.kadence-column5654_7545ce-13 > .kt-inside-inner-col:before{border-top-left-radius:0px;border-top-right-radius:0px;border-bottom-right-radius:0px;border-bottom-left-radius:0px;}.kadence-column5654_7545ce-13 > .kt-inside-inner-col{column-gap:var(--global-kb-gap-sm, 1rem);}.kadence-column5654_7545ce-13 > .kt-inside-inner-col{flex-direction:column;}.kadence-column5654_7545ce-13 > .kt-inside-inner-col > .aligncenter{width:100%;}.kadence-column5654_7545ce-13 > .kt-inside-inner-col:before{opacity:0.3;}.kadence-column5654_7545ce-13{position:relative;}@media all and (max-width: 1024px){.kadence-column5654_7545ce-13 > .kt-inside-inner-col{flex-direction:column;justify-content:center;}}@media all and (max-width: 767px){.kadence-column5654_7545ce-13 > .kt-inside-inner-col{flex-direction:column;justify-content:center;}}<\/style>\n<div class=\"wp-block-kadence-column kadence-column5654_7545ce-13\"><div class=\"kt-inside-inner-col\"><style>.wp-block-kadence-advancedheading.kt-adv-heading5654_c1ded5-ea, .wp-block-kadence-advancedheading.kt-adv-heading5654_c1ded5-ea[data-kb-block=\"kb-adv-heading5654_c1ded5-ea\"]{text-align:left;font-style:normal;}.wp-block-kadence-advancedheading.kt-adv-heading5654_c1ded5-ea mark.kt-highlight, .wp-block-kadence-advancedheading.kt-adv-heading5654_c1ded5-ea[data-kb-block=\"kb-adv-heading5654_c1ded5-ea\"] mark.kt-highlight{font-style:normal;color:#f76a0c;-webkit-box-decoration-break:clone;box-decoration-break:clone;padding-top:0px;padding-right:0px;padding-bottom:0px;padding-left:0px;}.wp-block-kadence-advancedheading.kt-adv-heading5654_c1ded5-ea img.kb-inline-image, .wp-block-kadence-advancedheading.kt-adv-heading5654_c1ded5-ea[data-kb-block=\"kb-adv-heading5654_c1ded5-ea\"] img.kb-inline-image{width:150px;vertical-align:baseline;}<\/style>\n<p class=\"kt-adv-heading5654_c1ded5-ea wp-block-kadence-advancedheading\" data-kb-block=\"kb-adv-heading5654_c1ded5-ea\">Seit November 2021 ist das Thema ChatGPT und die dahinterliegende Technologie der Large Language Models (LLMs) all\u00fcbergreifend. Aus einer rein technischen Perspektive basiert die Funktionsweise von LLMs auf Next-Word-Prediction. Das bedeutet, dass die Vorhersage eines Wortes auf der Grundlage vorangegangener W\u00f6rter erfolgt. Aufgrund der enormen Gr\u00f6\u00dfe an Trainingsdaten und Modellparameter \u2013 GPT 3.5 beispielsweise besteht aus 175 Milliarden Parameter \u2013 k\u00f6nnen die LLMs verschiedene Sachverhalte \u201everstehen\u201c und komplexere linguistische sowie semantische Muster erkennen\/reproduzieren.<\/p>\n<\/div><\/div>\n\n\n<style>.kadence-column5654_bcdae7-e1 > .kt-inside-inner-col,.kadence-column5654_bcdae7-e1 > .kt-inside-inner-col:before{border-top-left-radius:0px;border-top-right-radius:0px;border-bottom-right-radius:0px;border-bottom-left-radius:0px;}.kadence-column5654_bcdae7-e1 > .kt-inside-inner-col{column-gap:var(--global-kb-gap-sm, 1rem);}.kadence-column5654_bcdae7-e1 > .kt-inside-inner-col{flex-direction:column;}.kadence-column5654_bcdae7-e1 > .kt-inside-inner-col > .aligncenter{width:100%;}.kadence-column5654_bcdae7-e1 > .kt-inside-inner-col:before{opacity:0.3;}.kadence-column5654_bcdae7-e1{position:relative;}.kadence-column5654_bcdae7-e1, .kt-inside-inner-col > .kadence-column5654_bcdae7-e1:not(.specificity){margin-top:var(--global-kb-spacing-lg, 3rem);}@media all and (max-width: 1024px){.kadence-column5654_bcdae7-e1 > .kt-inside-inner-col{flex-direction:column;justify-content:center;}}@media all and (max-width: 767px){.kadence-column5654_bcdae7-e1 > .kt-inside-inner-col{flex-direction:column;justify-content:center;}}<\/style>\n<div class=\"wp-block-kadence-column kadence-column5654_bcdae7-e1\"><div class=\"kt-inside-inner-col\"><style>.wp-block-kadence-advancedheading.kt-adv-heading5654_061441-82, .wp-block-kadence-advancedheading.kt-adv-heading5654_061441-82[data-kb-block=\"kb-adv-heading5654_061441-82\"]{text-align:left;font-size:var(--global-kb-font-size-lg, 2rem);font-style:normal;}.wp-block-kadence-advancedheading.kt-adv-heading5654_061441-82 mark.kt-highlight, .wp-block-kadence-advancedheading.kt-adv-heading5654_061441-82[data-kb-block=\"kb-adv-heading5654_061441-82\"] mark.kt-highlight{font-style:normal;color:#f76a0c;-webkit-box-decoration-break:clone;box-decoration-break:clone;padding-top:0px;padding-right:0px;padding-bottom:0px;padding-left:0px;}.wp-block-kadence-advancedheading.kt-adv-heading5654_061441-82 img.kb-inline-image, .wp-block-kadence-advancedheading.kt-adv-heading5654_061441-82[data-kb-block=\"kb-adv-heading5654_061441-82\"] img.kb-inline-image{width:150px;vertical-align:baseline;}<\/style>\n<h2 class=\"kt-adv-heading5654_061441-82 wp-block-kadence-advancedheading\" data-kb-block=\"kb-adv-heading5654_061441-82\"><strong><strong>Limitationen<\/strong><\/strong><\/h2>\n\n\n<style>.wp-block-kadence-advancedheading.kt-adv-heading5654_fdb34c-1a, .wp-block-kadence-advancedheading.kt-adv-heading5654_fdb34c-1a[data-kb-block=\"kb-adv-heading5654_fdb34c-1a\"]{text-align:left;font-style:normal;}.wp-block-kadence-advancedheading.kt-adv-heading5654_fdb34c-1a mark.kt-highlight, .wp-block-kadence-advancedheading.kt-adv-heading5654_fdb34c-1a[data-kb-block=\"kb-adv-heading5654_fdb34c-1a\"] mark.kt-highlight{font-style:normal;color:#f76a0c;-webkit-box-decoration-break:clone;box-decoration-break:clone;padding-top:0px;padding-right:0px;padding-bottom:0px;padding-left:0px;}.wp-block-kadence-advancedheading.kt-adv-heading5654_fdb34c-1a img.kb-inline-image, .wp-block-kadence-advancedheading.kt-adv-heading5654_fdb34c-1a[data-kb-block=\"kb-adv-heading5654_fdb34c-1a\"] img.kb-inline-image{width:150px;vertical-align:baseline;}<\/style>\n<p class=\"kt-adv-heading5654_fdb34c-1a wp-block-kadence-advancedheading\" data-kb-block=\"kb-adv-heading5654_fdb34c-1a\">Aber hinter dem ganzen Hype rund um LLMs gibt es auch gerechtfertigte Kritik und Performance-Probleme:<br>LLMs werden auf einer enorm gro\u00dfen Datenmenge trainiert. Hierzu werden zum Beispiel Texte aus Wikipedia, Nachrichten oder dem sogenannten CommonCrawl verwendet. Die Inhalte dieser Datens\u00e4tze sind in den meisten F\u00e4llen sehr allgemein und wenig spezifisch. Nischen-Themen und Informationen, die beispielsweise aus weniger bekannten oder linzenztechnisch gesch\u00fctzten B\u00fcchern stammen, sind hier nicht zu finden.