promptfoo 0.79.0 → 0.80.1

package/dist/src/web/nextui/_next/static/chunks/258-1e2bd16e6c0b6b4a.js

@@ -1 +0,0 @@
-(self.webpackChunk_N_E=self.webpackChunk_N_E||[]).push([[258],{49532:function(e,t,n){"use strict";let s,l;async function a(){if(s||(l||(l=fetch("/api/config").then(e=>e.json()).then(e=>s=e.apiBaseUrl)),await l),void 0===s)throw Error("API base URL is undefined");return s}n.d(t,{b:function(){return a}})},55258:function(e,t,n){"use strict";n.r(t),n.d(t,{default:function(){return to}});var s=n(24004),l=n(14978),a=n(77580);let r=a.env.NEXT_PUBLIC_PROMPTFOO_REMOTE_API_BASE_URL||a.env.NEXT_PUBLIC_PROMPTFOO_BASE_URL||a.env.PROMPTFOO_REMOTE_API_BASE_URL||"https://api.promptfoo.dev",i=a.env.NEXT_PUBLIC_PROMPTFOO_BASE_URL||a.env.PROMPTFOO_REMOTE_APP_BASE_URL||"https://app.promptfoo.dev";a?.stdout?.columns&&a?.stdout?.columns>10&&a?.stdout?.columns;var o=n(49532);let c=(0,l.createContext)(void 0),d=e=>{let{children:t}=e,[n,a]=(0,l.useState)(!1);return(0,l.useEffect)(()=>{let e=e=>{"Shift"===e.key&&a(!0)},t=e=>{"Shift"===e.key&&a(!1)};return window.addEventListener("keydown",e),window.addEventListener("keyup",t),()=>{window.removeEventListener("keydown",e),window.removeEventListener("keyup",t)}},[]),(0,s.jsx)(c.Provider,{value:n,children:t})};var u=n(18960),h=n(67070);let p=(0,l.createContext)(void 0),m=()=>{let e=(0,l.useContext)(p);if(!e)throw Error("useToast must be used within a ToastProvider");return e},x=e=>{let{children:t}=e,[n,a]=(0,l.useState)({message:"",severity:"info",open:!1}),r=()=>{a(e=>({...e,open:!1}))};return(0,s.jsxs)(p.Provider,{value:{showToast:function(e){let t=arguments.length>1&&void 0!==arguments[1]?arguments[1]:"info";a({message:e,severity:t,open:!0})}},children:[t,(0,s.jsx)(u.Z,{open:n.open,autoHideDuration:6e3,onClose:r,children:(0,s.jsx)(h.Z,{onClose:r,severity:n.severity,children:n.message})})]})};var v=n(52428),g=n(6882),j=n(12594),f=n(28891),b=n(10804),y=n(93179),Z=n(47887),w=n(11116),C=n(46446),S=n(13391),k=n(98507),E=n(93295),R=n(72467),I=n(95198),N=n(49605),P=n(79245),F=n(49050),T=n(27760),D=n(47827),M=n(84081),O=n(81679),L=n(50819),A=n(8864),U=n(38212),W=n(33932),z=n(39114),V=n(90923),_=n(10654),B=n(13457),J=n(21975),H=n(78276),K=n(72303),X=n(72053),G=n(94106),Y=n(89394),q=n(42834),$=n(26337),Q=n(91797),ee=n(29872),et=n(73701),en=n(39279),es=n(30666),el=n(15795),ea=n(66988),er=n(98489),ei=n(43226),eo=n(32162),ec=n.n(eo),ed=e=>{let{open:t,onClose:n,recentEvals:a,onRecentEvalSelected:r,title:i,description:o}=e,[c,d]=(0,l.useState)(""),[u,h]=(0,l.useState)(-1),p=l.useRef(null),m=l.useRef(null),x=()=>{n(),d(""),h(-1)},v=a.filter(e=>ec()(c.toLowerCase(),e.label.toLowerCase())||"string"==typeof e.description&&ec()(c.toLowerCase(),e.description.toLowerCase())),g=e=>{r(e),x()},j=l.useCallback(()=>{if(u>=0&&m.current){let e=m.current.querySelectorAll("tbody tr"),t=Math.min(u+3,e.length-1);e[t]&&e[t].scrollIntoView({behavior:"smooth",block:"nearest"})}},[u]);l.useEffect(()=>{j()},[j]),l.useEffect(()=>{t&&(h(0),setTimeout(()=>{var e;null===(e=p.current)||void 0===e||e.focus()},0))},[t]);let f=l.useId();return(0,s.jsxs)(Y.Z,{open:t,onClose:x,maxWidth:"md",fullWidth:!0,children:[i?