synthetic_rag.de-c5b3de45295e892f.js

js promptingguide.ai collected 2026-09-24 08:29:03 UTC 55,977 bytes, 1 lines download raw bytes

1(self.webpackChunk_N_E=self.webpackChunk_N_E||[]).push([[32200],{66169:function(e,n,s){(window.__NEXT_P=window.__NEXT_P||[]).push(["/applications/synthetic_rag.de",function(){return s(46288)}])},86996:function(e,n){"use strict";n.Z={src:"/_next/static/media/synthetic_rag_1.15948626.png",height:266,width:1579,blurDataURL:"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAgAAAABCAMAAADU3h9xAAAACVBMVEXw8/b6+vnu7u4KD/EJAAAACXBIWXMAAAsTAAALEwEAmpwYAAAAEUlEQVR4nGNgYGBgZGJkYAAAABkABVEQXL0AAAAASUVORK5CYII=",blurWidth:8,blurHeight:1}},74269:function(e,n){"use strict";n.Z={src:"/_next/static/media/synthetic_rag_2.75c57819.png",height:624,width:2310,blurDataURL:"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAgAAAACCAMAAABSSm3fAAAADFBMVEXu7/H8/P309fbz7+qBwNmhAAAACXBIWXMAAAsTAAALEwEAmpwYAAAAF0lEQVR4nGNgZmZgYmBgZACRDEyMjIwAALoAEY7BYLQAAAAASUVORK5CYII=",blurWidth:8,blurHeight:2}},76502:function(e,n){"use strict";n.Z={src:"/_next/static/media/synthetic_rag_3.e13724fc.png",height:828,width:816,blurDataURL:"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAgAAAAICAMAAADz0U65AAAAD1BMVEX+/v73+fnx8fHm5+bp7O87btc4AAAACXBIWXMAAAsTAAALEwEAmpwYAAAAK0lEQVR4nC2KQQ4AIACCEPv/m5utC44JACEQY2tIZkDPCPgXEs/sxe5Xa70IyABM9gG8+gAAAABJRU5ErkJggg==",blurWidth:8,blurHeight:8}},57207:function(e,n){"use strict";n.Z={src:"/_next/static/media/synthetic_rag_4.615b2dcf.png",height:778,width:1844,blurDataURL:"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAgAAAADCAMAAACZFr56AAAACVBMVEXm5ubu7u739/eqJe69AAAACXBIWXMAABYlAAAWJQFJUiTwAAAAG0lEQVR4nDXGsQEAAAjDoND/j3aSiRKsVPaxAwEEABONwN8PAAAAAElFTkSuQmCC",blurWidth:8,blurHeight:3}},46288:function(e,n,s){"use strict";s.r(n),s.d(n,{__toc:function(){return g}});var a=s(52676),i=s(45118),t=s(23742),r=s(40241);s(21729),s(29814);var l=s(41619),d=s(16631),c=s(86996),o=s(74269),m=s(76502),h=s(57207);let g=[{depth:2,value:"Synthetische Daten f\xfcr RAG-Setup",id:"synthetische-daten-f\xfcr-rag-setup"},{depth:2,value:"Dom\xe4nenspezifische Datensatzgenerierung",id:"dom\xe4nenspezifische-datensatzgenerierung"}];function _createMdxContent(e){let n=Object.assign({h1:"h1",h2:"h2",p:"p",a:"a",em:"em",pre:"pre",code:"code",span:"span",math:"math",semantics:"semantics",mrow:"mrow",mo:"mo",msub:"msub",mi:"mi",mn:"mn",annotation:"annotation"},(0,l.a)(),e.components);return(0,a.jsxs)(a.Fragment,{children:[(0,a.jsx)(n.h1,{children:"Generierung eines synthetischen Datensatzes f\xfcr RAG"}),"\n","\n","\n",(0,a.jsx)(n.h2,{id:"synthetische-daten-f\xfcr-rag-setup",children:"Synthetische Daten f\xfcr RAG-Setup"}),"\n",(0,a.jsx)(n.p,{children:"Leider gibt es im Leben eines Machine Learning Ingenieurs oft einen Mangel an gelabelten Daten oder sehr wenige davon. Typischerweise beginnen Projekte, nachdem sie dies bemerkt haben, mit einem langwierigen Prozess der Datensammlung und -kennzeichnung. Erst nach einigen Monaten kann man mit der Entwicklung einer L\xf6sung beginnen."}),"\n",(0,a.jsx)(n.p,{children:"Mit dem Aufkommen der LLMs hat sich das Paradigma bei einigen Produkten jedoch verschoben: Nun kann man sich auf die Generalisierungsf\xe4higkeit von LLMs verlassen und fast sofort eine Idee testen oder ein KI-gesteuertes Feature entwickeln. Wenn sich herausstellt, dass es (beinahe) wie beabsichtigt funktioniert, kann der traditionelle Entwicklungsprozess beginnen."}),"\n",(0,a.jsx)(d.w,{src:c.Z,alt:"Paradigmenwechsel bei KI-gesteuerten Produkten."}),"\n",(0,a.jsxs)(n.p,{children:["Bildquelle: ",(0,a.jsx)(n.a,{href:"https://www.latent.space/p/ai-engineer",children:"The Rise of the AI Engineer, von S. Wang"})]}),"\n",(0,a.jsxs)(n.p,{children:["Einer der aufkommenden Ans\xe4tze ist ",(0,a.jsx)(n.a,{href:"https://www.promptingguide.ai/techniques/rag",children:"Retrieval Augmented Generation (RAG)"}),". Es wird f\xfcr wissensintensive Aufgaben verwendet, bei denen man sich nicht allein auf das Wissen des Modells verlassen kann. RAG kombiniert eine Informationswiederfindungskomponente mit einem Textgenerierungsmodell. Um mehr \xfcber diesen Ansatz zu erfahren, lesen Sie bitte ",(0,a.jsx)(n.a,{href:"https://www.promptingguide.ai/techniques/rag",children:"den entsprechenden Abschnitt im Leitfaden"}),"."]}),"\n",(0,a.jsx)(n.p,{children:"Die Schl\xfcsselkomponente von RAG ist ein Retrieval-Modell, das relevante Dokumente identifiziert und an

1LLMs zur weiteren Verarbeitung weiterleitet. Je besser die Leistung des Retrieval-Modells ist, desto besser ist das Ergebnis des Produkts oder Features. Idealweise funktioniert Retrieval sofort gut. Allerdings sinkt dessen Leistung oft in verschiedenen Sprachen oder spezifischen Dom\xe4nen."}),"\n",(0,a.jsx)(n.p,{children:"Stellen Sie sich vor: Sie m\xfcssen einen Chatbot erstellen, der Fragen basierend auf tschechischen Gesetzen und rechtlichen Praktiken beantwortet (nat\xfcrlich auf Tschechisch). Oder Sie entwerfen einen Steuerassistenten (ein Anwendungsfall, der von OpenAI w\xe4hrend der Pr\xe4sentation von GPT-4 vorgestellt wurde), der f\xfcr den indischen Markt ma\xdfgeschneidert ist. Sie werden wahrscheinlich feststellen, dass das Retrieval-Modell oft nicht die relevantesten Dokumente findet und insgesamt nicht so gut funktioniert, was die Qualit\xe4t des Systems einschr\xe4nkt."}),"\n",(0,a.jsx)(n.p,{children:"Aber es gibt eine L\xf6sung. Ein aufkommender Trend besteht darin, bestehende LLMs zu nutzen, um Daten f\xfcr das Training neuer Generationen von LLMs/Retrievers/anderen Modellen zu synthetisieren. Dieser Prozess kann als Destillieren von LLMs in standardgro\xdfe Encoder \xfcber prompt-basierte Abfragegenerierung betrachtet werden. Obwohl die Destillation rechenintensiv ist, reduziert sie die Inferenzkosten erheblich und k\xf6nnte die Leistung, besonders in spracharmen oder spezialisierten Dom\xe4nen, erheblich steigern."}),"\n",(0,a.jsxs)(n.p,{children:["In diesem Leitfaden verlassen wir uns auf die neuesten Textgenerierungsmodelle, wie ChatGPT und GPT-4, welche gro\xdfe Mengen an synthetischen Inhalten nach Anweisungen produzieren k\xf6nnen. ",(0,a.jsx)(n.a,{href:"https://arxiv.org/abs/2209.11755",children:"Dai et al. (2022)"})," schlugen eine Methode vor, bei der mit nur 8 manuell gelabelten Beispielen und einem gro\xdfen Korpus an ungelabelten Daten (Dokumente f\xfcr das Retrieval, z. B. alle verarbeiteten Gesetze) eine nahezu State-of-the-Art-Leistung erzielt werden kann. Diese Forschung best\xe4tigt, dass synthetisch generierte Daten das Training von aufgabenspezifischen Retrieval-Modellen f\xfcr Aufgaben erleichtern, bei denen supervised in-domain Fine-Tuning eine Herausforderung aufgrund von Datenknappheit ist."]