<\/p>\n\n\n<style>.wp-block-kadence-image.kb-image5654_0eaead-9c:not(.kb-specificity-added):not(.kb-extra-specificity-added){margin-top:-40px;margin-right:64px;margin-bottom:-10px;margin-left:64px;}.kb-image5654_0eaead-9c.kb-image-is-ratio-size, .kb-image5654_0eaead-9c .kb-image-is-ratio-size{max-width:-200px;width:100%;}.wp-block-kadence-column > .kt-inside-inner-col > .kb-image5654_0eaead-9c.kb-image-is-ratio-size, .wp-block-kadence-column > .kt-inside-inner-col > .kb-image5654_0eaead-9c .kb-image-is-ratio-size{align-self:unset;}.kb-image5654_0eaead-9c figure{max-width:-200px;}.kb-image5654_0eaead-9c .image-is-svg, .kb-image5654_0eaead-9c .image-is-svg img{width:100%;}.kb-image5654_0eaead-9c .kb-image-has-overlay:after{opacity:0.3;}<\/style>\n<div class=\"wp-block-kadence-image kb-image5654_0eaead-9c\"><figure class=\"aligncenter\"><img loading=\"lazy\" decoding=\"async\" width=\"942\" height=\"633\" src=\"https:\/\/laizee.ai\/wp-content\/uploads\/2024\/04\/GPT_png.png\" alt=\"Tabelle der Trainingsdaten f\u00fcr GPT 3.5\" class=\"kb-img wp-image-5814\" srcset=\"https:\/\/laizee.ai\/wp-content\/uploads\/2024\/04\/GPT_png.png 942w, https:\/\/laizee.ai\/wp-content\/uploads\/2024\/04\/GPT_png-300x202.png 300w, https:\/\/laizee.ai\/wp-content\/uploads\/2024\/04\/GPT_png-768x516.png 768w\" sizes=\"auto, (max-width: 942px) 100vw, 942px\" \/><\/figure><\/div>\n\n\n\n<p>Sp\u00e4testens, wenn man einem Chatbot wie ChatGPT Anfragen zu Themen der eigenen Expertise &#8211; auch Prompts genannt &#8211; stellt, sieht man, dass diese Systeme schnell an Ihre Grenzen sto\u00dfen. Da die zugrundeliegenden KI-Modelle nicht ausreichend mit Informationen zu etwaigen Themen trainiert wurden, erfinden diese Modelle Antworten, die auf den ersten Blick sehr plausibel erscheinen. Im Bereich der KI bezeichnet man dieses Ph\u00e4nomen als eine \u201eHalluzination\u201c. Anders ausgedr\u00fcckt k\u00f6nnte man auch sagen, dass das Modell <em>\u00fcberzeugend<\/em> Fakten erfindet und damit Falschinformationen erzeugt.<\/p>\n\n\n<style>.wp-block-kadence-advancedheading.kt-adv-heading5654_abe9b1-6d, .wp-block-kadence-advancedheading.kt-adv-heading5654_abe9b1-6d[data-kb-block=\"kb-adv-heading5654_abe9b1-6d\"]{text-align:left;font-style:normal;}.wp-block-kadence-advancedheading.kt-adv-heading5654_abe9b1-6d mark.kt-highlight, .wp-block-kadence-advancedheading.kt-adv-heading5654_abe9b1-6d[data-kb-block=\"kb-adv-heading5654_abe9b1-6d\"] mark.kt-highlight{font-style:normal;color:#f76a0c;-webkit-box-decoration-break:clone;box-decoration-break:clone;padding-top:0px;padding-right:0px;padding-bottom:0px;padding-left:0px;}.wp-block-kadence-advancedheading.kt-adv-heading5654_abe9b1-6d img.kb-inline-image, .wp-block-kadence-advancedheading.kt-adv-heading5654_abe9b1-6d[data-kb-block=\"kb-adv-heading5654_abe9b1-6d\"] img.kb-inline-image{width:150px;vertical-align:baseline;}<\/style>\n<p class=\"kt-adv-heading5654_abe9b1-6d wp-block-kadence-advancedheading\" data-kb-block=\"kb-adv-heading5654_abe9b1-6d\">Problematisch ist au\u00dferdem, dass LLMs nur mit Daten trainiert werden, die bis zu einem bestimmten Stichtag reichen. Bei GPT3.5 beispielsweise liegt dieser im Juni 2021. Alle Informationen und Texte, die nach diesem Datum ver\u00f6ffentlicht wurden, stehen dem KI-Modell ensprechend nicht zur Verf\u00fcgung. Ein regelm\u00e4\u00dfiges Nachtrainieren der Modelle mit weiteren Daten kann dabei helfen, die Datenl\u00fccke zu schlie\u00dfen, ist jedoch sehr kosten- und zeitintensiv.<\/p>\n\n\n<style>.wp-block-kadence-advancedheading.kt-adv-heading5654_18f334-2d, .wp-block-kadence-advancedheading.kt-adv-heading5654_18f334-2d[data-kb-block=\"kb-adv-heading5654_18f334-2d\"]{text-align:left;font-style:normal;}.wp-block-kadence-advancedheading.kt-adv-heading5654_18f334-2d mark.kt-highlight, .wp-block-kadence-advancedheading.kt-adv-heading5654_18f334-2d[data-kb-block=\"kb-adv-heading5654_18f334-2d\"] mark.kt-highlight{font-style:normal;color:#f76a0c;-webkit-box-decoration-break:clone;box-decoration-break:clone;padding-top:0px;padding-right:0px;padding-bottom:0px;padding-left:0px;}.wp-block-kadence-advancedheading.kt-adv-heading5654_18f334-2d img.kb-inline-image, .wp-block-kadence-advancedheading.kt-adv-heading5654_18f334-2d[data-kb-block=\"kb-adv-heading5654_18f334-2d\"] img.kb-inline-image{width:150px;vertical-align:baseline;}<\/style>\n<p class=\"kt-adv-heading5654_18f334-2d wp-block-kadence-advancedheading\" data-kb-block=\"kb-adv-heading5654_18f334-2d\">Die Limitationen und Probleme von LLMs schr\u00e4nken die Nutzbarkeit f\u00fcr viele User stark ein, insbesondere in hochspezialisierten Dom\u00e4nen.<br>Wie kann man als Unternehmen aus etwaigen Dom\u00e4nen trotzdem die St\u00e4rken von LLMs nutzen und das Wissen dieser um unternehmensinterne Informationen erweitern? Eine M\u00f6glichkeit fehlendes Wissen schnell und kosteng\u00fcnstig in das eigene System\/Modell zu integrieren ist RAG. <\/p>\n<\/div><\/div>\n\n\n<style>.kadence-column5654_4cb713-3f > .kt-inside-inner-col,.kadence-column5654_4cb713-3f > .kt-inside-inner-col:before{border-top-left-radius:0px;border-top-right-radius:0px;border-bottom-right-radius:0px;border-bottom-left-radius:0px;}.kadence-column5654_4cb713-3f > .kt-inside-inner-col{column-gap:var(--global-kb-gap-sm, 1rem);}.kadence-column5654_4cb713-3f > .kt-inside-inner-col{flex-direction:column;}.kadence-column5654_4cb713-3f > .kt-inside-inner-col > .aligncenter{width:100%;}.kadence-column5654_4cb713-3f > .kt-inside-inner-col:before{opacity:0.3;}.kadence-column5654_4cb713-3f{position:relative;}.kadence-column5654_4cb713-3f, .kt-inside-inner-col > .kadence-column5654_4cb713-3f:not(.specificity){margin-top:var(--global-kb-spacing-lg, 3rem);}@media all and (max-width: 1024px){.kadence-column5654_4cb713-3f > .kt-inside-inner-col{flex-direction:column;justify-content:center;}}@media all and (max-width: 767px){.kadence-column5654_4cb713-3f > .kt-inside-inner-col{flex-direction:column;justify-content:center;}}<\/style>\n<div class=\"wp-block-kadence-column kadence-column5654_4cb713-3f\"><div class=\"kt-inside-inner-col\"><style>.wp-block-kadence-advancedheading.kt-adv-heading5654_3bc5de-14, .wp-block-kadence-advancedheading.kt-adv-heading5654_3bc5de-14[data-kb-block=\"kb-adv-heading5654_3bc5de-14\"]{text-align:left;font-size:var(--global-kb-font-size-lg, 2rem);font-style:normal;}.wp-block-kadence-advancedheading.kt-adv-heading5654_3bc5de-14 mark.kt-highlight, .wp-block-kadence-advancedheading.kt-adv-heading5654_3bc5de-14[data-kb-block=\"kb-adv-heading5654_3bc5de-14\"] mark.kt-highlight{font-style:normal;color:#f76a0c;-webkit-box-decoration-break:clone;box-decoration-break:clone;padding-top:0px;padding-right:0px;padding-bottom:0px;padding-left:0px;}.wp-block-kadence-advancedheading.kt-adv-heading5654_3bc5de-14 img.kb-inline-image, .wp-block-kadence-advancedheading.kt-adv-heading5654_3bc5de-14[data-kb-block=\"kb-adv-heading5654_3bc5de-14\"] img.kb-inline-image{width:150px;vertical-align:baseline;}<\/style>\n<h2 class=\"kt-adv-heading5654_3bc5de-14 wp-block-kadence-advancedheading\" data-kb-block=\"kb-adv-heading5654_3bc5de-14\"><strong><strong>Was ist RAG und wie funktioniert es?<\/strong><\/strong><\/h2>\n\n\n<style>.wp-block-kadence-advancedheading.kt-adv-heading5654_f2ebb1-1c, .wp-block-kadence-advancedheading.kt-adv-heading5654_f2ebb1-1c[data-kb-block=\"kb-adv-heading5654_f2ebb1-1c\"]{text-align:left;font-style:normal;}.wp-block-kadence-advancedheading.kt-adv-heading5654_f2ebb1-1c mark.kt-highlight, .wp-block-kadence-advancedheading.kt-adv-heading5654_f2ebb1-1c[data-kb-block=\"kb-adv-heading5654_f2ebb1-1c\"] mark.kt-highlight{font-style:normal;color:#f76a0c;-webkit-box-decoration-break:clone;box-decoration-break:clone;padding-top:0px;padding-right:0px;padding-bottom:0px;padding-left:0px;}.wp-block-kadence-advancedheading.kt-adv-heading5654_f2ebb1-1c img.kb-inline-image, .wp-block-kadence-advancedheading.kt-adv-heading5654_f2ebb1-1c[data-kb-block=\"kb-adv-heading5654_f2ebb1-1c\"] img.kb-inline-image{width:150px;vertical-align:baseline;}<\/style>\n<p class=\"kt-adv-heading5654_f2ebb1-1c wp-block-kadence-advancedheading\" data-kb-block=\"kb-adv-heading5654_f2ebb1-1c\">RAG steht f\u00fcr <em>Retrieval Augmented Generation<\/em>. Auf Deutsch k\u00f6nne man es \u00fcbersetzen als \u201eum Abfragen erweiterte Generierung\u201c. RAG ist ein Prozess zur Optimierung und Spezialisierung von LLM-Ausgaben. Im Wesentlichen geht es dabei darum, dass wir das implizit gelernte Wissen eines LLMs um anwendungsspezifische Informationen erweitern. Das LLM ist anschlie\u00dfen in der Lage, anwendungsspezifische Anfragen eines Users auf Grundlage seiner Wissenserweiteurng korrekt zu beantworten. <\/p>\n\n\n<style>.wp-block-kadence-advancedheading.kt-adv-heading5654_ea4d64-33, .wp-block-kadence-advancedheading.kt-adv-heading5654_ea4d64-33[data-kb-block=\"kb-adv-heading5654_ea4d64-33\"]{text-align:left;font-style:normal;}.wp-block-kadence-advancedheading.kt-adv-heading5654_ea4d64-33 mark.kt-highlight, .wp-block-kadence-advancedheading.kt-adv-heading5654_ea4d64-33[data-kb-block=\"kb-adv-heading5654_ea4d64-33\"] mark.kt-highlight{font-style:normal;color:#f76a0c;-webkit-box-decoration-break:clone;box-decoration-break:clone;padding-top:0px;padding-right:0px;padding-bottom:0px;padding-left:0px;}.wp-block-kadence-advancedheading.kt-adv-heading5654_ea4d64-33 img.kb-inline-image, .wp-block-kadence-advancedheading.kt-adv-heading5654_ea4d64-33[data-kb-block=\"kb-adv-heading5654_ea4d64-33\"] img.kb-inline-image{width:150px;vertical-align:baseline;}<\/style>\n<p class=\"kt-adv-heading5654_ea4d64-33 wp-block-kadence-advancedheading\" data-kb-block=\"kb-adv-heading5654_ea4d64-33\">In der einfachsten Form baut der RAG-Prozess auf drei Komponenten auf:<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>Einer Vektodatenbank, in der anwendungsspezifische Informationsbl\u00f6cke gespeichert werden.<\/li>\n\n\n\n<li>Einem Retrieval-System, welches die Vektordatenbank nach anfragerelevanten Informationen durchsucht.<\/li>\n\n\n\n<li>Dem LLM, das userfreundliche Antworten aus den gefundenen Informationen generiert.<\/li>\n<\/ul>\n\n\n<style>.wp-block-kadence-advancedheading.kt-adv-heading5654_375ee8-e7, .wp-block-kadence-advancedheading.kt-adv-heading5654_375ee8-e7[data-kb-block=\"kb-adv-heading5654_375ee8-e7\"]{text-align:left;font-style:normal;}.wp-block-kadence-advancedheading.kt-adv-heading5654_375ee8-e7 mark.kt-highlight, .wp-block-kadence-advancedheading.kt-adv-heading5654_375ee8-e7[data-kb-block=\"kb-adv-heading5654_375ee8-e7\"] mark.kt-highlight{font-style:normal;color:#f76a0c;-webkit-box-decoration-break:clone;box-decoration-break:clone;padding-top:0px;padding-right:0px;padding-bottom:0px;padding-left:0px;}.wp-block-kadence-advancedheading.kt-adv-heading5654_375ee8-e7 img.kb-inline-image, .wp-block-kadence-advancedheading.kt-adv-heading5654_375ee8-e7[data-kb-block=\"kb-adv-heading5654_375ee8-e7\"] img.kb-inline-image{width:150px;vertical-align:baseline;}<\/style>\n<p class=\"kt-adv-heading5654_375ee8-e7 wp-block-kadence-advancedheading\" data-kb-block=\"kb-adv-heading5654_375ee8-e7\">Das R\u00fcckgrat eines RAG-Systems bildet die Vektordatenbank, in welcher die Informationen, die zur Beantwortung von anwendungsspezifischen Anfragen ben\u00f6tigt werden, gespeichert sind. Um diese aufzubauen, werden im wesentlichen folgende sechs Schritte durchgef\u00fchrt:<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li>Sammlung von Dokumenten (z.B. PDF-Dokumente) mit anwendungsspezifischen Informationen und Daten.<\/li>\n\n\n\n<li>Aufbereitung der Dokumente, also beispielsweise Konvertierung von PDF-Dokumenten in maschinenlesbaren Text.<\/li>\n\n\n\n<li>Festlegung einer sogenannten Chunking Strategie, mithilfe derer die Dokumente in Informationsbl\u00f6cke zerlegt werden.