(0,s.jsx)(Q.Z,{children:i}):null,(0,s.jsxs)($.Z,{children:[o?(0,s.jsx)(P.Z,{sx:{mb:4},children:o}):null,(0,s.jsxs)(P.Z,{sx:{width:"100%",mt:2},children:[(0,s.jsx)(J.Z,{fullWidth:!0,variant:"outlined",placeholder:"Search",value:c,onChange:e=>{d(e.target.value),h(0)},onKeyDown:e=>{if(e.stopPropagation(),t)switch(e.key){case"ArrowDown":e.preventDefault(),h(e=>Math.min(e+1,v.length-1));break;case"ArrowUp":e.preventDefault(),h(e=>Math.max(e-1,0));break;case"Enter":e.preventDefault(),u>=0&&u<v.length?g(v[u].evalId):v.length>0&&g(v[0].evalId);break;case"Escape":e.preventDefault(),x()}},sx:{mb:2},inputRef:p,id:"eval-selector-search-".concat(f)}),(0,s.jsx)(el.Z,{component:ee.Z,sx:{height:"600px",overflow:"auto"},ref:m,children:(0,s.jsxs)(et.Z,{stickyHeader:!0,children:[(0,s.jsx)(ea.Z,{children:(0,s.jsxs)(er.Z,{children:[(0,s.jsx)(es.Z,{children:"Created"}),(0,s.jsx)(es.Z,{children:"Description"}),(0,s.jsx)(es.Z,{children:"# Tests"})]})}),(0,s.jsx)(en.Z,{children:v.length>0?v.map((e,t)=>(0,s.jsxs)(er.Z,{hover:!0,onClick:()=>g(e.evalId),sx:{cursor:"pointer",backgroundColor:t===u?"rgba(255, 255, 0, 0.1)":"inherit"},children:[(0,s.jsx)(es.Z,{children:new Date(e.createdAt).toLocaleString()}),(0,s.jsx)(es.Z,{children:e.description||e.label}),(0,s.jsx)(es.Z,{children:e.numTests})]},e.evalId)):(0,s.jsx)(er.Z,{children:(0,s.jsx)(es.Z,{colSpan:3,align:"center",sx:{py:4},children:(0,s.jsxs)(P.Z,{sx:{textAlign:"center",color:"text.secondary"},children:[(0,s.jsx)(P.Z,{sx:{fontSize:"3rem",mb:2},children:"\uD83D\uDD0D"}),(0,s.jsx)(ei.Z,{variant:"h6",gutterBottom:!0,children:"No evaluations found"}),(0,s.jsx)(ei.Z,{variant:"body2",children:"Try adjusting your search or create a new evaluation"})]})})})})]})})]})]}),(0,s.jsx)(q.Z,{children:(0,s.jsx)(F.Z,{onClick:x,children:"Cancel"})})]})},eu=n(37204),eh=n(94660),ep=n(74810);let em={getItem:async e=>await (0,eu.U2)(e)||null,setItem:async(e,t)=>{await (0,eu.t8)(e,t)},removeItem:async e=>{await (0,eu.IV)(e)}},ex=(0,eh.Ue)()((0,ep.tJ)((e,t)=>({evalId:null,setEvalId:t=>e(()=>({evalId:t})),author:null,setAuthor:t=>e(()=>({author:t})),table:null,setTable:t=>e(()=>({table:t})),config:null,setConfig:t=>e(()=>({config:t})),maxTextLength:250,setMaxTextLength:t=>e(()=>({maxTextLength:t})),wordBreak:"break-word",setWordBreak:t=>e(()=>({wordBreak:t})),showInferenceDetails:!0,setShowInferenceDetails:t=>e(()=>({showInferenceDetails:t})),renderMarkdown:!1,setRenderMarkdown:t=>e(()=>({renderMarkdown:t})),prettifyJson:!1,setPrettifyJson:t=>e(()=>({prettifyJson:t})),showPrompts:!1,setShowPrompts:t=>e(()=>({showPrompts:t})),showPassFail:!0,setShowPassFail:t=>e(()=>({showPassFail:t})),inComparisonMode:!1,setInComparisonMode:t=>e(()=>({inComparisonMode:t})),columnStates:{},setColumnState:(t,n)=>e(e=>({columnStates:{...e.columnStates,[t]:n}}))}),{name:"ResultsViewStorage",storage:(0,ep.FL)(()=>em)}));var