}),"\n",(0,a.jsx)(n.h2,{id:"dom\xe4nenspezifische-datensatzgenerierung",children:"Dom\xe4nenspezifische Datensatzgenerierung"}),"\n",(0,a.jsxs)(n.p,{children:['Um LLMs zu nutzen, muss man eine kurze Beschreibung liefern und einige Beispiele manuell kennzeichnen. Es ist wichtig zu beachten, dass verschiedene Retrieval-Aufgaben unterschiedliche Suchintentionen besitzen, was bedeutet, dass sich die Definition von "Relevanz" unterscheidet. Anders ausgedr\xfcckt, f\xfcr dasselbe Paar (Abfrage, Dokument) k\xf6nnte ihre Relevanz v\xf6llig unterschiedlich sein, basierend auf der Suchintention. Beispielsweise sucht eine Argumentfindungsaufgabe nach unterst\xfctzenden Argumenten, w\xe4hrend andere Aufgaben Gegenargumente erfordern (wie im ',(0,a.jsx)(n.a,{href:"https://aclanthology.org/P18-1023/",children:"ArguAna-Datensatz"})," zu sehen)."]}),"\n",(0,a.jsx)(n.p,{children:"Betrachten Sie das folgende Beispiel. Obwohl es zur leichteren Verst\xe4ndnis auf Englisch geschrieben ist, erinnern Sie sich daran, dass Daten in jeder Sprache sein k\xf6nnen, da ChatGPT/GPT-4 auch weniger verbreitete Sprachen effizient verarbeiten kann."}),"\n",(0,a.jsx)(n.p,{children:(0,a.jsx)(n.em,{children:"Prompt:"})}),"\n",(0,a.jsx)(n.pre,{"data-language":"text","data-theme":"default",children:(0,a.jsxs)(n.code,{"data-language":"text","data-theme":"default",children:[(0,a.jsx)(n.span,{className:"line",children:(0,a.jsx)(n.span,{style:{color:"var(--shiki-color-text)"},children:"Task: Identify a counter-argument for the given argument."})}),"\n",(0,a.jsx)(n.span,{className:"line",children:(0,a.jsx)(n.span,{style:{color:"var(--shiki-color-text)"}})}),"\n",(0,a.jsx)(n.span,{className:"line",children:(0,a.jsx)(n.span,{style:{color:"var(--shiki-color-text)"},children:"Argument #1: {insert passage X1 here}"})}),"\n",(0,a.jsx)(n.span,{className:"line",children:(0,a.jsx)(n.span,{style:{color:"var(--shiki-color-text)"}})}),"\n",(0,a.jsx)(n.span,{className:"line",children:(0,a.jsx)(n.span,{style:{color:"var(--shiki-color-text)"},children:"A concise counter-argument query related to the argument #1: {insert manually prepared query Y1 here}"})}),"\n",(0,a.jsx)(n.span,{className:"line",children:(0,a.jsx)(n.span,{style:{color:"var(--shiki-color-text)"}})}),"\n",(0,a.jsx)(n.span,{className:"line",children:(0,a.jsx)(n.span,{style:{color:"var(--shiki-color-text)"},children:"Argument #2: {insert passage X2 here}"})}),"\n",(0,a.jsx)(n.span,{className:"line",children:(0,a.jsx)(n.span,{style:{color:"var(--shiki-color-text)"},children:"A concise counter-argument query related to the argument #2: {insert manually prepared query Y2 here}"})}),"\n",(0,a.jsx)(n.span,{className:"line",children:(0,a.jsx)(n.span,{style:{color:"var(--shiki-color-text)"}})}),"\n",(0,a.jsx)(n.span,{className:"line",children:(0,a.jsx)(n.span,{style:{color:"var(--shiki-color-text)"},children:"<- paste your examples here ->"})}),"\n",(0,a.jsx)(n.span,{className:"line",children:(0,a.jsx)(n.span,{style:{color:"var(--shiki-color-text)"}})}),"\n",(0,a.jsx)(n.span,{className:"line",children:(0,a.jsx)(n.span,{style:{color:"var(--shiki-color-text)"},children:"Argument N: Even if a fine is made proportional to income, you will not get the equality of impact you desire. This is because the impact is not proportional simply to income, but must take into account a number of other factors. For example, someone supporting a family will face a greater impact than someone who is not, because they have a smaller disposable income. Further, a fine based on income ignores overall wealth (i.e. how much money someone actually has: someone might have a lot of assets but not have"})}),"\n",(0,a.jsx)(n.span,{className:"line",children:(0,a.jsx)(n.span,{style:{color:"var(--shiki-color-text)"},children:"a high income). The proposition does not cater for these inequalities, which may well have a much greater skewing effect, and therefore the argument is being applied inconsistently."})}),"\n",(0,a.jsx)(n.span,{className:"line",children:(0,a.jsx)(n.span,{style:{color:"var(--shiki-color-text)"}})}),"\n",(0,a.jsx)(n.span,{className:"line",children:(0,a.jsx)(n.span,{style:{color:"var(--shiki-color-text)"},children:"A concise counter-argument query related to the argument #N:"})})]})}),"\n",(0,a.jsx)(n.p,{children:(0,a.jsx)(n.em,{children:"Output:"})}),"\n",(0,a.jsx)(n.pre,{"data-language":"text","data-theme":"default",children:(0,a.jsx)(n.code,{"data-language":"text","data-theme":"default",children:(0,a.jsx)(n.span,{className:"line",children:(0,a.jsx)(n.span,{style:{color:"var(--shiki-color-text)"},children:"punishment house would make fines relative income"})})})}),"\n",(0,a.jsx)(n.p,{children:"Generell kann ein solcher Prompt wie folgt ausgedr\xfcckt werden:"}),"\n",(0,a.jsx)(n.p,{children:(0,a.jsxs)(n.span,{className:"katex",children:[(0,a.jsx)(n.span,{className:"katex-mathml",children:(0,a.jsx)(n.math,{xmlns:"http://www.w3.org/1998/Math/MathML",children:(0,a.jsxs)(n.semantics,{children:[(0,a.jsxs)(n.mrow,{children:[(0,a.jsx)(n.mo,{stretchy:"false",children:"("}),(0,a.jsxs)(n.msub,{children:[(0,a.jsx)(n.mi,{children:"e"}),(0,a.jsxs)(n.mrow,{children:[(0,a.jsx)(n.mi,{children:"p"}),(0,a.jsx)(n.mi,{children:"r"}),(0,a.jsx)(n.mi,{children:"o"}),(0,a.jsx)(n.mi,{children:"m"}),(0,a.jsx)(n.mi,{children:"p"}),(0,a.jsx)(n.mi,{children:"t"})]})]}),(0,a.jsx)(n.mo,{separator:"true",children:","}),(0,a.jsxs)(n.msub,{children:[(0,a.jsx)(n.mi,{children:"e"}),(0,a.jsxs)(n.mrow,{children:[(0,a.jsx)(n.mi,{children:"d"}),(0,a.jsx)(n.mi,{children:"o"}),(0,a.jsx)(n.mi,{children:"c"})]})]}),(0,a.jsx)(n.mo,{stretchy:"false",children:"("}),(0,a.jsxs)(n.msub,{children:[(0,a.jsx)(n.mi,{children:"d"}),(0,a.jsx)(n.mn,{children:"1"})]}),(0,a.jsx)(n.mo,{stretchy:"false",children:")"}),(0,a.jsx)(n.mo,{separator:"true",children:","}),(0,a.jsxs)(n.msub,{children:[(0,a.jsx)(n.mi,{children:"e"}),(0,a.jsxs)(n.mrow,{children:[(0,a.jsx)(n.mi,{children:"q"}),(0,a.jsx)(n.mi,{children:"u"}),(0,a.jsx)(n.mi,{children:"e"}),(0,a.jsx)(n.mi,{children:"r"}),(0,a.jsx)(n.mi,{children:"y"})]})]}),(0,a.jsx)(n.mo,{stretchy:"false",children:"("}),(0,a.jsxs)(n.msub,{children:[(0,a.jsx)(n.mi,{children:"q"}),(0,a.jsx)(n.mn,{children:"1"})]}),(0,a.jsx)(n.mo,{stretchy:"false",children:")"}),(0,a.jsx)(n.mo,{separator:"true",children:","}),(0,a.jsx)(n.mi,{mathvariant:"normal",children:"."}),(0,a.jsx)(n.mi,{mathvariant:"normal",children:"."}),(0,a.jsx)(n.mi,{mathvariant:"normal",children:"."