<\/li>\n\n\n\n<li>Generierung der Informationsbl\u00f6cke (\u201eChunks\u201c) auf Grundlage der zuvor festgelegten Strategie.<\/li>\n\n\n\n<li>\u00dcberf\u00fchrung der Chunks in Zahlenreihen (Vektoren). Dieser Schritt wird auch als Embedding bezeichnet.<\/li>\n\n\n\n<li>Speichern der \u00fcberf\u00fchrten Chunks in der Vektordatenbank.<\/li>\n<\/ol>\n\n\n<style>.kb-image5654_f970c5-ac .kb-image-has-overlay:after{opacity:0.3;}<\/style>\n<figure class=\"wp-block-kadence-image kb-image5654_f970c5-ac\"><img loading=\"lazy\" decoding=\"async\" width=\"1608\" height=\"1036\" src=\"https:\/\/laizee.ai\/wp-content\/uploads\/2024\/04\/Ablauf_RAG.png\" alt=\"Schritte zum Aufbau eines RAG Retrieval Augmented Generation Systems\" class=\"kb-img wp-image-5816\" srcset=\"https:\/\/laizee.ai\/wp-content\/uploads\/2024\/04\/Ablauf_RAG.png 1608w, https:\/\/laizee.ai\/wp-content\/uploads\/2024\/04\/Ablauf_RAG-300x193.png 300w, https:\/\/laizee.ai\/wp-content\/uploads\/2024\/04\/Ablauf_RAG-1024x660.png 1024w, https:\/\/laizee.ai\/wp-content\/uploads\/2024\/04\/Ablauf_RAG-768x495.png 768w, https:\/\/laizee.ai\/wp-content\/uploads\/2024\/04\/Ablauf_RAG-1536x990.png 1536w\" sizes=\"auto, (max-width: 1608px) 100vw, 1608px\" \/><\/figure>\n\n\n<style>.wp-block-kadence-advancedheading.kt-adv-heading5654_a3a154-40, .wp-block-kadence-advancedheading.kt-adv-heading5654_a3a154-40[data-kb-block=\"kb-adv-heading5654_a3a154-40\"]{text-align:left;font-style:normal;}.wp-block-kadence-advancedheading.kt-adv-heading5654_a3a154-40 mark.kt-highlight, .wp-block-kadence-advancedheading.kt-adv-heading5654_a3a154-40[data-kb-block=\"kb-adv-heading5654_a3a154-40\"] mark.kt-highlight{font-style:normal;color:#f76a0c;-webkit-box-decoration-break:clone;box-decoration-break:clone;padding-top:0px;padding-right:0px;padding-bottom:0px;padding-left:0px;}.wp-block-kadence-advancedheading.kt-adv-heading5654_a3a154-40 img.kb-inline-image, .wp-block-kadence-advancedheading.kt-adv-heading5654_a3a154-40[data-kb-block=\"kb-adv-heading5654_a3a154-40\"] img.kb-inline-image{width:150px;vertical-align:baseline;}<\/style>\n<h6 class=\"kt-adv-heading5654_a3a154-40 wp-block-kadence-advancedheading\" data-kb-block=\"kb-adv-heading5654_a3a154-40\">Wie funktioniert das Retrieval System?<\/h6>\n\n\n<style>.wp-block-kadence-advancedheading.kt-adv-heading5654_f0ed02-3e, .wp-block-kadence-advancedheading.kt-adv-heading5654_f0ed02-3e[data-kb-block=\"kb-adv-heading5654_f0ed02-3e\"]{text-align:left;font-style:normal;}.wp-block-kadence-advancedheading.kt-adv-heading5654_f0ed02-3e mark.kt-highlight, .wp-block-kadence-advancedheading.kt-adv-heading5654_f0ed02-3e[data-kb-block=\"kb-adv-heading5654_f0ed02-3e\"] mark.kt-highlight{font-style:normal;color:#f76a0c;-webkit-box-decoration-break:clone;box-decoration-break:clone;padding-top:0px;padding-right:0px;padding-bottom:0px;padding-left:0px;}.wp-block-kadence-advancedheading.kt-adv-heading5654_f0ed02-3e img.kb-inline-image, .wp-block-kadence-advancedheading.kt-adv-heading5654_f0ed02-3e[data-kb-block=\"kb-adv-heading5654_f0ed02-3e\"] img.kb-inline-image{width:150px;vertical-align:baseline;}<\/style>\n<p class=\"kt-adv-heading5654_f0ed02-3e wp-block-kadence-advancedheading\" data-kb-block=\"kb-adv-heading5654_f0ed02-3e\">W\u00f6rter werden in numerische Vektoren konvertiert. Ein einzelnes Wort wird entsprechend als Zahlenreihe dargestellt. Das spannende an Vektoren ist, dass sich mithilfe dieser semantische \u00c4hnlichkeiten und Zusammenh\u00e4nge zwischen W\u00f6rtern berechnen lassen. Da Vektoren sehr schnell verglichen werden k\u00f6nnen, lassen sich mit einem Retrieval System inhaltlich \u00e4hnliche Dokumente auffinden.<\/p>\n<\/div><\/div>\n\n\n<style>.kadence-column5654_20ea9c-db > .kt-inside-inner-col,.kadence-column5654_20ea9c-db > .kt-inside-inner-col:before{border-top-left-radius:0px;border-top-right-radius:0px;border-bottom-right-radius:0px;border-bottom-left-radius:0px;}.kadence-column5654_20ea9c-db > .kt-inside-inner-col{column-gap:var(--global-kb-gap-sm, 1rem);}.kadence-column5654_20ea9c-db > .kt-inside-inner-col{flex-direction:column;}.kadence-column5654_20ea9c-db > .kt-inside-inner-col > .aligncenter{width:100%;}.kadence-column5654_20ea9c-db > .kt-inside-inner-col:before{opacity:0.3;}.kadence-column5654_20ea9c-db{position:relative;}.kadence-column5654_20ea9c-db, .kt-inside-inner-col > .kadence-column5654_20ea9c-db:not(.specificity){margin-top:var(--global-kb-spacing-lg, 3rem);}@media all and (max-width: 1024px){.kadence-column5654_20ea9c-db > .kt-inside-inner-col{flex-direction:column;justify-content:center;}}@media all and (max-width: 767px){.kadence-column5654_20ea9c-db > .kt-inside-inner-col{flex-direction:column;justify-content:center;}}<\/style>\n<div class=\"wp-block-kadence-column kadence-column5654_20ea9c-db\"><div class=\"kt-inside-inner-col\"><style>.wp-block-kadence-advancedheading.kt-adv-heading5654_e60a56-6b, .wp-block-kadence-advancedheading.kt-adv-heading5654_e60a56-6b[data-kb-block=\"kb-adv-heading5654_e60a56-6b\"]{text-align:left;font-size:var(--global-kb-font-size-lg, 2rem);font-style:normal;}.wp-block-kadence-advancedheading.kt-adv-heading5654_e60a56-6b mark.kt-highlight, .wp-block-kadence-advancedheading.kt-adv-heading5654_e60a56-6b[data-kb-block=\"kb-adv-heading5654_e60a56-6b\"] mark.kt-highlight{font-style:normal;color:#f76a0c;-webkit-box-decoration-break:clone;box-decoration-break:clone;padding-top:0px;padding-right:0px;padding-bottom:0px;padding-left:0px;}.wp-block-kadence-advancedheading.kt-adv-heading5654_e60a56-6b img.kb-inline-image, .wp-block-kadence-advancedheading.kt-adv-heading5654_e60a56-6b[data-kb-block=\"kb-adv-heading5654_e60a56-6b\"] img.kb-inline-image{width:150px;vertical-align:baseline;}<\/style>\n<h2 class=\"kt-adv-heading5654_e60a56-6b wp-block-kadence-advancedheading\" data-kb-block=\"kb-adv-heading5654_e60a56-6b\"><strong><strong>RAG-Prozess<\/strong><\/strong><\/h2>\n\n\n<style>.wp-block-kadence-advancedheading.kt-adv-heading5654_11a351-e4, .wp-block-kadence-advancedheading.kt-adv-heading5654_11a351-e4[data-kb-block=\"kb-adv-heading5654_11a351-e4\"]{text-align:left;font-style:normal;}.wp-block-kadence-advancedheading.kt-adv-heading5654_11a351-e4 