ev=function(e){let{initialEvals:t,onComparisonEvalSelected:n}=e,{evalId:a}=ex(),[r,i]=(0,l.useState)(!1),[c,d]=(0,l.useState)(t),u=async()=>{try{let e=await (0,o.b)(),n=a||t[0].evalId,s=await fetch("".concat(e,"/api/results/").concat(n),{cache:"no-store"}),l=await s.json(),r=l.data.datasetId;if(!r){console.error("No datasetId found for current eval "+n);return}let i=await fetch("".concat(e,"/api/results?datasetId=").concat(r),{cache:"no-store"}),c=await i.json();d(c.data.filter(e=>e.evalId!==n))}catch(e){console.error("Error fetching recent evals:",e)}},h=()=>{i(!1)};return(0,s.jsxs)(s.Fragment,{children:[(0,s.jsx)(H.Z,{title:"Combine this eval with another eval run",placement:"left",children:(0,s.jsxs)(z.Z,{onClick:()=>{u(),i(!0)},children:[(0,s.jsx)(A.Z,{children:(0,s.jsx)(G.Z,{fontSize:"small"})}),(0,s.jsx)(U.Z,{children:"Compare with another eval"})]})}),(0,s.jsx)(ed,{open:r,onClose:h,recentEvals:c,onRecentEvalSelected:e=>{n(e),h()},title:"Select an eval to compare",description:"Only evals with the same dataset can be compared."})]})},eg=n(54719),ej=n(52481),ef=n(7175),eb=n(52653);function ey(e){let{open:t,onClose:a}=e,{config:r}=ex(),i=l.useRef(null),[o,c]=l.useState(!1),[d,u]=l.useState("");l.useEffect(()=>{t&&(async()=>{let{default:e}=await Promise.resolve().then(n.bind(n,34235));u(e.dump(r))})()},[t,r]);let h=()=>{c(!1),a()};return(0,s.jsxs)(Y.Z,{open:t,onClose:h,"aria-labelledby":"config-dialog-title",maxWidth:"md",fullWidth:!0,children:[(0,s.jsx)(Q.Z,{id:"config-dialog-title",children:(0,s.jsxs)(P.Z,{display:"flex",justifyContent:"space-between",alignItems:"center",children:[(0,s.jsx)(ei.Z,{variant:"h6",style:{flexGrow:1},children:"Config"}),(0,s.jsxs)(P.Z,{children:[(0,s.jsx)(H.Z,{title:"Copy to clipboard",children:(0,s.jsx)(eb.Z,{onClick:()=>{i.current&&(i.current.select(),document.execCommand("copy"),c(!0))},children:o?(0,s.jsx)(eg.Z,{}):(0,s.jsx)(ef.Z,{})})}),(0,s.jsx)(H.Z,{title:"Download .yaml",children:(0,s.jsx)(eb.Z,{onClick:()=>{let e=new Blob([d],{type:"text/yaml;charset=utf-8"}),t=URL.createObjectURL(e),n=document.createElement("a");n.href=t,n.download="config.yaml",document.body.appendChild(n),n.click(),document.body.removeChild(n),URL.revokeObjectURL(t)},children:(0,s.jsx)(ej.Z,{})})})]})]})}),(0,s.jsx)($.Z,{children:(0,s.jsx)(ei.Z,{variant:"body1",component:"div",children:(0,s.jsx)("textarea",{ref:i,readOnly:!0,value:d,style:{width:"100%",minHeight:"400px",fontFamily:"monospace",border:"1px solid #ccc"}})})}),(0,s.jsx)(q.Z,{children:(0,s.jsx)(F.Z,{onClick:h,color:"primary",children:"Close"})})]})}var eZ=n(79715),ew=n(34235),eC=function(){let{table:e,config:t,evalId:n}=ex(),[a,r]=l.useState(!1),i=(e,t)=>{let