}),(0,a.jsx)(n.mo,{separator:"true",children:","}),(0,a.jsxs)(n.msub,{children:[(0,a.jsx)(n.mi,{children:"e"}),(0,a.jsxs)(n.mrow,{children:[(0,a.jsx)(n.mi,{children:"d"}),(0,a.jsx)(n.mi,{children:"o"}),(0,a.jsx)(n.mi,{children:"c"})]})]}),(0,a.jsx)(n.mo,{stretchy:"false",children:"("}),(0,a.jsxs)(n.msub,{children:[(0,a.jsx)(n.mi,{children:"d"}),(0,a.jsx)(n.mi,{children:"k"})]}),(0,a.jsx)(n.mo,{stretchy:"false",children:")"}),(0,a.jsx)(n.mo,{separator:"true",children:","}),(0,a.jsxs)(n.msub,{children:[(0,a.jsx)(n.mi,{children:"e"}),(0,a.jsxs)(n.mrow,{children:[(0,a.jsx)(n.mi,{children:"q"}),(0,a.jsx)(n.mi,{children:"u"}),(0,a.jsx)(n.mi,{children:"e"}),(0,a.jsx)(n.mi,{children:"r"}),(0,a.jsx)(n.mi,{children:"y"})]})]}),(0,a.jsx)(n.mo,{stretchy:"false",children:"("}),(0,a.jsxs)(n.msub,{children:[(0,a.jsx)(n.mi,{children:"q"}),(0,a.jsx)(n.mi,{children:"k"})]}),(0,a.jsx)(n.mo,{stretchy:"false",children:")"}),(0,a.jsx)(n.mo,{separator:"true",children:","}),(0,a.jsxs)(n.msub,{children:[(0,a.jsx)(n.mi,{children:"e"}),(0,a.jsxs)(n.mrow,{children:[(0,a.jsx)(n.mi,{children:"d"}),(0,a.jsx)(n.mi,{children:"o"}),(0,a.jsx)(n.mi,{children:"c"})]})]}),(0,a.jsx)(n.mo,{stretchy:"false",children:"("}),(0,a.jsx)(n.mi,{children:"d"}),(0,a.jsx)(n.mo,{stretchy:"false",children:")"}),(0,a.jsx)(n.mo,{stretchy:"false",children:")"})]}),(0,a.jsx)(n.annotation,{encoding:"application/x-tex",children:"(e_{prompt}, e_{doc}(d_{1}), e_{query}(q_1), . . . , e_{doc}(d_k), e_{query}(q_k), e_{doc}(d))"})]})})}),(0,a.jsx)(n.span,{className:"katex-html","aria-hidden":"true",children:(0,a.jsxs)(n.span,{className:"base",children:[(0,a.jsx)(n.span,{className:"strut",style:{height:"1.0361em",verticalAlign:"-0.2861em"}}),(0,a.jsx)(n.span,{className:"mopen",children:"("}),(0,a.jsxs)(n.span,{className:"mord",children:[(0,a.jsx)(n.span,{className:"mord mathnormal",children:"e"}),(0,a.jsx)(n.span,{className:"msupsub",children:(0,a.jsxs)(n.span,{className:"vlist-t vlist-t2",children:[(0,a.jsxs)(n.span,{className:"vlist-r",children:[(0,a.jsx)(n.span,{className:"vlist",style:{height:"0.2806em"}

1,children:(0,a.jsxs)(n.span,{style:{top:"-2.55em",marginLeft:"0em",marginRight:"0.05em"},children:[(0,a.jsx)(n.span,{className:"pstrut",style:{height:"2.7em"}}),(0,a.jsx)(n.span,{className:"sizing reset-size6 size3 mtight",children:(0,a.jsxs)(n.span,{className:"mord mtight",children:[(0,a.jsx)(n.span,{className:"mord mathnormal mtight",children:"p"}),(0,a.jsx)(n.span,{className:"mord mathnormal mtight",children:"ro"}),(0,a.jsx)(n.span,{className:"mord mathnormal mtight",children:"m"}),(0,a.jsx)(n.span,{className:"mord mathnormal mtight",children:"pt"})]})})]})}),(0,a.jsx)(n.span,{className:"vlist-s",children:"â"})]}),(0,a.jsx)(n.span,{className:"vlist-r",children:(0,a.jsx)(n.span,{className:"vlist",style:{height:"0.2861em"},children:(0,a.jsx)(n.span,{})})})]})})]}),(0,a.jsx)(n.span,{className:"mpunct",children:","}),(0,a.jsx)(n.span,{className:"mspace",style:{marginRight:"0.1667em"}}),(0,a.jsxs)(n.span,{className:"mord",children:[(0,a.jsx)(n.span,{className:"mord mathnormal",children:"e"}),(0,a.jsx)(n.span,{className:"msupsub",children:(0,a.jsxs)(n.span,{className:"vlist-t vlist-t2",children:[(0,a.jsxs)(n.span,{className:"vlist-r",children:[(0,a.jsx)(n.span,{className:"vlist",style:{height:"0.3361em"},children:(0,a.jsxs)(n.span,{style:{top:"-2.55em",marginLeft:"0em",marginRight:"0.05em"},children:[(0,a.jsx)(n.span,{className:"pstrut",style:{height:"2.7em"}}),(0,a.jsx)(n.span,{className:"sizing reset-size6 size3 mtight",children:(0,a.jsxs)(n.span,{className:"mord mtight",children:[(0,a.jsx)(n.span,{className:"mord mathnormal mtight",children:"d"}),(0,a.jsx)(n.span,{className:"mord mathnormal mtight",children:"oc"})]})})]})}),(0,a.jsx)(n.span,{className:"vlist-s",children:"â"})]}),(0,a.jsx)(n.span,{className:"vlist-r",children:(0,a.jsx)(n.span,{className:"vlist",style:{height:"0.15em"},children:(0,a.jsx)(n.span,{})})})]})})]}),(0,a.jsx)(n.span,{className:"mopen",children:"("}),(0,a.jsxs)(n.span,{className:"mord",children:[(0,a.jsx)(n.span,{className:"mord mathnormal",children:"d"}),(0,a.jsx)(n.span,{className:"msupsub",children:(0,a.jsxs)(n.span,{className:"vlist-t vlist-t2",children:[(0,a.jsxs)(n.span,{className:"vlist-r",children:[(0,a.jsx)(n.span,{className:"vlist",style:{height:"0.3011em"},children:(0,a.jsxs)(n.span,{style:{top:"-2.55em",marginLeft:"0em",marginRight:"0.05em"},children:[(0,a.jsx)(n.span,{className:"pstrut",style:{height:"2.7em"}}),(0,a.jsx)(n.span,{className:"sizing reset-size6 size3 mtight",children:(0,a.jsx)(n.span,{className:"mord mtight",children:(0,a.jsx)(n.span,{className:"mord mtight",children:"1"})})})]})}),(0,a.jsx)(n.span,{className:"vlist-s",children:"â"})]}),(0,a.jsx)(n.span,{className:"vlist-r",children:(0,a.jsx)(n.span,{className:"vlist",style:{height:"0.15em"},children:(0,a.jsx)(n.span,{})})})]})})]}),(0,a.jsx)(n.span,{className:"mclose",children:")"}),(0,a.jsx)(n.span,{className:"mpunct",children:","}),(0,a.jsx)(n.span,{className:"mspace",style:{marginRight:"0.1667em"}}),(0,a.jsxs)(n.span,{className:"mord",children:[(0,a.jsx)(n.span,{className:"mord mathnormal",children:"e"}),(0,a.jsx)(n.span,{className:"msupsub",children:(0,a.jsxs)(n.span,{className:"vlist-t vlist-t2",children:[(0,a.jsxs)(n.span,{className:"vlist-r",children:[(0,a.jsx)(n.span,{className:"vlist",style:{height:"0.1514em"},children:(0,a.jsxs)(n.span,{style:{top:"-2.55em",marginLeft:"0em",marginRight:"0.05em"},children:[(0,a.jsx)(n.span,{className:"pstrut",style:{height:"2.7em"}}),(0,a.jsx)(n.span,{className:"sizing reset-size6 size3 mtight",children:(0,a.jsxs)(n.span,{className:"mord mtight",children:[(0,a.jsx)(n.span,{className:"mord mathnormal mtight",style:{marginRight:"0.03588em"},children:"q"}),(0,a.jsx)(n.span,{className:"mord mathnormal mtight",children:"u"}),(0,a.jsx)(n.span,{className:"mord mathnormal mtight",style:{marginRight:"0.03588em"},children:"ery"})]})})]})}),(0,a.jsx)(n.span,{className:"vlist-s",children:"â"})]}),(0,a.jsx)(n.span,{className:"vlist-r",children:(0,a.jsx)(n.span,{className:"vlist",style:{height:"0.2861em"},children:(0,a.jsx)(n.span,{})})})]})})]}),(0,a.jsx)(n.span,{className:"mopen",children:"("}),(0,a.jsxs)(n.span,{className:"mord",children:[(0,a.jsx)(n.span,{className:"mord mathnormal",style:{marginRight:"0.03588em"},children:"q"}),(0,a.jsx)(n.span,{className:"msupsub",children:(0,a.jsxs)(n.span,{className:"vlist-t vlist-t2",children:[(0,a.jsxs)(n.span,{className:"vlist-r",children:[(0,a.jsx)(n.span,{className:"vlist",style:{height:"0.3011em"}

1,children:(0,a.jsxs)(n.span,{style:{top:"-2.55em",marginLeft:"-0.0359em",marginRight:"0.05em"},children:[(0,a.jsx)(n.span,{className:"pstrut",style:{height:"2.7em"}}),(0,a.jsx)(n.span,{className:"sizing reset-size6 size3 mtight",children:(0,a.jsx)(n.span,{className:"mord mtight",children:"1"})})]})}),(0,a.jsx)(n.span,{className:"vlist-s",children:"â"})]}),(0,a.jsx)(n.span,{className:"vlist-r",children:(0,a.jsx)(n.span,{className:"vlist",style:{height:"0.15em"},children:(0,a.jsx)(n.span,{})})})]})})]}),(0,a.jsx)(n.span,{className:"mclose",children:")"}),(0,a.jsx)(n.span,{className:"mpunct",children:","}),(0,a.jsx)(n.span,{className:"mspace",style:{marginRight:"0.1667em"}}),(0,a.jsx)(n.span,{className:"mord",children:"..."