mark.kt-highlight, .wp-block-kadence-advancedheading.kt-adv-heading5654_11a351-e4[data-kb-block=\"kb-adv-heading5654_11a351-e4\"] mark.kt-highlight{font-style:normal;color:#f76a0c;-webkit-box-decoration-break:clone;box-decoration-break:clone;padding-top:0px;padding-right:0px;padding-bottom:0px;padding-left:0px;}.wp-block-kadence-advancedheading.kt-adv-heading5654_11a351-e4 img.kb-inline-image, .wp-block-kadence-advancedheading.kt-adv-heading5654_11a351-e4[data-kb-block=\"kb-adv-heading5654_11a351-e4\"] img.kb-inline-image{width:150px;vertical-align:baseline;}<\/style>\n<p class=\"kt-adv-heading5654_11a351-e4 wp-block-kadence-advancedheading\" data-kb-block=\"kb-adv-heading5654_11a351-e4\">Der User schickt eine Anfrage an das System. Diese Anfrage wird mit demselben Verfahren aus Schritt 5 in einen Vektor umgewandelt.<br>Basierend auf dem Vektor der Anfrage wird in der Vektordatenbank nach relevanten Informationen gesucht.<br>Diese Daten werden dann abgerufen (retrieved) und nach unterschiedlichen Faktoren bewertet, geranked und die Top-Ergebnisse anschlie\u00dfend konsolidiert. Diese Daten und Dokumente stellen den <em>Kontext <\/em>dar, mit welchem die Anfrage des Users angereichert wird.<br>Die Anfrage sowie der zuvor ermittelte Kontext werden anschlie\u00dfend an das LLM \u00fcbergeben. Dieses generiert mithilfe der zus\u00e4tzlichen Informationen eine passende Antwort und spielt sie dem User abschlie\u00dfend zur\u00fcck.<\/p>\n\n\n<style>.wp-block-kadence-advancedheading.kt-adv-heading5654_a60ff8-bf, .wp-block-kadence-advancedheading.kt-adv-heading5654_a60ff8-bf[data-kb-block=\"kb-adv-heading5654_a60ff8-bf\"]{text-align:left;font-style:normal;}.wp-block-kadence-advancedheading.kt-adv-heading5654_a60ff8-bf mark.kt-highlight, .wp-block-kadence-advancedheading.kt-adv-heading5654_a60ff8-bf[data-kb-block=\"kb-adv-heading5654_a60ff8-bf\"] mark.kt-highlight{font-style:normal;color:#f76a0c;-webkit-box-decoration-break:clone;box-decoration-break:clone;padding-top:0px;padding-right:0px;padding-bottom:0px;padding-left:0px;}.wp-block-kadence-advancedheading.kt-adv-heading5654_a60ff8-bf img.kb-inline-image, .wp-block-kadence-advancedheading.kt-adv-heading5654_a60ff8-bf[data-kb-block=\"kb-adv-heading5654_a60ff8-bf\"] img.kb-inline-image{width:150px;vertical-align:baseline;}<\/style>\n<h6 class=\"kt-adv-heading5654_a60ff8-bf wp-block-kadence-advancedheading\" data-kb-block=\"kb-adv-heading5654_a60ff8-bf\">\u00dcbersicht der RAG Architektur und Prozess<\/h6>\n\n\n<style>.kb-image5654_c72542-d2 .kb-image-has-overlay:after{opacity:0.3;}<\/style>\n<figure class=\"wp-block-kadence-image kb-image5654_c72542-d2\"><img loading=\"lazy\" decoding=\"async\" width=\"2784\" height=\"1569\" src=\"https:\/\/laizee.ai\/wp-content\/uploads\/2024\/04\/RAG_Prozess.png\" alt=\"\u00dcbersicht der RAG Architektur und Prozess\" class=\"kb-img wp-image-5817\" srcset=\"https:\/\/laizee.ai\/wp-content\/uploads\/2024\/04\/RAG_Prozess.png 2784w, https:\/\/laizee.ai\/wp-content\/uploads\/2024\/04\/RAG_Prozess-300x169.png 300w, https:\/\/laizee.ai\/wp-content\/uploads\/2024\/04\/RAG_Prozess-1024x577.png 1024w, https:\/\/laizee.ai\/wp-content\/uploads\/2024\/04\/RAG_Prozess-768x433.png 768w, https:\/\/laizee.ai\/wp-content\/uploads\/2024\/04\/RAG_Prozess-1536x866.png 1536w, https:\/\/laizee.ai\/wp-content\/uploads\/2024\/04\/RAG_Prozess-2048x1154.png 2048w\" sizes=\"auto, (max-width: 2784px) 100vw, 2784px\" \/><\/figure>\n<\/div><\/div>\n\n\n<style>.kadence-column5654_b97220-b1 > .kt-inside-inner-col,.kadence-column5654_b97220-b1 > .kt-inside-inner-col:before{border-top-left-radius:0px;border-top-right-radius:0px;border-bottom-right-radius:0px;border-bottom-left-radius:0px;}.kadence-column5654_b97220-b1 > .kt-inside-inner-col{column-gap:var(--global-kb-gap-sm, 1rem);}.kadence-column5654_b97220-b1 > .kt-inside-inner-col{flex-direction:column;}.kadence-column5654_b97220-b1 > .kt-inside-inner-col > .aligncenter{width:100%;}.kadence-column5654_b97220-b1 > .kt-inside-inner-col:before{opacity:0.3;}.kadence-column5654_b97220-b1{position:relative;}.kadence-column5654_b97220-b1, .kt-inside-inner-col > .kadence-column5654_b97220-b1:not(.specificity){margin-top:var(--global-kb-spacing-lg, 3rem);}@media all and (max-width: 1024px){.kadence-column5654_b97220-b1 > .kt-inside-inner-col{flex-direction:column;justify-content:center;}}@media all and (max-width: 767px){.kadence-column5654_b97220-b1 > .kt-inside-inner-col{flex-direction:column;justify-content:center;}}<\/style>\n<div class=\"wp-block-kadence-column kadence-column5654_b97220-b1\"><div class=\"kt-inside-inner-col\"><style>.wp-block-kadence-advancedheading.kt-adv-heading5654_954a56-06, .wp-block-kadence-advancedheading.kt-adv-heading5654_954a56-06[data-kb-block=\"kb-adv-heading5654_954a56-06\"]{text-align:left;font-size:var(--global-kb-font-size-lg, 2rem);font-style:normal;}.wp-block-kadence-advancedheading.kt-adv-heading5654_954a56-06 mark.kt-highlight, .wp-block-kadence-advancedheading.kt-adv-heading5654_954a56-06[data-kb-block=\"kb-adv-heading5654_954a56-06\"] mark.kt-highlight{font-style:normal;color:#f76a0c;-webkit-box-decoration-break:clone;box-decoration-break:clone;padding-top:0px;padding-right:0px;padding-bottom:0px;padding-left:0px;}.wp-block-kadence-advancedheading.kt-adv-heading5654_954a56-06 img.kb-inline-image, .wp-block-kadence-advancedheading.kt-adv-heading5654_954a56-06[data-kb-block=\"kb-adv-heading5654_954a56-06\"] img.kb-inline-image{width:150px;vertical-align:baseline;}<\/style>\n<h2 class=\"kt-adv-heading5654_954a56-06 wp-block-kadence-advancedheading\" data-kb-block=\"kb-adv-heading5654_954a56-06\">RAG vs. Finetuning<\/h2>\n\n\n<style>.wp-block-kadence-advancedheading.kt-adv-heading5654_c5cc02-fe, .wp-block-kadence-advancedheading.kt-adv-heading5654_c5cc02-fe[data-kb-block=\"kb-adv-heading5654_c5cc02-fe\"]{text-align:left;font-style:normal;}.wp-block-kadence-advancedheading.kt-adv-heading5654_c5cc02-fe