n=URL.createObjectURL(e),s=document.createElement("a");s.href=n,s.download=t,document.body.appendChild(s),s.click(),document.body.removeChild(s),URL.revokeObjectURL(n)},o=()=>{r(!1)};return(0,s.jsxs)(s.Fragment,{children:[(0,s.jsxs)(z.Z,{onClick:()=>{r(!0)},children:[(0,s.jsx)(A.Z,{children:(0,s.jsx)(ej.Z,{fontSize:"small"})}),(0,s.jsx)(U.Z,{children:"Download"})]}),(0,s.jsx)(Y.Z,{onClose:o,open:a,children:(0,s.jsx)($.Z,{children:(0,s.jsxs)(B.Z,{direction:"column",spacing:2,sx:{width:"100%"},children:[(0,s.jsx)(F.Z,{onClick:()=>{let e=ew.default.dump(t),n=new Blob([e],{type:"text/yaml;charset=utf-8"});i(n,"promptfooconfig.yaml"),o()},startIcon:(0,s.jsx)(ej.Z,{}),fullWidth:!0,sx:{justifyContent:"flex-start"},children:"Download YAML Config"}),(0,s.jsx)(F.Z,{onClick:()=>{if(!e){alert("No table data");return}let t=[],s=[...e.head.vars,...e.head.prompts.map(e=>"[".concat(e.provider,"] ").concat(e.label))];t.push(s),e.body.forEach(e=>{let n=[...e.vars,...e.outputs.map(e=>{let{pass:t,text:n}=e;return(t?"[PASS] ":"[FAIL] ")+n})];t.push(n)});let l=(0,eZ.P)(t),a=new Blob([l],{type:"text/csv;charset=utf-8;"});i(a,"".concat(n,"-table.csv")),o()},startIcon:(0,s.jsx)(ej.Z,{}),fullWidth:!0,sx:{justifyContent:"flex-start"},children:"Download Table CSV"}),(0,s.jsx)(F.Z,{onClick:()=>{if(!e){alert("No table data");return}let t=new Blob([JSON.stringify(e,null,2)],{type:"application/json"});i(t,"".concat(n,"-table.json")),o()},startIcon:(0,s.jsx)(ej.Z,{}),fullWidth:!0,sx:{justifyContent:"flex-start"},children:"Download Table JSON"}),(0,s.jsx)(F.Z,{onClick:()=>{if(!e){alert("No table data");return}let t=e.body.map((t,n)=>({chosen:t.outputs.filter(e=>e.pass).map(e=>e.text),rejected:t.outputs.filter(e=>!e.pass).map(e=>e.text),vars:t.test.vars,providers:e.head.prompts.map(e=>e.provider),prompts:e.head.prompts.map(e=>e.label||e.display||e.raw)})),s=new Blob([JSON.stringify(t,null,2)],{type:"application/json"});i(s,"".concat(n,"-dpo.json")),o()},startIcon:(0,s.jsx)(ej.Z,{}),fullWidth:!0,sx:{justifyContent:"flex-start"},children:"Download DPO JSON"})]})})})]})},eS=n(58991),ek=e=>{let{recentEvals:t,onRecentEvalSelected:n}=e,[a,r]=(0,l.useState)(!1),i="undefined"!=typeof navigator&&/Mac|iPod|iPhone|iPad/.test(navigator.userAgent),o=()=>{r(!0)};return l.useEffect(()=>{let e=e=>{(e.ctrlKey||e.metaKey)&&"k"===e.key&&(e.preventDefault(),o())};return window.addEventListener("keydown",e),()=>{window.removeEventListener("keydown",e)}},[]),(0,s.jsxs)(s.Fragment,{children:[(0,s.jsx)(H.Z,{title:i?"Search for Evals (⌘ + K)":"Search for Evals (Ctrl + K)",children:(0,s.jsx)(eb.Z,{onClick:o,size:"large",children:(0,s.jsx)(eS.Z,{})})}),(0,s.jsx)(ed,{title:"Open an Eval",open:a,onClose:()=>{r(!1)},recentEvals:t,onRecentEvalSelected:n})]})},eE=n(69179),eR=n(33533),eI=n(41101),eN=n(36001);let eP=["#fd7f6f","#7eb0d5","#b2e061","#bd7ebe","#ffb55a","#ffee65","#beb9db","#fdcce5","#8bd3c7"];function eF(e){let{table:t}=e,n=(0,l.useRef)(null),a=(0,l.useRef)(null);return(0,l.useEffect)(()=>{if(!n.current)return;a.current&&a.current.destroy();let e=t.body.flatMap(e=>e.outputs.map(e=>e.score)),s=Math.min(...e),l=(Math.ceil(Math.max(...e))-Math.floor(s))/10,r=Array.from({length:11},(e,t)=>parseFloat((Math.floor(s)+t*l).toFixed(2))),i=t.head.prompts.map((e,n)=>{let