}),(0,a.jsx)(n.span,{className:"mpunct",children:","}),(0,a.jsx)(n.span,{className:"mspace",style:{marginRight:"0.1667em"}}),(0,a.jsxs)(n.span,{className:"mord",children:[(0,a.jsx)(n.span,{className:"mord mathnormal",children:"e"}),(0,a.jsx)(n.span,{className:"msupsub",children:(0,a.jsxs)(n.span,{className:"vlist-t vlist-t2",children:[(0,a.jsxs)(n.span,{className:"vlist-r",children:[(0,a.jsx)(n.span,{className:"vlist",style:{height:"0.3361em"},children:(0,a.jsxs)(n.span,{style:{top:"-2.55em",marginLeft:"0em",marginRight:"0.05em"},children:[(0,a.jsx)(n.span,{className:"pstrut",style:{height:"2.7em"}}),(0,a.jsx)(n.span,{className:"sizing reset-size6 size3 mtight",children:(0,a.jsxs)(n.span,{className:"mord mtight",children:[(0,a.jsx)(n.span,{className:"mord mathnormal mtight",children:"d"}),(0,a.jsx)(n.span,{className:"mord mathnormal mtight",children:"oc"})]})})]})}),(0,a.jsx)(n.span,{className:"vlist-s",children:"â"})]}),(0,a.jsx)(n.span,{className:"vlist-r",children:(0,a.jsx)(n.span,{className:"vlist",style:{height:"0.15em"},children:(0,a.jsx)(n.span,{})})})]})})]}),(0,a.jsx)(n.span,{className:"mopen",children:"("}),(0,a.jsxs)(n.span,{className:"mord",children:[(0,a.jsx)(n.span,{className:"mord mathnormal",children:"d"}),(0,a.jsx)(n.span,{className:"msupsub",children:(0,a.jsxs)(n.span,{className:"vlist-t vlist-t2",children:[(0,a.jsxs)(n.span,{className:"vlist-r",children:[(0,a.jsx)(n.span,{className:"vlist",style:{height:"0.3361em"},children:(0,a.jsxs)(n.span,{style:{top:"-2.55em",marginLeft:"0em",marginRight:"0.05em"},children:[(0,a.jsx)(n.span,{className:"pstrut",style:{height:"2.7em"}}),(0,a.jsx)(n.span,{className:"sizing reset-size6 size3 mtight",children:(0,a.jsx)(n.span,{className:"mord mathnormal mtight",style:{marginRight:"0.03148em"},children:"k"})})]})}),(0,a.jsx)(n.span,{className:"vlist-s",children:"â"})]}),(0,a.jsx)(n.span,{className:"vlist-r",children:(0,a.jsx)(n.span,{className:"vlist",style:{height:"0.15em"},children:(0,a.jsx)(n.span,{})})})]})})]}),(0,a.jsx)(n.span,{className:"mclose",children:")"}),(0,a.jsx)(n.span,{className:"mpunct",children:","}),(0,a.jsx)(n.span,{className:"mspace",style:{marginRight:"0.1667em"}}),(0,a.jsxs)(n.span,{className:"mord",children:[(0,a.jsx)(n.span,{className:"mord mathnormal",children:"e"}),(0,a.jsx)(n.span,{className:"msupsub",children:(0,a.jsxs)(n.span,{className:"vlist-t vlist-t2",children:[(0,a.jsxs)(n.span,{className:"vlist-r",children:[(0,a.jsx)(n.span,{className:"vlist",style:{height:"0.1514em"},children:(0,a.jsxs)(n.span,{style:{top:"-2.55em",marginLeft:"0em",marginRight:"0.05em"},children:[(0,a.jsx)(n.span,{className:"pstrut",style:{height:"2.7em"}}),(0,a.jsx)(n.span,{className:"sizing reset-size6 size3 mtight",children:(0,a.jsxs)(n.span,{className:"mord mtight",children:[(0,a.jsx)(n.span,{className:"mord mathnormal mtight",style:{marginRight:"0.03588em"},children:"q"}),(0,a.jsx)(n.span,{className:"mord mathnormal mtight",children:"u"}),(0,a.jsx)(n.span,{className:"mord mathnormal mtight",style:{marginRight:"0.03588em"},children:"ery"})]})})]})}),(0,a.jsx)(n.span,{className:"vlist-s",children:"â"})]}),(0,a.jsx)(n.span,{className:"vlist-r",children:(0,a.jsx)(n.span,{className:"vlist",style:{height:"0.2861em"},children:(0,a.jsx)(n.span,{})})})]})})]}),(0,a.jsx)(n.span,{className:"mopen",children:"("}),(0,a.jsxs)(n.span,{className:"mord",children:[(0,a.jsx)(n.span,{className:"mord mathnormal",style:{marginRight:"0.03588em"},children:"q"}),(0,a.jsx)(n.span,{className:"msupsub",children:(0,a.jsxs)(n.span,{className:"vlist-t vlist-t2",children:[(0,a.jsxs)(n.span,{className:"vlist-r",children:[(0,a.jsx)(n.span,{className:"vlist",style:{height:"0.3361em"}

1,children:(0,a.jsxs)(n.span,{style:{top:"-2.55em",marginLeft:"-0.0359em",marginRight:"0.05em"},children:[(0,a.jsx)(n.span,{className:"pstrut",style:{height:"2.7em"}}),(0,a.jsx)(n.span,{className:"sizing reset-size6 size3 mtight",children:(0,a.jsx)(n.span,{className:"mord mathnormal mtight",style:{marginRight:"0.03148em"},children:"k"})})]})}),(0,a.jsx)(n.span,{className:"vlist-s",children:"â"})]}),(0,a.jsx)(n.span,{className:"vlist-r",children:(0,a.jsx)(n.span,{className:"vlist",style:{height:"0.15em"},children:(0,a.jsx)(n.span,{})})})]})})]}),(0,a.jsx)(n.span,{className:"mclose",children:")"}),(0,a.jsx)(n.span,{className:"mpunct",children:","}),(0,a.jsx)(n.span,{className:"mspace",style:{marginRight:"0.1667em"}}),(0,a.jsxs)(n.span,{className:"mord",children:[(0,a.jsx)(n.span,{className:"mord mathnormal",children:"e"}),(0,a.jsx)(n.span,{className:"msupsub",children:(0,a.jsxs)(n.span,{className:"vlist-t vlist-t2",children:[(0,a.jsxs)(n.span,{className:"vlist-r",children:[(0,a.jsx)(n.span,{className:"vlist",style:{height:"0.3361em"},children:(0,a.jsxs)(n.span,{style:{top:"-2.55em",marginLeft:"0em",marginRight:"0.05em"},children:[(0,a.jsx)(n.span,{className:"pstrut",style:{height:"2.7em"}}),(0,a.jsx)(n.span,{className:"sizing reset-size6 size3 mtight",children:(0,a.jsxs)(n.span,{className:"mord mtight",children:[(0,a.jsx)(n.span,{className:"mord mathnormal mtight",children:"d"}),(0,a.jsx)(n.span,{className:"mord mathnormal mtight",children:"oc"})]})})]})}),(0,a.jsx)(n.span,{className:"vlist-s",children:"â"})]}),(0,a.jsx)(n.span,{className:"vlist-r",children:(0,a.jsx)(n.span,{className:"vlist",style:{height:"0.15em"},children:(0,a.jsx)(n.span,{})})})]})})]}),(0,a.jsx)(n.span,{className:"mopen",children:"("}),(0,a.jsx)(n.span,{className:"mord mathnormal",children:"d"}),(0,a.jsx)(n.span,{className:"mclose",children:"))"})]})})]})}),"\n",(0,a.jsxs)(n.p,{children:[", wobei ",(0,a.jsxs)(n.span,{className:"katex",children:[(0,a.jsx)(n.span,{className:"katex-mathml",children:(0,a.jsx)(n.math,{xmlns:"http://www.w3.org/1998/Math/MathML",children:(0,a.jsxs)(n.semantics,{children:[(0,a.jsx)(n.mrow,{children:(0,a.jsxs)(n.msub,{children:[(0,a.jsx)(n.mi,{children:"e"}),(0,a.jsxs)(n.mrow,{children:[(0,a.jsx)(n.mi,{children:"d"}),(0,a.jsx)(n.mi,{children:"o"}),(0,a.jsx)(n.mi,{children:"c"})]})]})}),(0,a.jsx)(n.annotation,{encoding:"application/x-tex",children:"e_{doc}"})]})})}),(0,a.jsx)(n.span,{className:"katex-html","aria-hidden":"true",children:(0,a.jsxs)(n.span,{className:"base",children:[(0,a.jsx)(n.span,{className:"strut",style:{height:"0.5806em",verticalAlign:"-0.15em"}}),(0,a.jsxs)(n.span,{className:"mord",children:[(0,a.jsx)(n.span,{className:"mord mathnormal",children:"e"}),(0,a.jsx)(n.span,{className:"msupsub",children:(0,a.jsxs)(n.span,{className:"vlist-t vlist-t2",children:[(0,a.jsxs)(n.span,{className:"vlist-r",children:[(0,a.jsx)(n.span,{className:"vlist",style:{height:"0.3361em"},children:(0,a.jsxs)(n.span,{style:{top:"-2.55em",marginLeft:"0em",marginRight:"0.05em"},children:[(0,a.jsx)(n.span,{className:"pstrut",style:{height:"2.7em"}}),(0,a.jsx)(n.span,{className:"sizing reset-size6 size3 mtight",children:(0,a.jsxs)(n.span,{className:"mord mtight",children:[(0,a.jsx)(n.span,{className:"mord mathnormal mtight",children:"d"}),(0,a.jsx)(n.span,{className:"mord mathnormal mtight",children:"oc"})]})})]})}),(0,a.jsx)(n.span,{className:"vlist-s",children:"â"})]}),(0,a.jsx)(n.span,{className:"vlist-r",children:(0,a.jsx)(n.span,{className:"vlist",style:{height:"0.15em"},children:(0,a.jsx)(n.span,{})})})]})})]})]})})]})," und ",(0,a.jsxs)(n.span,{className:"katex",children:[(0,a.jsx)(n.span,{className:"katex-mathml",children:(0,a.jsx)(n.math,{xmlns:"http://www.w3.org/1998/Math/MathML",children:(0,a.jsxs)(n.semantics,{children:[(0,a.jsx)(n.mrow,{children:(0,a.jsxs)(n.msub,{children:[(0,a.jsx)(n.mi,{children:"e"}),(0,a.jsxs)(n.mrow,{children:[(0,a.jsx)(n.mi,{children:"q"}),(0,a.jsx)(n.mi,{children:"u"}),(0,a.jsx)(n.mi,{children:"e"}),(0,a.jsx)(n.mi,{children:"r"}),(0,a.jsx)(n.mi,{children:"y"})]})]})}),(0,a.jsx)(n.annotation,{encoding:"application/x-tex",children:"e_{query}"})]})})}),(0,a.jsx)(n.span,{className:"katex-html","aria-hidden":"true",children:(0,a.jsxs)(n.span,{className:"base",children:[(0,a.jsx)(n.span,{className:"strut",style:{height:"0.7167em",verticalAlign:"-0.2861em"}}),(0,a.jsxs)(n.span,{className:"mord",children:[(0,a.jsx)(n.span,{className:"mord mathnormal",children:"e"}),(0,a.jsx)(n.span,{className:"msupsub",children:(0,a.jsxs)(n.span,{className:"vlist-t vlist-t2",children:[(0,a.jsxs)(n.span,{className:"vlist-r",children:[(0,a.jsx)(n.span,{className:"vlist",style:{height:"0.1514em"}