mark.kt-highlight, .wp-block-kadence-advancedheading.kt-adv-heading5654_c5cc02-fe[data-kb-block=\"kb-adv-heading5654_c5cc02-fe\"] mark.kt-highlight{font-style:normal;color:#f76a0c;-webkit-box-decoration-break:clone;box-decoration-break:clone;padding-top:0px;padding-right:0px;padding-bottom:0px;padding-left:0px;}.wp-block-kadence-advancedheading.kt-adv-heading5654_c5cc02-fe img.kb-inline-image, .wp-block-kadence-advancedheading.kt-adv-heading5654_c5cc02-fe[data-kb-block=\"kb-adv-heading5654_c5cc02-fe\"] img.kb-inline-image{width:150px;vertical-align:baseline;}<\/style>\n<p class=\"kt-adv-heading5654_c5cc02-fe wp-block-kadence-advancedheading\" data-kb-block=\"kb-adv-heading5654_c5cc02-fe\">Ein anderer Weg, um ein LLM mit anwendungsspezifischen Informationen zu erweitern, ist die des sogenannten <em>Finetunings<\/em>. Hierf\u00fcr nimmt man ein bereits vortrainiertes LLM und trainiert dieses auf der Grundlage anwendungsspezifischer Informationen nach. Im Rahmen dieses Nachtrainierens werden die bestehenden Modellparameter optimiert. Durch das Finetuning der Modellparameter ist das nachtrainierte Modell im Gegensatz zum Ursprungsmodell in der Lage, auch anwendungsspezifische Aufgaben l\u00f6sen. Das Finetuning bestehender LLMs ist in der Regel deutlich kosten- und zeitintensiver als der RAG-Ansatz.<\/p>\n\n\n<style>.wp-block-kadence-advancedheading.kt-adv-heading5654_57021a-14, .wp-block-kadence-advancedheading.kt-adv-heading5654_57021a-14[data-kb-block=\"kb-adv-heading5654_57021a-14\"]{text-align:left;font-style:normal;}.wp-block-kadence-advancedheading.kt-adv-heading5654_57021a-14 mark.kt-highlight, .wp-block-kadence-advancedheading.kt-adv-heading5654_57021a-14[data-kb-block=\"kb-adv-heading5654_57021a-14\"] mark.kt-highlight{font-style:normal;color:#f76a0c;-webkit-box-decoration-break:clone;box-decoration-break:clone;padding-top:0px;padding-right:0px;padding-bottom:0px;padding-left:0px;}.wp-block-kadence-advancedheading.kt-adv-heading5654_57021a-14 img.kb-inline-image, .wp-block-kadence-advancedheading.kt-adv-heading5654_57021a-14[data-kb-block=\"kb-adv-heading5654_57021a-14\"] img.kb-inline-image{width:150px;vertical-align:baseline;}<\/style>\n<h6 class=\"kt-adv-heading5654_57021a-14 wp-block-kadence-advancedheading\" data-kb-block=\"kb-adv-heading5654_57021a-14\">Vorteile beim Finetuning Ansatz<\/h6>\n\n\n\n<ul class=\"wp-block-list\">\n<li>Das finetuned Modell kann f\u00fcr alle LLM Tasks (z.B. Sentiment Analysen, Entity Recognition) und nicht nur ausschlie\u00dflich f\u00fcr Frage-Antwort Systeme, also Q&amp;A Tasks, eingesetzt werden<\/li>\n\n\n\n<li>Einmalige Kosten (sofern nicht regelm\u00e4\u00dfig neue Informationen angelernt werden m\u00fcssen)<\/li>\n\n\n\n<li>Kein dauerhaft zus\u00e4tzlicher Infrastrukturaufwand<\/li>\n<\/ul>\n\n\n<style>.wp-block-kadence-advancedheading.kt-adv-heading5654_ae6211-ad, .wp-block-kadence-advancedheading.kt-adv-heading5654_ae6211-ad[data-kb-block=\"kb-adv-heading5654_ae6211-ad\"]{text-align:left;font-style:normal;}.wp-block-kadence-advancedheading.kt-adv-heading5654_ae6211-ad mark.kt-highlight, .wp-block-kadence-advancedheading.kt-adv-heading5654_ae6211-ad[data-kb-block=\"kb-adv-heading5654_ae6211-ad\"] mark.kt-highlight{font-style:normal;color:#f76a0c;-webkit-box-decoration-break:clone;box-decoration-break:clone;padding-top:0px;padding-right:0px;padding-bottom:0px;padding-left:0px;}.wp-block-kadence-advancedheading.kt-adv-heading5654_ae6211-ad img.kb-inline-image, .wp-block-kadence-advancedheading.kt-adv-heading5654_ae6211-ad[data-kb-block=\"kb-adv-heading5654_ae6211-ad\"] img.kb-inline-image{width:150px;vertical-align:baseline;}<\/style>\n<h6 class=\"kt-adv-heading5654_ae6211-ad wp-block-kadence-advancedheading\" data-kb-block=\"kb-adv-heading5654_ae6211-ad\">Challenges beim Ansatz des Finetunings<\/h6>\n\n\n\n<ul class=\"wp-block-list\">\n<li>Prinzipiell nicht m\u00f6glich bei closed-sourced 3rd Party Modell Anbietern wie OpenAI<\/li>\n\n\n\n<li>Trainingsprozess tendenziell kosten &amp; zeitintensiv<\/li>\n\n\n\n<li>nicht funktional, wenn Real-Time oder dynamische Daten ber\u00fccksichtigt werden muss<\/li>\n\n\n\n<li>Halluzinationen sind weiterhin ein Problem<\/li>\n\n\n\n<li>Quelle der Antworten nicht r\u00fcckverfolgbar\/nicht nachvollziehbar<\/li>\n<\/ul>\n\n\n<style>.wp-block-kadence-advancedheading.kt-adv-heading5654_4e09bb-fe, .wp-block-kadence-advancedheading.kt-adv-heading5654_4e09bb-fe[data-kb-block=\"kb-adv-heading5654_4e09bb-fe\"]{text-align:left;font-style:normal;}.wp-block-kadence-advancedheading.kt-adv-heading5654_4e09bb-fe mark.kt-highlight, .wp-block-kadence-advancedheading.kt-adv-heading5654_4e09bb-fe[data-kb-block=\"kb-adv-heading5654_4e09bb-fe\"] mark.kt-highlight{font-style:normal;color:#f76a0c;-webkit-box-decoration-break:clone;box-decoration-break:clone;padding-top:0px;padding-right:0px;padding-bottom:0px;padding-left:0px;}.wp-block-kadence-advancedheading.kt-adv-heading5654_4e09bb-fe img.kb-inline-image, .wp-block-kadence-advancedheading.kt-adv-heading5654_4e09bb-fe[data-kb-block=\"kb-adv-heading5654_4e09bb-fe\"] img.kb-inline-image{width:150px;vertical-align:baseline;}<\/style>\n<h6 class=\"kt-adv-heading5654_4e09bb-fe wp-block-kadence-advancedheading\" data-kb-block=\"kb-adv-heading5654_4e09bb-fe\">Vorteile des RAG Prozesses<\/h6>\n\n\n\n<ul class=\"wp-block-list\">\n<li>Schnelle M\u00f6glichkeit LLMs um internes und dom\u00e4nenspezifisches Wissen zu erweitern<\/li>\n\n\n\n<li>G\u00fcnstiges initiatives Setup<\/li>\n\n\n\n<li>Unterst\u00fctzt die dynamische Einbindung von Real-Time-Data<\/li>\n\n\n\n<li>Erm\u00f6glicht die Angabe und Nachvollziehbarkeit der Quellen<\/li>\n\n\n\n<li>Ben\u00f6tigt keine gelabelten Daten<\/li>\n\n\n\n<li>Zugriffsverwaltung von Quellen<\/li>\n<\/ul>\n\n\n<style>.wp-block-kadence-advancedheading.kt-adv-heading5654_5e4712-e0, .wp-block-kadence-advancedheading.kt-adv-heading5654_5e4712-e0[data-kb-block=\"kb-adv-heading5654_5e4712-e0\"]{text-align:left;font-style:normal;}.wp-block-kadence-advancedheading.kt-adv-heading5654_5e4712-e0 