s=t.body.flatMap(e=>e.outputs[n].score),a=r.map(e=>s.filter(t=>t>=e&&t<e+l).length);return{label:"Column ".concat(n+1),data:a,backgroundColor:eP[n%eP.length]}});a.current=new eN.kL(n.current,{type:"bar",data:{labels:r,datasets:i},options:{animation:!1,plugins:{title:{display:!0,text:"Score Distribution"},legend:{display:!1},tooltip:{callbacks:{title:function(e){let t=e[0].datasetIndex;return"Column ".concat(t+1)},label:function(e){let t=e.dataIndex,n=r[t],s=r[t+1];return s?"".concat(n," <= score < ").concat(s):"".concat(n," <= score")}}}}}})},[t]),(0,s.jsx)("canvas",{ref:n,style:{maxHeight:"300px"}})}function eT(e){let{table:t}=e,n=(0,l.useRef)(null),a=(0,l.useRef)(null);return(0,l.useEffect)(()=>{if(!n.current)return;a.current&&a.current.destroy();let e=t.head.prompts.map((e,n)=>{let s=t.body.flatMap(e=>e.outputs[n]),l=s.filter(e=>e.pass).length,a=l/s.length*100;return{label:"Column ".concat(n+1),data:[a],backgroundColor:eP[n%eP.length]}});a.current=new eN.kL(n.current,{type:"bar",data:{labels:["Pass Rate (%)"],datasets:e},options:{animation:!1,plugins:{title:{display:!0,text:"Pass rate"},legend:{display:!0}}}})},[t]),(0,s.jsx)("canvas",{ref:n,style:{maxHeight:"300px"}})}function eD(e){let{table:t}=e,n=(0,l.useRef)(null),a=(0,l.useRef)(null),[r,i]=(0,l.useState)(0),[o,c]=(0,l.useState)(1),[d,u]=(0,l.useState)(!1);return(0,l.useEffect)(()=>{if(!n.current)return;a.current&&a.current.destroy();let e=t.body.flatMap(e=>e.outputs.map(e=>e.score)),s=Math.min(...e),l=Math.max(...e),i=t.body.map(e=>{let t=e.outputs[r].score,n=e.outputs[o].score;return{x:t,y:n,backgroundColor:n>t?"green":n<t?"red":"gray"}});a.current=new eN.kL(n.current,{type:"scatter",data:{datasets:[{data:i,backgroundColor:i.map(e=>e.backgroundColor)},{type:"line",data:[{x:s,y:s},{x:l,y:l}],borderColor:"gray",borderWidth:1,borderDash:[5,5],pointRadius:0}]},options:{animation:!1,plugins:{legend:{display:!1},tooltip:{callbacks:{label:function(e){let n=t.body[e.dataIndex],s=n.outputs[0].text,l=n.outputs[1].text;return s.length>30&&(s=s.substring(0,30)+"..."),l.length>30&&(l=l.substring(0,30)+"..."),"Output 1: ".concat(s,"\nOutput 2: ").concat(l)}}}},scales:{x:{title:{display:!0,text:"Prompt ".concat(r+1," Score")},ticks:{callback:function(e,t,n){let s=String(Math.round(100*Number(e)));return t===n.length-1&&(s+="%"),s}}},y:{title:{display:!0,text:"Prompt ".concat(o+1," Score")},ticks:{callback:function(e,t,n){let s=String(Math.round(100*Number(e)));return t===n.length-1&&(s+="%"),s}}}}}})},[t,r,o]),(0,s.jsxs)(s.Fragment,{children:[(0,s.jsxs)(Y.Z,{open:d,onClose:()=>u(!1),children:[(0,s.jsx)(Q.Z,{children:"Compare prompt outputs"}),(0,s.jsxs)($.Z,{children:[(0,s.jsx)(M.Z,{sx:{m:1,minWidth:120},children:(0,s.jsx)(_.Z,{value:r,onChange:e=>i(Number(e.target.value)),children:t.head.prompts.map((e,t)=>(0,s.jsxs)(z.Z,{value:t,children:["Prompt ",t+1]},t))})}),(0,s.jsx)(M.Z,{sx:{m:1,minWidth:120},children:(0,s.jsx)(_.Z,{value:o,onChange:e=>c(Number(e.target.value)),children:t.head.prompts.map((e,t)=>(0,s.jsxs)(z.Z,{value:t,children:["Prompt ",t+1]},t))})})]})]}),(0,s.jsx)("canvas",{ref:n,style:{maxHeight:"300px",cursor:"pointer"},onClick:()=>u(!0)})]})}function eM(e){let{table:t}=e,n=(0,l.useRef)(null),a=(0,l.useRef)(null);return(0,l.useEffect)(()=>{var e;if(!n.current)return;a.current&&a.current.destroy();let s=Object.keys((null===(e=t.head.prompts[0].metrics)||void 0===e?void 0:e.namedScores)||{}),l=t.head.prompts.map((e,n)=>{let l=s.map(n=>{var s;let l=(null===(s=e.metrics)||void 0===s?void 0:s.namedScores[n])||0,a=Math.max(...t.head.prompts.map(e=>{var t;return(null===(t=e.metrics)||void 0===t?void 0:t.namedScores[n])||0}));return l/a});return{label:"".concat(t.head.prompts[n].provider),data:l,backgroundColor:eP[n%eP.length]}});a.current=new eN.kL(n.current,{type:"bar",data:{labels:s,datasets:l},options:{scales:{x:{grid:{display:!1}},y:{ticks:{callback:function(e,t,n){let s=String(Math.round(100*Number(e)));return t===n.length-1&&(s+="%"),s}}}},plugins:{tooltip:{callbacks:{title:function(e){return e[0].dataset.label},label:function(e){let t=e.parsed.y;return"".concat(s[e.dataIndex],": ").concat((100*t).toFixed(2),"% pass rate")}}}}}})},[t]),(0,s.jsx)("canvas",{ref:n,style:{maxHeight:"300px"}})}eN.kL.register(eN.vn,eN.ST,eN.ho,eN.uw,eN.f$,eN.ZL,eN.jn,eN.od,eN.u,eN.wL);var eO=l.memo(function(e){var t;let{columnVisibility:n}=e,a=(0,eI.Z)();eN.kL.defaults.color="dark"===a.palette.mode?"#aaa":"#666";let[r,i]=(0,l.useState)(!0),{table:o}=ex();if(!o||!r||o.head.prompts.length<2)return null;let c=o.body.flatMap(e=>e.outputs.map(e=>e.score)),d=new Set(c);return 1===d.size?null:(0,s.jsx)(eE.SV,{fallback:null,children:(0,s.jsxs)(ee.Z,{sx:{position:"relative",padding:3,mt:2},children:[(0,s.jsx)(eb.Z,{style:{position:"absolute",right:0,top:0},onClick:()=>i(!1),children:(0,s.jsx)(eR.Z,{})}),(0,s.jsxs)("div",{style:{display:"flex",justifyContent:"space-between",width:"100%"},children:[(0,s.jsx)("div",{style:{width:"33%"},children:(0,s.jsx)(eT,{table:o})}),(0,s.jsx)("div",{style:{width:"33%"},children:d.size<=3&&Object.keys((null===(t=o.head.prompts[0].metrics)||void 0===t?void 0:t.namedScores)||{}).length>1?(0,s.jsx)(eM,{table:o}):(0,s.jsx)(eF,{table:o})}),(0,s.jsx)("div",{style:{width:"33%"},children:(0,s.jsx)(eD,{table:o})})]})]})})}),eL=n(7660),eA=n(3216),eU=n(5971);n(55974);var eW=e=>{let{lookup:t,metricTotals:n,onSearchTextChange:a}=e,[r,i]=l.useState(!1);if(!t||!Object.keys(t).length)return null;let o=Object.entries(t),c=r?o:o.slice(0,10);return(0,s.jsxs)("div",{className:"custom-metric-container",children:[c.map(e=>{let[t,l]=e;return t&&void 0!==l?(0,s.jsxs)("span",{onClick:()=>a&&a("metric=".concat(t,":")),className:a?"clickable":"",children:[t,":"," ",n&&n[t]?(0,s.jsxs)(s.Fragment,{children:[(l/n[t]*100).toFixed(2),"% (",l.toFixed(2),"/",n[t].toFixed(2),")"]}):l.toFixed(2)]},t):null}),o.length>10&&(0,s.jsx)("span",{className:"clickable",onClick:()=>i(!r),children:r?"Show less":"Show more..."