1,children:(0,a.jsxs)(n.span,{style:{top:"-2.55em",marginLeft:"0em",marginRight:"0.05em"},children:[(0,a.jsx)(n.span,{className:"pstrut",style:{height:"2.7em"}}),(0,a.jsx)(n.span,{className:"sizing reset-size6 size3 mtight",children:(0,a.jsxs)(n.span,{className:"mord mtight",children:[(0,a.jsx)(n.span,{className:"mord mathnormal mtight",style:{marginRight:"0.03588em"},children:"q"}),(0,a.jsx)(n.span,{className:"mord mathnormal mtight",children:"u"}),(0,a.jsx)(n.span,{className:"mord mathnormal mtight",style:{marginRight:"0.03588em"},children:"ery"})]})})]})}),(0,a.jsx)(n.span,{className:"vlist-s",children:"â"})]}),(0,a.jsx)(n.span,{className:"vlist-r",children:(0,a.jsx)(n.span,{className:"vlist",style:{height:"0.2861em"},children:(0,a.jsx)(n.span,{})})})]})})]})]})})]})," jeweils aufgabenspezifische Dokument- und Abfragebeschreibungen sind, ",(0,a.jsxs)(n.span,{className:"katex",children:[(0,a.jsx)(n.span,{className:"katex-mathml",children:(0,a.jsx)(n.math,{xmlns:"http://www.w3.org/1998/Math/MathML",children:(0,a.jsxs)(n.semantics,{children:[(0,a.jsx)(n.mrow,{children:(0,a.jsxs)(n.msub,{children:[(0,a.jsx)(n.mi,{children:"e"}),(0,a.jsxs)(n.mrow,{children:[(0,a.jsx)(n.mi,{children:"p"}),(0,a.jsx)(n.mi,{children:"r"}),(0,a.jsx)(n.mi,{children:"o"}),(0,a.jsx)(n.mi,{children:"m"}),(0,a.jsx)(n.mi,{children:"p"}),(0,a.jsx)(n.mi,{children:"t"})]})]})}),(0,a.jsx)(n.annotation,{encoding:"application/x-tex",children:"e_{prompt}"})]})})}),(0,a.jsx)(n.span,{className:"katex-html","aria-hidden":"true",children:(0,a.jsxs)(n.span,{className:"base",children:[(0,a.jsx)(n.span,{className:"strut",style:{height:"0.7167em",verticalAlign:"-0.2861em"}}),(0,a.jsxs)(n.span,{className:"mord",children:[(0,a.jsx)(n.span,{className:"mord mathnormal",children:"e"}),(0,a.jsx)(n.span,{className:"msupsub",children:(0,a.jsxs)(n.span,{className:"vlist-t vlist-t2",children:[(0,a.jsxs)(n.span,{className:"vlist-r",children:[(0,a.jsx)(n.span,{className:"vlist",style:{height:"0.2806em"},children:(0,a.jsxs)(n.span,{style:{top:"-2.55em",marginLeft:"0em",marginRight:"0.05em"},children:[(0,a.jsx)(n.span,{className:"pstrut",style:{height:"2.7em"}}),(0,a.jsx)(n.span,{className:"sizing reset-size6 size3 mtight",children:(0,a.jsxs)(n.span,{className:"mord mtight",children:[(0,a.jsx)(n.span,{className:"mord mathnormal mtight",children:"p"}),(0,a.jsx)(n.span,{className:"mord mathnormal mtight",children:"ro"}),(0,a.jsx)(n.span,{className:"mord mathnormal mtight",children:"m"}),(0,a.jsx)(n.span,{className:"mord mathnormal mtight",children:"pt"})]})})]})}),(0,a.jsx)(n.span,{className:"vlist-s",children:"â"})]}),(0,a.jsx)(n.span,{className:"vlist-r",children:(0,a.jsx)(n.span,{className:"vlist",style:{height:"0.2861em"},children:(0,a.jsx)(n.span,{})})})]})})]})]})})]})," ist eine aufgabenspezifische Prompt/Anweisung f\xfcr ChatGPT/GPT-4, und ",(0,a.jsxs)(n.span,{className:"katex",children:[(0,a.jsx)(n.span,{className:"katex-mathml",children:(0,a.jsx)(n.math,{xmlns:"http://www.w3.org/1998/Math/MathML",children:(0,a.jsxs)(n.semantics,{children:[(0,a.jsx)(n.mrow,{children:(0,a.jsx)(n.mi,{children:"d"})}),(0,a.jsx)(n.annotation,{encoding:"application/x-tex",children:"d"})]})})}),(0,a.jsx)(n.span,{className:"katex-html","aria-hidden":"true",children:(0,a.jsxs)(n.span,{className:"base",children:[(0,a.jsx)(n.span,{className:"strut",style:{height:"0.6944em"}}),(0,a.jsx)(n.span,{className:"mord mathnormal",children:"d"})]})})]})," ist ein neues Dokument, f\xfcr welches LLM eine Abfrage generieren wird."]}),"\n",(0,a.jsxs)(n.p,{children:["Von diesem Prompt werden nur das letzte Dokument ",(0,a.jsxs)(n.span,{className:"katex",children:[(0,a.jsx)(n.span,{className:"katex-mathml",children:(0,a.jsx)(n.math,{xmlns:"http://www.w3.org/1998/Math/MathML",children:(0,a.jsxs)(n.semantics,{children:[(0,a.jsx)(n.mrow,{children:(0,a.jsx)(n.mi,{children:"d"})}),(0,a.jsx)(n.annotation,{encoding:"application/x-tex",children:"d"})]})})}),(0,a.jsx)(n.span,{className:"katex-html","aria-hidden":"true",children:(0,a.jsxs)(n.span,{className:"base",children:[(0,a.jsx)(n.span,{className:"strut",style:{height:"0.6944em"}}),(0,a.jsx)(n.span,{className:"mord mathnormal",children:"d"})]})})]})," und die generierte Abfrage f\xfcr das weitere Training des lokalen Modells verwendet. Dieser Ansatz kann angewendet werden, wenn ein zielspezifischer Retrieval-Korpus ",(0,a.jsxs)(n.span,{className:"katex",children:[(0,a.jsx)(n.span,{className:"katex-mathml",children:(0,a.jsx)(n.math,{xmlns:"http://www.w3.org/1998/Math/MathML",children:(0,a.jsxs)(n.semantics,{children:[(0,a.jsx)(n.mrow,{children:(0,a.jsx)(n.mi,{children:"D"})}),(0,a.jsx)(n.annotation,{encoding:"application/x-tex",children:"D"})]})})}),(0,a.jsx)(n.span,{className:"katex-html","aria-hidden":"true",children:(0,a.jsxs)(n.span,{className:"base",children:[(0,a.jsx)(n.span,{className:"strut",style:{height:"0.6833em"}}),(0,a.jsx)(n.span,{className:"mord mathnormal",style:{marginRight:"0.02778em"},children:"D"})]})})]})," verf\xfcgbar ist, aber die Anzahl der annotierten Abfrage-Dokument-Paare f\xfcr die neue Aufgabe begrenzt ist."]}),"\n",(0,a.jsx)(n.p,{children:"Der Gesamt\xfcberblick \xfcber die Pipeline:"}),"\n",(0,a.jsx)(d.w,{src:o.Z,alt:"PROMPTGATOR Datensatzgenerierung & Training-\xdcberblick."}),"\n",(0,a.jsxs)(n.p,{children:["Bildquelle: ",(0,a.jsx)(n.a,{href:"https://arxiv.org/abs/2209.11755",children:"Dai et al. (2022)"})]}),"\n",(0,a.jsx)(n.p,{children:"Es ist entscheidend, die manuelle Annotation von Beispielen verantwortungsbewusst zu handhaben. Es ist besser, mehr vorzubereiten (beispielsweise 20) und zuf\xe4llig 2-8 davon zum Prompt hinzuzuf\xfcgen. Dies erh\xf6ht die Vielfalt der generierten Daten ohne signifikante Zeitkosten beim Annotieren. Diese Beispiele sollten allerdings repr\xe4sentativ sein, korrekt formatiert und sogar Details wie die angestrebte Abfragel\xe4nge oder deren Ton spezifizieren. Je pr\xe4ziser die Beispiele und Anweisungen sind, desto besser wird die synthetische Datenqualit\xe4t f\xfcr das Training des Retrievers sein. Beispiele von schlechter Qualit\xe4t k\xf6nnen sich negativ auf die resultierende Qualit\xe4t des trainierten Modells auswirken."