mark.kt-highlight, .wp-block-kadence-advancedheading.kt-adv-heading5654_5e4712-e0[data-kb-block=\"kb-adv-heading5654_5e4712-e0\"] mark.kt-highlight{font-style:normal;color:#f76a0c;-webkit-box-decoration-break:clone;box-decoration-break:clone;padding-top:0px;padding-right:0px;padding-bottom:0px;padding-left:0px;}.wp-block-kadence-advancedheading.kt-adv-heading5654_5e4712-e0 img.kb-inline-image, .wp-block-kadence-advancedheading.kt-adv-heading5654_5e4712-e0[data-kb-block=\"kb-adv-heading5654_5e4712-e0\"] img.kb-inline-image{width:150px;vertical-align:baseline;}<\/style>\n<h6 class=\"kt-adv-heading5654_5e4712-e0 wp-block-kadence-advancedheading\" data-kb-block=\"kb-adv-heading5654_5e4712-e0\">Challenges des RAG Prozesses<\/h6>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u201eSearch and Retrieval\u201c-Prozess hat einen gro\u00dfen Einfluss auf die Outputqualit\u00e4t<\/li>\n\n\n\n<li>RAG haupts\u00e4chlich relevant f\u00fcr Frage-Antwort Systeme<\/li>\n\n\n\n<li>Aufbau und dauerhafter Betrieb der Vektordatenbank verursacht laufende Kosten<\/li>\n\n\n\n<li>Anzahl an Input Tokens, die dem LLM \u00fcbergeben werden, vergr\u00f6\u00dfert sich durch den Kontext<\/li>\n<\/ul>\n<\/div><\/div>\n\n\n<style>.kadence-column5654_40e474-e3 > .kt-inside-inner-col,.kadence-column5654_40e474-e3 > .kt-inside-inner-col:before{border-top-left-radius:0px;border-top-right-radius:0px;border-bottom-right-radius:0px;border-bottom-left-radius:0px;}.kadence-column5654_40e474-e3 > .kt-inside-inner-col{column-gap:var(--global-kb-gap-sm, 1rem);}.kadence-column5654_40e474-e3 > .kt-inside-inner-col{flex-direction:column;}.kadence-column5654_40e474-e3 > .kt-inside-inner-col > .aligncenter{width:100%;}.kadence-column5654_40e474-e3 > .kt-inside-inner-col:before{opacity:0.3;}.kadence-column5654_40e474-e3{position:relative;}.kadence-column5654_40e474-e3, .kt-inside-inner-col > .kadence-column5654_40e474-e3:not(.specificity){margin-top:var(--global-kb-spacing-lg, 3rem);}@media all and (max-width: 1024px){.kadence-column5654_40e474-e3 > .kt-inside-inner-col{flex-direction:column;justify-content:center;}}@media all and (max-width: 767px){.kadence-column5654_40e474-e3 > .kt-inside-inner-col{flex-direction:column;justify-content:center;}}<\/style>\n<div class=\"wp-block-kadence-column kadence-column5654_40e474-e3\"><div class=\"kt-inside-inner-col\"><style>.wp-block-kadence-advancedheading.kt-adv-heading5654_55862d-b8, .wp-block-kadence-advancedheading.kt-adv-heading5654_55862d-b8[data-kb-block=\"kb-adv-heading5654_55862d-b8\"]{text-align:left;font-style:normal;}.wp-block-kadence-advancedheading.kt-adv-heading5654_55862d-b8 mark.kt-highlight, .wp-block-kadence-advancedheading.kt-adv-heading5654_55862d-b8[data-kb-block=\"kb-adv-heading5654_55862d-b8\"] mark.kt-highlight{font-style:normal;color:#f76a0c;-webkit-box-decoration-break:clone;box-decoration-break:clone;padding-top:0px;padding-right:0px;padding-bottom:0px;padding-left:0px;}.wp-block-kadence-advancedheading.kt-adv-heading5654_55862d-b8 img.kb-inline-image, .wp-block-kadence-advancedheading.kt-adv-heading5654_55862d-b8[data-kb-block=\"kb-adv-heading5654_55862d-b8\"] img.kb-inline-image{width:150px;vertical-align:baseline;}<\/style>\n<h2 class=\"kt-adv-heading5654_55862d-b8 wp-block-kadence-advancedheading\" data-kb-block=\"kb-adv-heading5654_55862d-b8\">Use Cases<\/h2>\n\n\n<style>.wp-block-kadence-advancedheading.kt-adv-heading5654_2d89b9-4c, .wp-block-kadence-advancedheading.kt-adv-heading5654_2d89b9-4c[data-kb-block=\"kb-adv-heading5654_2d89b9-4c\"]{text-align:left;font-style:normal;}.wp-block-kadence-advancedheading.kt-adv-heading5654_2d89b9-4c mark.kt-highlight, .wp-block-kadence-advancedheading.kt-adv-heading5654_2d89b9-4c[data-kb-block=\"kb-adv-heading5654_2d89b9-4c\"] mark.kt-highlight{font-style:normal;color:#f76a0c;-webkit-box-decoration-break:clone;box-decoration-break:clone;padding-top:0px;padding-right:0px;padding-bottom:0px;padding-left:0px;}.wp-block-kadence-advancedheading.kt-adv-heading5654_2d89b9-4c img.kb-inline-image, .wp-block-kadence-advancedheading.kt-adv-heading5654_2d89b9-4c[data-kb-block=\"kb-adv-heading5654_2d89b9-4c\"] img.kb-inline-image{width:150px;vertical-align:baseline;}<\/style>\n<p class=\"kt-adv-heading5654_2d89b9-4c wp-block-kadence-advancedheading\" data-kb-block=\"kb-adv-heading5654_2d89b9-4c\">RAG ist vor allem bei Q&amp;A Tasks &amp; Informationsextraktion enorm performant, also Aufgaben in denen wir von dem KI-System eine klare Antwort auf eine Frage erwarten. Die Antwort kann zudem mit den Quellen belegt werden, welches das LLM zur Beantwortung einer Anfrage genutzt hat. Das erm\u00f6glicht dem User eine m\u00f6gliche Nachverfolgung oder Einholung weiterer Informationen.<\/p>\n\n\n<style>.wp-block-kadence-advancedheading.kt-adv-heading5654_5c3525-7e, .wp-block-kadence-advancedheading.kt-adv-heading5654_5c3525-7e[data-kb-block=\"kb-adv-heading5654_5c3525-7e\"]{text-align:left;font-style:normal;}.wp-block-kadence-advancedheading.kt-adv-heading5654_5c3525-7e mark.kt-highlight, .wp-block-kadence-advancedheading.kt-adv-heading5654_5c3525-7e[data-kb-block=\"kb-adv-heading5654_5c3525-7e\"] mark.kt-highlight{font-style:normal;color:#f76a0c;-webkit-box-decoration-break:clone;box-decoration-break:clone;padding-top:0px;padding-right:0px;padding-bottom:0px;padding-left:0px;}.wp-block-kadence-advancedheading.kt-adv-heading5654_5c3525-7e img.kb-inline-image, .wp-block-kadence-advancedheading.kt-adv-heading5654_5c3525-7e[data-kb-block=\"kb-adv-heading5654_5c3525-7e\"] img.kb-inline-image{width:150px;vertical-align:baseline;}<\/style>\n<p class=\"kt-adv-heading5654_5c3525-7e wp-block-kadence-advancedheading\" data-kb-block=\"kb-adv-heading5654_5c3525-7e\">Konkrete Szenarien im Business Kontext sind die Erweiterung eines Company Intranets oder einer Knowledge Database um einen Chatbot, mit dem Mitarbeiter*innen einfach interagieren und diesem Fragen stellen k\u00f6nnen. Der RAG-Chatbot erm\u00f6glicht den Mitarbeitenden, schnell auf unternehmensinternes Wissen zuzugreifen. Zus\u00e4tzlich bereitet der Chatbot das gefundene Wissen auf und stellt es