})]})},ez=n(39329),eV=n(64748);function e_(e,t){return e.length<=t?e:e.slice(0,t)+"..."}function eB(e){let{gradingResults:t}=e,[n,a]=(0,l.useState)({});if(!t)return null;let r=e=>{a(t=>({...t,[e]:!t[e]}))};return(0,s.jsxs)(P.Z,{mt:2,children:[(0,s.jsx)(ei.Z,{variant:"subtitle1",children:"Assertions"}),(0,s.jsx)(el.Z,{children:(0,s.jsxs)(et.Z,{children:[(0,s.jsx)(ea.Z,{children:(0,s.jsxs)(er.Z,{children:[(0,s.jsx)(es.Z,{style:{fontWeight:"bold"},children:"Pass"}),(0,s.jsx)(es.Z,{style:{fontWeight:"bold"},children:"Score"}),(0,s.jsx)(es.Z,{style:{fontWeight:"bold"},children:"Type"}),(0,s.jsx)(es.Z,{style:{fontWeight:"bold"},children:"Value"}),(0,s.jsx)(es.Z,{style:{fontWeight:"bold"},children:"Reason"})]})}),(0,s.jsx)(en.Z,{children:t.map((e,t)=>{var l,a,i;if(!e)return null;let o=(null===(l=e.assertion)||void 0===l?void 0:l.value)?"object"==typeof e.assertion.value?JSON.stringify(e.assertion.value,null,2):String(e.assertion.value):"-",c=e_(o,300),d=n[t]||!1;return(0,s.jsxs)(er.Z,{children:[(0,s.jsx)(es.Z,{children:e.pass?"✅":"❌"}),(0,s.jsx)(es.Z,{children:null===(a=e.score)||void 0===a?void 0:a.toFixed(2)}),(0,s.jsx)(es.Z,{children:(null===(i=e.assertion)||void 0===i?void 0:i.type)||""}),(0,s.jsx)(es.Z,{style:{whiteSpace:"pre-wrap",cursor:"pointer"},onClick:()=>r(t),children:d?o:c}),(0,s.jsx)(es.Z,{style:{whiteSpace:"pre-wrap"},children:e.reason})]},t)})})]})})]})}function eJ(e){let{open:t,onClose:n,prompt:a,provider:r,output:i,gradingResults:o,metadata:c}=e,[d,u]=(0,l.useState)(!1),[h,p]=(0,l.useState)({});(0,l.useEffect)(()=>{u(!1)},[a]);let m=async e=>{await navigator.clipboard.writeText(e),u(!0)},x=e=>{p(t=>({...t,[e]:!t[e]}))};return(0,s.jsxs)(Y.Z,{open:t,onClose:n,fullWidth:!0,maxWidth:"lg",children:[(0,s.jsxs)(Q.Z,{children:["Details",r&&": ".concat(r)]}),(0,s.jsxs)($.Z,{children:[(0,s.jsxs)(P.Z,{mb:2,children:[(0,s.jsx)(ei.Z,{variant:"subtitle1",style:{marginBottom:"1rem"},children:"Prompt"}),(0,s.jsx)(eV.Z,{readOnly:!0,value:a,style:{width:"100%",padding:"0.75rem"},maxRows:20}),(0,s.jsx)(eb.Z,{onClick:()=>m(a),style:{position:"absolute",right:"10px",top:"10px"},children:d?(0,s.jsx)(eg.Z,{}):(0,s.jsx)(ez.Z,{})})]}),(null==c?void 0:c.redteamFinalPrompt)&&(0,s.jsxs)(P.Z,{my:2,children:[(0,s.jsx)(ei.Z,{variant:"subtitle1",style:{marginBottom:"1rem",marginTop:"1rem"},children:"Modified User Input (Red Team)"}),(0,s.jsx)(eV.Z,{readOnly:!0,maxRows:20,value:c.redteamFinalPrompt,style:{width:"100%",padding:"0.75rem"}})]}),i&&(0,s.jsxs)(P.Z,{my:2,children:[(0,s.jsx)(ei.Z,{variant:"subtitle1",style:{marginBottom:"1rem",marginTop:"1rem"},children:"Output"}),(0,s.jsx)(eV.Z,{readOnly:!0,maxRows:20,value:i,style:{width:"100%",padding:"0.75rem"}})]}),(0,s.jsx)(eB,{gradingResults:o}),c&&Object.keys(c).length>0&&(0,s.jsxs)(P.Z,{my:2,children:[(0,s.jsx)(ei.Z,{variant:"subtitle1",style:{marginBottom:"1rem",marginTop:"1rem"},children:"Metadata"}),(0,s.jsx)(el.Z,{children:(0,s.jsxs)(et.Z,{size:"small",children:[(0,s.jsx)(ea.Z,{children:(0,s.jsxs)(er.Z,{children:[(0,s.jsx)(es.Z,{children:(0,s.jsx)("strong",{children:"Key"})}),(0,s.jsx)(es.Z,{children:(0,s.jsx)("strong",{children:"Value"})})]})}),(0,s.jsx)(en.Z,{children:Object.entries(c).map(e=>{let[t,n]=e,l="string"==typeof