}),"\n",(0,a.jsxs)(n.p,{children:["In den meisten F\xe4llen ist die Verwendung eines kosteng\xfcnstigeren Modells wie ChatGPT ausreichend, da es gut mit ungew\xf6hnlichen Dom\xe4nen und Sprachen, die nicht Englisch sind, zurechtkommt. Angenommen, ein Prompt mit Anweisungen und 4-5 Beispielen ben\xf6tigt typischerweise 700 Token (wobei davon ausgegangen wird, dass jeder Abschnitt aufgrund von Retrieval-Einschr\xe4nkungen nicht l\xe4nger als 128 Token ist) und die Generierung ist 25 Token. Somit w\xfcrden die Kosten f\xfcr die Erstellung eines synthetischen Datensatzes f\xfcr ein Korpus von 50.000 Dokumenten f\xfcr das lokale Modell-Fine-Tuning betragen: ",(0,a.jsx)(n.code,{children:"50.000 * (700 * 0.001 * $0.0015 + 25 * 0.001 * $0.002) = 55"}),", wobei ",(0,a.jsx)(n.code,{children:"$0.0015"})," und ",(0,a.jsx)(n.code,{children:"$0.002"})," die Kosten pro 1.000 Token in der GPT-3.5 Turbo-API sind. Es ist sogar m\xf6glich, 2-4 Abfragebeispiele f\xfcr dasselbe Dokument zu generieren. Dennoch sind die Vorteile des weiteren Trainings oft lohnenswert, besonders wenn Sie Retriever nicht f\xfcr eine allgemeine Dom\xe4ne (wie Nachrichtensuche auf Englisch) sondern f\xfcr eine spezifische verwenden (wie tschechische Gesetze, wie erw\xe4hnt)."]}),"\n",(0,a.jsxs)(n.p,{children:["Die Zahl von 50.000 ist nicht willk\xfcrlich. In der Forschung von ",(0,a.jsx)(n.a,{href:"https://arxiv.org/abs/2209.11755",children:"Dai et al. (2022)"})," wird angegeben, dass dies ungef\xe4hr die Anzahl an manuell gelabelten Daten ist, die ein Modell ben\xf6tigt, um die Qualit\xe4t eines auf synthetischen Daten trainierten Modells zu erreichen. Stellen Sie sich vor, Sie m\xfcssten mindestens 10.000 Beispiele sammeln, bevor Sie Ihr Produkt auf den Markt bringen! Das w\xfcrde nicht weniger als einen Monat dauern und die Arbeitskosten w\xfcrden sicherlich tausend Dollar \xfcbersteigen, viel mehr als das Erzeugen von synthetischen Daten und das Training eines lokalen Retriever-Modells. Jetzt k\xf6nnen Sie mit der Technik, die Sie heute gelernt haben, innerhalb weniger Tage ein zweistelliges Wachstum der Metriken erreichen!"]}),"\n",(0,a.jsx)(d.w,{src:m.Z,alt:"Synthetischer Datensatz VS Manuell Gelabelter Datensatz"}),"\n",(0,a.jsxs)(n.p,{children:["Bildquelle: ",(0,a.jsx)(n.a,{href:"https://arxiv.org/abs/2209.11755",children:"Dai et al. (2022)"})]}),"\n",(0,a.jsx)(n.p,{children:"Und hier sind Prompt-Vorlagen aus demselben Papier f\xfcr einige der Datens\xe4tze im BeIR-Benchmark."}),"\n",(0,a.jsx)(d.w,{src:h.Z,alt:"Prompt-Vorlagen aus dem PROMPTGATOR-Artikel."}),"\n",(0,a.jsxs)(n.p,{children:["Bildquelle: ",(0,a.jsx)(n.a,{href:"https://arxiv.org/abs/2209.11755",children:"Dai et al. (2022)"})]})]})}let p={MDXContent:function(){let e=arguments.length>0&&void 0!==arguments[0]?arguments[0]:{},{wrapper:n}=Object.assign({},(0,l.a)(),e.components);return n?(0,a.jsx)(n,{...e,children:(0,a.jsx)(_createMdxContent,{...e})}):_createMdxContent(e)},pageOpts:{filePath:"pages/applications/synthetic_rag.de.mdx",route:"/applications/synthetic_rag",timestamp:1766942504e3,pageMap:[{kind:"Meta",locale:"de",data:{index:"Prompt Engineering",introduction:"Einleitung",techniques:"Techniken",applications:"Anwendungen",prompts:"Prompt Hub",models:"Modelle",risks:"Risiken & Missbrauch",research:"LLM Forschungsergebnisse",papers:"Papers",tools:"Werkzeuge & Bibliotheken",notebooks:"Notebooks",datasets:"Datens\xe4tze",readings:"Zusatzlekt\xfcre",about:{title:"\xdcber",type:"page"},course:"Course",services:"Services"}},{kind:"MdxPage",name:"about",route:"/about",locale:"de"},{kind:"Folder",name:"agents",route:"/agents",children:[{kind:"Meta",locale:"en",data:{introduction:"Introduction to Agents",components:"Agent Components","ai-wor

1kflows-vs-ai-agents":"AI Workflows vs AI Agents","context-engineering":"Context Engineering for AI Agents","context-engineering-deep-dive":"Context Engineering Deep Dive","function-calling":"Function Calling","deep-agents":"Deep Agents"}},{kind:"MdxPage",name:"ai-workflows-vs-ai-agents",route:"/agents/ai-workflows-vs-ai-agents",locale:"en"},{kind:"MdxPage",name:"components",route:"/agents/components",locale:"en"},{kind:"MdxPage",name:"context-engineering-deep-dive",route:"/agents/context-engineering-deep-dive",locale:"en"},{kind:"MdxPage",name:"context-engineering",route:"/agents/context-engineering",locale:"en"},{kind:"MdxPage",name:"deep-agents",route:"/agents/deep-agents",locale:"en"},{kind:"MdxPage",name:"function-calling",route:"/agents/function-calling",locale:"en"},{kind:"MdxPage",name:"introduction",route:"/agents/introduction",locale:"en"}]},{kind:"Folder",name:"applications",route:"/applications",children:[{kind:"Meta",locale:"de",data:{function_calling:"Funktionsaufrufe",generating:"Generierung von Daten",synthetic_rag:"Generierung eines synthetischen Datensatzes f\xfcr RAG",generating_textbooks:"Umgang mit generierten Datens\xe4tzen und deren Vielfalt",coding:"Codegenerierung",workplace_casestudy:"Fallstudie zur Klassifizierung von Absolventenjobs",pf:"Prompt-Funktion"}},{kind:"MdxPage",name:"coding",route:"/applications/coding",locale:"de"},{kind:"MdxPage",name:"function_calling",route:"/applications/function_calling",locale:"de"},{kind:"MdxPage",name:"generating",route:"/applications/generating",locale:"de"},{kind:"MdxPage",name:"generating_textbooks",route:"/applications/generating_textbooks",locale:"de"},{kind:"MdxPage",name:"pf",route:"/applications/pf",locale:"de"},{kind:"MdxPage",name:"synthetic_rag",route:"/applications/synthetic_rag",locale:"de"},{kind:"MdxPage",name:"workplace_casestudy",route:"/applications/workplace_casestudy",locale:"de"},{kind:"MdxPage",name:"context-caching",route:"/applications/context-caching",locale:"en"},{kind:"MdxPage",name:"finetuning-gpt4o",route:"/applications/finetuning-gpt4o",locale:"en"}]},{kind:"MdxPage",name:"applications",route:"/applications",locale:"de"},{kind:"MdxPage",name:"course",route:"/course",locale:"de"},{kind:"MdxPage",name:"datasets",route:"/datasets",locale:"de"},{kind:"Folder",name:"guides",route:"/guides",children:[{kind:"MdxPage",name:"4o-image-generation",route:"/guides/4o-image-generation",locale:"en"},{kind:"Meta",locale:"en",data:{"optimizing-prompts":"Optimizing Prompts","deep-research":"OpenAI Deep Research","reasoning-llms":"Reasoning LLMs","4o-image-generation":"4o Image Generation","context-engineering-guide":"Context Engineering Guide"}},{kind:"MdxPage",name:"context-engineering-guide",route:"/guides/context-engineering-guide",locale:"en"},{kind:"MdxPage",name:"deep-research",route:"/guides/deep-research",locale:"en"},{kind:"MdxPage",name:"optimizing-prompts",route:"/guides/optimizing-prompts",locale:"en"},{kind:"MdxPage",name:"reasoning-llms",route:"/guides/reasoning-llms",locale:"en"}]},{kind:"MdxPage",name:"index",route:"/",locale:"de"},{kind:"Folder",name:"introduction",route:"/introduction",children:[{kind:"Meta",locale:"de",data:{settings:"LLM Einstellungen",basics:"Grundlagen des Promptings",elements:"Elemente eines Prompts",tips:"Allgemeine Tipps f\xfcr das Entwerfen von Prompts",examples:"Beispiel f\xfcr