zusammengefasst dar. Dies erm\u00f6glicht eine effiziente Nutzung von Unternehmenswissen, steigert die Arbeitsqualit\u00e4t der Mitarbeitenden und kann dazu beitragen, langfristig  einen Wettbewerbsvorteil zu sichern. <\/p>\n\n\n<style>.wp-block-kadence-advancedheading.kt-adv-heading5654_819a2b-03, .wp-block-kadence-advancedheading.kt-adv-heading5654_819a2b-03[data-kb-block=\"kb-adv-heading5654_819a2b-03\"]{text-align:left;font-style:normal;}.wp-block-kadence-advancedheading.kt-adv-heading5654_819a2b-03 mark.kt-highlight, .wp-block-kadence-advancedheading.kt-adv-heading5654_819a2b-03[data-kb-block=\"kb-adv-heading5654_819a2b-03\"] mark.kt-highlight{font-style:normal;color:#f76a0c;-webkit-box-decoration-break:clone;box-decoration-break:clone;padding-top:0px;padding-right:0px;padding-bottom:0px;padding-left:0px;}.wp-block-kadence-advancedheading.kt-adv-heading5654_819a2b-03 img.kb-inline-image, .wp-block-kadence-advancedheading.kt-adv-heading5654_819a2b-03[data-kb-block=\"kb-adv-heading5654_819a2b-03\"] img.kb-inline-image{width:150px;vertical-align:baseline;}<\/style>\n<p class=\"kt-adv-heading5654_819a2b-03 wp-block-kadence-advancedheading\" data-kb-block=\"kb-adv-heading5654_819a2b-03\">Beispielhafte Use-Cases und Produktanwendungen f\u00fcr einen RAG-Chatbot sind:<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>Onboarding neuer Mitarbeiter*innen<\/li>\n\n\n\n<li>T\u00e4gliche Nutzung und Gewinnung von Unternehmensinsights<\/li>\n\n\n\n<li>Chatbot als weiterer Zugang zu digitalen Lerninhalten<\/li>\n<\/ul>\n\n\n<style>.wp-block-kadence-advancedheading.kt-adv-heading5654_5aff31-e8, .wp-block-kadence-advancedheading.kt-adv-heading5654_5aff31-e8[data-kb-block=\"kb-adv-heading5654_5aff31-e8\"]{text-align:left;font-style:normal;}.wp-block-kadence-advancedheading.kt-adv-heading5654_5aff31-e8 mark.kt-highlight, .wp-block-kadence-advancedheading.kt-adv-heading5654_5aff31-e8[data-kb-block=\"kb-adv-heading5654_5aff31-e8\"] mark.kt-highlight{font-style:normal;color:#f76a0c;-webkit-box-decoration-break:clone;box-decoration-break:clone;padding-top:0px;padding-right:0px;padding-bottom:0px;padding-left:0px;}.wp-block-kadence-advancedheading.kt-adv-heading5654_5aff31-e8 img.kb-inline-image, .wp-block-kadence-advancedheading.kt-adv-heading5654_5aff31-e8[data-kb-block=\"kb-adv-heading5654_5aff31-e8\"] img.kb-inline-image{width:150px;vertical-align:baseline;}<\/style>\n<p class=\"kt-adv-heading5654_5aff31-e8 wp-block-kadence-advancedheading\" data-kb-block=\"kb-adv-heading5654_5aff31-e8\">&#8230; oder ihr baut euch euer digtales Alter-Ego \u00e0 la Tom Riddle mit euren Tagebucheintr\u00e4gen . <\/p>\n<\/div><\/div>\n","protected":false},"excerpt":{"rendered":"<p>Seit November 2021 ist das Thema ChatGPT und die dahinterliegende Technologie der&#8230;<\/p>\n","protected":false},"author":2,"featured_media":5982,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"inline_featured_image":false,"_kad_blocks_custom_css":"","_kad_blocks_head_custom_js":"","_kad_blocks_body_custom_js":"","_kad_blocks_footer_custom_js":"","_monsterinsights_skip_tracking":false,"_monsterinsights_sitenote_active":false,"_monsterinsights_sitenote_note":"","_monsterinsights_sitenote_category":0,"_kadence_starter_templates_imported_post":false,"_kad_post_transparent":"","_kad_post_title":"","_kad_post_layout":"","_kad_post_sidebar_id":"","_kad_post_content_style":"","_kad_post_vertical_padding":"","_kad_post_feature":"","_kad_post_feature_position":"","_kad_post_header":false,"_kad_post_footer":false,"_kad_post_classname":"","footnotes":""},"categories":[13],"tags":[23,16,15],"class_list":["post-5654","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-natural-language-processing","tag-ai","tag-data-science","tag-nlp"],"acf":[],"taxonomy_info":{"category":[{"value":13,"label":"Natural Language Processing"}],"post_tag":[{"value":23,"label":"AI"},{"value":16,"label":"Data Science"},{"value":15,"label":"NLP"}]},"featured_image_src_large":["https:\/\/laizee.ai\/wp-content\/uploads\/2024\/04\/RAG-Header-1024x511.png",1024,511,true],"author_info":{"display_name":"Kai Fedin","author_link":"https:\/\/laizee.ai\/author\/kai"},"comment_info":380,"category_info":[{"term_id":13,"name":"Natural Language Processing","slug":"natural-language-processing","term_group":0,"term_taxonomy_id":13,"taxonomy":"category","description":"","parent":0,"count":3,"filter":"raw","cat_ID":13,"category_count":3,"category_description":"","cat_name":"Natural Language Processing","category_nicename":"natural-language-processing","category_parent":0}],"tag_info":[{"term_id":23,"name":"AI","slug":"ai","term_group":0,"term_taxonomy_id":23,"taxonomy":"post_tag","description":"","parent":0,"count":7,"filter":"raw"},{"term_id":16,"name":"Data Science","slug":"data-science","term_group":0,"term_taxonomy_id":16,"taxonomy":"post_tag","description":"","parent":0,"count":2,"filter":"raw"},{"term_id":15,"name":"NLP","slug":"nlp","term_group":0,"term_taxonomy_id":15,"taxonomy":"post_tag","description":"","parent":0,"count":5,"filter":"raw"}],"_links":{"self":[{"href":"https:\/\/laizee.ai\/wp-json\/wp\/v2\/posts\/5654","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/laizee.ai\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/laizee.ai\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/laizee.ai\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/laizee.ai\/wp-json\/wp\/v2\/comments?post=5654"}],"version-history":[{"count":37,"href":"https:\/\/laizee.ai\/wp-json\/wp\/v2\/posts\/5654\/revisions"}],"predecessor-version":[{"id":5988,"href":"https:\/\/laizee.ai\/wp-json\/wp\/v2\/posts\/5654\/revisions\/5988"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/laizee.ai\/wp-json\/wp\/v2\/media\/5982"}],"wp:attachment":[{"href":"https:\/\/laizee.ai\/wp-json\/wp\/v2\/media?parent=5654"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/laizee.ai\/wp-json\/wp\/v2\/categories?post=5654"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/laizee.ai\/wp-json\/wp\/v2\/tags?post=5654"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}