Prompts"}},{kind:"MdxPage",name:"basics",route:"/introduction/basics",locale:"de"},{kind:"MdxPage",name:"elements",route:"/introduction/elements",locale:"de"},{kind:"MdxPage",name:"examples",route:"/introduction/examples",locale:"de"},{kind:"MdxPage",name:"settings",route:"/introduction/settings",locale:"de"},{kind:"MdxPage",name:"tips",route:"/introduction/tips",locale:"de"}]},{kind:"MdxPage",name:"introduction",route:"/introduction",locale:"de"},{kind:"Folder",name:"models",route:"/models",children:[{kind:"Meta",locale:"de",data:{chatgpt:"ChatGPT","claude-3":"Claude 3","code-llama":"Code Llama",flan:"Flan",gemini:"Gemini","gemini-advanced":"Gemini Advanced","gemini-pro":"Gemini 1.5 Pro",gemma:"Gemma","gpt-4":"GPT-4","grok-1":"Grok-1",llama:"LLaMA","llama-3":"Llama 3","mistral-7b":"Mistral 7B","mistral-large":"Mistral Large",mixtral:"Mixtral","mixtral-8x22b":"Mixtral 8x22B",olmo:"OLMo","phi-2":"Phi-2",sora:"Sora",collection:"LLM-Sammlung"}},{kind:"MdxPage",name:"chatgpt",route:"/models/chatgpt",locale:"de"},{kind:"MdxPage",name:"claude-3",route:"/models/claude-3",locale:"de"},{kind:"MdxPage",name:"code-llama",route:"/models/code-llama",locale:"de"},{kind:"MdxPage",name:"collection",route:"/models/collection",locale:"de"},{kind:"MdxPage",name:"flan",route:"/models/flan",locale:"de"},{kind:"MdxPage",name:"gemini-advanced",route:"/models/gemini-advanced",locale:"de"},{kind:"MdxPage",name:"gemini-pro",route:"/models/gemini-pro",locale:"de"},{kind:"MdxPage",name:"gemini",route:"/models/gemini",locale:"de"},{kind:"MdxPage",name:"gemma",route:"/models/gemma",locale:"de"},{kind:"MdxPage",name:"gpt-4",route:"/models/gpt-4",locale:"de"},{kind:"MdxPage",name:"grok-1",route:"/models/grok-1",locale:"de"},{kind:"MdxPage",name:"llama-3",route:"/models/llama-3",locale:"de"},{kind:"MdxPage",name:"llama",route:"/models/llama",locale:"de"},{kind:"MdxPage",name:"mistral-7b",route:"/models/mistral-7b",locale:"de"},{kind:"MdxPage",name:"mistral-large",route:"/models/mistral-large",locale:"de"},{kind:"MdxPage",name:"mixtral-8x22b",route:"/models/mixtral-8x22b",locale:"de"},{kind:"MdxPage",name:"mixtral",route:"/models/mixtral",locale:"de"},{kind:"MdxPage",name:"olmo",route:"/models/olmo",locale:"de"},{kind:"MdxPage",name:"phi-2",route:"/models/phi-2",locale:"de"},{kind:"MdxPage",name:"sora",route:"/models/sora",locale:"de"},{kind:"MdxPage",name:"kimi-k2.5",route:"/models/kimi-k2.5",locale:"en"}]},{kind:"MdxPage",name:"models",route:"/models",locale:"de"},{kind:"MdxPage",name:"notebooks",route:"/notebooks",locale:"de"},{kind:"MdxPage",name:"papers",route:"/papers",locale:"de"},{kind:"Folder",name:"prompts",route:"/prompts",children:[{kind:"Meta",locale:"de",data:{classification:"Klassifizierung",coding:"Coding",creativity:"Kreativit\xe4t",evaluation:"Evaluation","information-extraction":"Informationsextraktion","image-generation":"Bildgenerierung",mathematics:"Mathematik  ","question-answering":"Fragebeantwortung",reasoning:"Schlussfolgerungen","text-summarization":"Textzusammenfassung",truthfulness:"Wahrhaftigkeit","adversarial-prompting":"Adversariales Prompting"}},{kind:"Folder",name:"adversarial-prompting",route:"/prompts/adversarial-prompting",children:[{kind:"Meta",locale:"de",data:{"prompt-injection":"Prompt Injection","prompt-leaking":"Prompt Leaking","jailbreaking-llms":"Jailbreaking"}},{kind:"MdxPage",name:"jailbreaking-llms",route:"/prompts/adversarial-prompting/jailbreaking-llms",locale:"de"},{kind:"MdxPage",name:"prompt-injection",route:"/prompts/adversarial-prompting/prompt-injection",locale:"de"},{kind:"MdxPage",name:"prompt-leaking",route:"/prompts/adversarial-prompting/prompt-leaking",locale:"de"}]},{kind:"MdxPage",name:"adversarial-prompting",route:"/prompts/adversarial-prompting",locale:"de"},{kind:"Folder",name:"classification",route:"/prompts/classification",children:[{kind:"Meta",locale:"de",data:{sentiment:"Sentimentklassifikation","sentiment-fewsh

1ot":"Few-Shot Sentimentklassifikation"}},{kind:"MdxPage",name:"sentiment-fewshot",route:"/prompts/classification/sentiment-fewshot",locale:"de"},{kind:"MdxPage",name:"sentiment",route:"/prompts/classification/sentiment",locale:"de"}]},{kind:"MdxPage",name:"classification",route:"/prompts/classification",locale:"de"},{kind:"Folder",name:"coding",route:"/prompts/coding",children:[{kind:"Meta",locale:"de",data:{"code-snippet":"Code-Snippets generieren","mysql-query":"Erzeugen von MySQL-Queries",tikz:"TiKZ-Diagramm zeichnen"}},{kind:"MdxPage",name:"code-snippet",route:"/prompts/coding/code-snippet",locale:"de"},{kind:"MdxPage",name:"mysql-query",route:"/prompts/coding/mysql-query",locale:"de"},{kind:"MdxPage",name:"tikz",route:"/prompts/coding/tikz",locale:"de"}]},{kind:"MdxPage",name:"coding",route:"/prompts/coding",locale:"de"},{kind:"Folder",name:"creativity",route:"/prompts/creativity",children:[{kind:"Meta",locale:"de",data:{rhymes:"Reime","infinite-primes":"Unendlichkeit der Primzahlen",interdisciplinary:"Interdisziplin\xe4re Aufgaben","new-words":"Erfindung neuer W\xf6rter"}},{kind:"MdxPage",name:"infinite-primes",route:"/prompts/creativity/infinite-primes",locale:"de"},{kind:"MdxPage",name:"interdisciplinary",route:"/prompts/creativity/interdisciplinary",locale:"de"},{kind:"MdxPage",name:"new-words",route:"/prompts/creativity/new-words",locale:"de"},{kind:"MdxPage",name:"rhymes",route:"/prompts/creativity/rhymes",locale:"de"}]},{kind:"MdxPage",name:"creativity",route:"/prompts/creativity",locale:"de"},{kind:"Folder",name:"evaluation",route:"/prompts/evaluation",children:[{kind:"Meta",locale:"de",data:{"plato-dialogue":"Platons Dialog bewerten"}},{kind:"MdxPage",name:"plato-dialogue",route:"/prompts/evaluation/plato-dialogue",locale:"de"}]},{kind:"MdxPage",name:"evaluation",route:"/prompts/evaluation",locale:"de"},{kind:"Folder",name:"image-generation",route:"/prompts/image-generation",children:[{kind:"Meta",locale:"de",data:{"alphabet-person":"Eine Person mit Alphabet-Buchstaben zeichnen"}},{kind:"MdxPage",name:"alphabet-person",route:"/prompts/image-generation/alphabet-person",locale:"de"}]},{kind:"MdxPage",name:"image-generation",route:"/prompts/image-generation",locale:"de"},{kind:"Folder",name:"information-extraction",route:"/prompts/information-extraction",children:[{kind:"Meta",locale:"de",data:{"extract-models":"Modellnamen extrahieren"}},{kind:"MdxPage",name:"extract-models",route:"/prompts/information-extraction/extract-models",locale:"de"}]},{kind:"MdxPage",name:"information-extraction",route:"/prompts/information-extraction",locale:"de"},{kind:"Folder",name:"mathematics",route:"/prompts/mathematics",children:[{kind:"Meta",locale:"de",data:{"composite-functions":"Auswertung zusammengesetzter Funktionen","odd-numbers":"Ungerade Zahlen addieren"}},{kind:"MdxPage",name:"composite-functions",route:"/prompts/mathematics/composite-functions",locale:"de"},{kind:"MdxPage",name:"odd-numbers",route:"/prompts/mathematics/odd-numbers",locale:"de"}]},{kind:"MdxPage",name:"mathematics",route:"/prompts/mathematics",locale:"de"},{kind:"Folder",name:"question-answering",route:"/prompts/question-answering",children:[{kind:"Meta",locale:"de",data:{"closed-domain":"Geschlossene Dom\xe4nen-Fragenbeantwortung","open-domain":"Offene Dom\xe4nen-Fragenbeantwortung","science-qa":"Wissenschaftliches Frage-Antworten"}},{kind:"MdxPage",name:"closed-domain",route:"/prompts/question-answering/closed-domain",locale:"de"},{kind:"MdxPage",name:"open-domain",route:"/prompts/question-answering/open-domain",locale:"de"},{kind:"MdxPage",name:"science-qa",route:"/prompts/question-answering/science-qa",locale:"de"}]},{kind:"MdxPage",name:"question-answering",route:"/prompts/question-answering",locale:"de"},{kind:"Folder",name:"reasoning",route:"/prompts/reasoning",children:[{kind:"Meta",locale:"de",data:{"indirect-reasoning":"Indirektes Reasoning","physical-reasoning":"Physisches Reasoning"}},{kind:"MdxPage",name:"indirect-reasoning",route:"/prompts/reasoning/indirect-reasoning",locale:"de"},{kind:"MdxPage",name:"physical-reasoning",route:"/prompts/reasoning/physical-reasoning",locale:"de"}]},{kind:"MdxPage",name:"reasoning",route:"/prompts/reasoning",locale:"de"},{kind:"Folder",name:"text-summarization",route:"/prompts/text-summarization",children:[{kind:"Meta",locale:"de",data:{"explain-concept":"Konzepte erkl\xe4ren"}},{kind:"MdxPage",name:"explain-concept",route:"/prompts/text-summarization/explain-concept",locale:"de"}]},{kind:"MdxPage",name:"text-summarization",route:"/prompts/text-summarization",locale:"de"},{kind:"Folder",name:"truthfulness",route:"/prompts/truthfulness",children:[{kind:"Meta",locale:"de",data:{"identify-hallucination":"Identifizieren von Halluzination"}},{kind:"MdxPage",name:"identify-hallucination",route:"/prompts/truthfulness/identify-hallucination",locale:"de"}]},{kind:"MdxPage",name:"truthfulness",route:"/prompts/truthfulness",locale:"de"}]},{kind:"MdxPage",name:"prompts",route:"/prompts",locale:"de"},{kind:"MdxPage",name:"readings",route:"/readings",locale:"de"},{kind:"Folder",name:"research",route:"/research",children:[{kind:"Meta",locale:"de",data:{"llm-agents":"LLM Agenten",rag:"RAG f\xfcr LLMs","llm-reasoning":"LLM Reasoning","guided-cot":"LM-gef\xfchrtes CoT",rag_hallucinations:"RAG Reduziert Halluzination",synthetic_data:"Synthetische Daten","rag-faithfulness":"RAG Zuverl\xe4ssigkeit","llm-recall":"LLM In-Context Recall",thoughtsculpt:"ThoughtSculpt","infini-attention":"Infini-Attention","trustworthiness-in-llms":"Vertrauensw\xfcrdigkeit in LLMs","llm-tokenization":"LLM Tokenisierung",groq:"Was ist Groq?"}},{kind:"MdxPage",name:"groq",route:"/research/groq",locale:"de"},{kind:"MdxPage",name:"guided-cot",route:"/research/guided-cot",locale:"de"},{kind:"MdxPage",name:"infini-attention",route:"/research/infini-attention",locale:"de"},{kind:"MdxPage",name:"llm-agents",route:"/research/llm-agents",locale:"de"},{kind:"MdxPage",name:"llm-reasoning",route:"/research/llm-reasoning",locale:"de"},{kind:"MdxPage",name:"llm-recall",route:"/research/llm-recall",locale:"de"},{kind:"MdxPage",name:"llm-tokenization",route:"/research/llm-tokenization",locale:"de"},{kind:"MdxPage",name:"rag-faithfulness",route:"/research/rag-faithfulness",locale:"de"},{kind:"MdxPage",name:"rag",route:"/research/rag",locale:"de"},{kind:"MdxPage",name:"rag_hallucinations",route:"/research/rag_hallucinations",locale:"de"},{kind:"MdxPage",name:"synthetic_data",route:"/research/synthetic_data",locale:"de"},{kind:"MdxPage",name:"thoughtsculpt",route:"/research/thoughtsculpt",locale:"de"},{kind:"MdxPage",name:"trustworthiness-in-llms",route:"/research/trustworthiness-in-llms",locale:"de"}]},{kind:"MdxPage",name:"research",route:"/research",locale:"de"},{kind:"Folder",name:"risks",route:"/risks",children:[{kind:"Meta",locale:"de",data:{adversarial:"Adversariales Prompting",factuality:"Faktentreue",biases:"Verzerrungen (biases)"}},{kind:"MdxPage",name:"adversarial",route:"/risks/adversarial",locale:"de"},{kind:"MdxPage",name:"biases",route:"/risks/biases",locale:"de"},{kind:"MdxPage",name:"factuality",route:"/risks/factuality",locale:"de"}]},{kind:"MdxPage",name:"risks",route:"/risks",locale:"de"},{kind:"MdxPage",name:"services",route:"/services",locale:"de"},{kind:"Folder",name:"techniques",route:"/techniques",children:[{kind:"Meta",locale:"de",data:{zeroshot:"Zero-Shot Prompting",fewshot:"Few-Shot Prompting",cot:"Chain-of-Thought Prompting",consistency:"Selbstkonsistenz",knowledge:"Generiertes Wissens-Prompting",prompt_chaining:"Prompt Chaining",tot:"Tree of Thoughts",rag:"Retrieval Augmented Generation",art:"Automatic Reasoning and Tool-use",ape:"Automatic Prompt Engineer",activeprompt:"Active-Prompt",dsp:"Directional Stimulus Prompting",pal:"Program-Aided Language Models",react:"ReAct",reflexion:"Reflexion",multimodalcot:"Multimodal CoT",graph:"Graph-Prompting"}},{kind:"MdxPage",name:"activeprompt",route:"/techniques/activeprompt",locale:"de"},{kind:"MdxPage",name:"ape",route:"/techniques/ape",locale:"de"},{kind:"MdxPage",name:"art",route:"/techniques/art",locale:"de"},{kind:"MdxPage",name:"consistency",route:"/techniques/consistency",locale:"de"},{kind:"MdxPage",name:"cot",route:"/techniques/cot",locale:"de"},{kind:"MdxPage",name:"dsp",route:"/techniques/dsp",locale:"de"},{kind:"MdxPage",name:"fewshot",route:"/techniques/fewshot",locale:"de"},{kind:"MdxPage",name:"graph",route:"/techniques/graph",locale:"de"},{kind:"MdxPage",name:"knowledge",route:"/techniques/knowledge",locale:"de"},{kind:"MdxPage",name:"multimodalcot",route:"/techniques/multimodalcot",locale:"de"},{kind:"MdxPage",name:"pal",route:"/techniques/pal",locale:"de"},{kind:"MdxPage",name:"prompt_chaining",route:"/techniques/prompt_chaining",locale:"de"},{kind:"MdxPage",name:"rag",route:"/techniques/rag",locale:"de"},{kind:"MdxPage",name:"react",route:"/techniques/react",locale:"de"},{kind:"MdxPage",name:"reflexion",route:"/techniques/reflexion",locale:"de"},{kind:"MdxPage",name:"tot",route:"/techniques/tot",locale:"de"},{kind:"MdxPage",name:"zeroshot",route:"/techniques/zeroshot",locale:"de"},{kind:"MdxPage",name:"meta-prompting",route:"/techniques/meta-prompting",locale:"en"}]},{kind:"MdxPage",name:"techniques",route:"/techniques",locale:"de"},{kind:"MdxPage",name:"tools",route:"/tools",locale:"de"},{kind:"MdxPage",name:"agents",route:"/agents",locale:"en"},{kind:"MdxPage",name:"courses",route:"/courses",locale:"en"}],flexsearch:{codeblocks:!0},title:"Generierung eines synthetischen Datensatzes f\xfcr RAG",headings:g},pageNextRoute:"/applications/synthetic_rag.de",nextraLayout:t.ZP,themeConfig:r.Z};n.default=(0,i.j)(p)},16631:function(e,n,s){"use strict";s.d(n,{w:function(){return Screenshot}});var a=s(52676),i=s(4814),t=s(408),r=s.n(t);function Screenshot(e){let{src:n,alt:s,full:t}=e;return(0,a.jsx)("div",{className:(0,i.Z)("mt-6 -mb-4 flex justify-center overflow-hidden rounded-xl border dark:border-zinc-800",t?"bg-white":"bg-zinc-100"),children:(0,a.jsx)(r(),{src:n,alt:s,className:(0,i.Z)("w-auto select-none bg-white",t?"":"ring-1 ring-gray-200")})})}}},function(e){e.O(0,[57914,26494,49774,92888,40179],function(){return e(e.s=66169)}),_N_E=e.O()}]);

Line numbers count LF bytes from the start of the resource, as the search results do. Vendor segments are library code the classifier recognised; they are stored but not indexed. Bytes are shown as Latin1 characters, one per byte.