n?n:JSON.stringify(n),a=e_(l,300),r=h[t]||!1;return(0,s.jsxs)(er.Z,{children:[(0,s.jsx)(es.Z,{children:t}),(0,s.jsx)(es.Z,{style:{whiteSpace:"pre-wrap",cursor:"pointer"},onClick:()=>x(t),children:r?l:a})]},t)})})]})})]})]}),(0,s.jsx)(q.Z,{children:(0,s.jsx)(F.Z,{onClick:n,children:"Close"})})]})}var eH=n(89396),eK=n(64173),eX=function(){let e=(0,eI.Z)(),[t,n]=l.useState(!1),a=()=>{n(!1)};return(0,s.jsxs)(s.Fragment,{children:[(0,s.jsx)("div",{style:{textAlign:"center",marginTop:20,marginBottom:40},children:(0,s.jsx)(F.Z,{variant:"text",color:"primary",startIcon:(0,s.jsx)(eH.Z,{}),onClick:()=>{n(!0)},children:"Generate test cases"})}),(0,s.jsxs)(Y.Z,{open:t,onClose:a,children:[(0,s.jsx)(Q.Z,{children:"Run on Command Line"}),(0,s.jsx)($.Z,{children:(0,s.jsxs)(eK.Z,{children:[(0,s.jsx)("p",{children:"This feature is in beta. UI coming soon."}),(0,s.jsxs)("p",{children:["Run"," ",(0,s.jsx)(P.Z,{component:"code",sx:{backgroundColor:"dark"===e.palette.mode?"#424242":"#f0f0f0",padding:"2px 4px",borderRadius:"4px"},children:"promptfoo generate dataset"}),"to generate test cases on the command line."]})]})}),(0,s.jsx)(q.Z,{children:(0,s.jsx)(F.Z,{onClick:a,color:"primary",children:"Close"})})]})]})};function eG(e){return"string"==typeof e||"number"==typeof e?e.toString().length:Array.isArray(e)?e.reduce((e,t)=>e+eG(t),0):l.isValidElement(e)&&e.props.children?l.Children.toArray(e.props.children).reduce((e,t)=>e+eG(t),0):0}let eY=l.memo(function(e){let t,{text:n,maxLength:a}=e,[r,i]=l.useState(!0),o=()=>{i(!r)},c=function(e){let t=arguments.length>1&&void 0!==arguments[1]?arguments[1]:0;if("string"==typeof e||"number"==typeof e){let n=e.toString();return n.slice(0,a-t)}if(Array.isArray(e)){let n=[],s=t;for(let t of e){let e=eG(t);if(s+e>a){n.push(c(t,s));break}n.push(t),s+=e}return n}if(l.isValidElement(e)&&e.props.children){let n=eG(e.props.children);if(n>a-t)return l.cloneElement(e,{...e.props,children:c(e.props.children,t)})}return e};t=l.isValidElement(n)||"string"==typeof n?n:JSON.stringify(n);let d=r?c(t):t,u=eG(t)>a;return(0,s.jsxs)("div",{style:{cursor:u?"pointer":"normal"},onMouseDown:e=>{let t=e.clientX,n=e.clientY,s=e=>{let l=e.clientX,a=e.clientY;5>Math.abs(l-t)&&5>Math.abs(a-n)&&o(),document.removeEventListener("mouseup",s)};document.addEventListener("mouseup",s)},children:[d,r&&eG(t)>a&&(0,s.jsx)("span",{children:"..."})]})});var eq=n(45391),e$=n(40182),eQ=n(38640),e0=n.n(eQ),e1=n(33352),e2=n(15014),e3=n(65101);n(16658);var e5=e=>{let{failReasons:t}=e,[n,a]=(0,l.useState)(0);return t.length<1?null:(0,s.jsxs)("div",{className:"fail-reason",children:[t.length>1&&(0,s.jsxs)("span",{className:"fail-reason-carousel-controls",children:[(0,s.jsx)(eb.Z,{onClick:()=>{a(e=>e>0?e-1:t.length-1)},children:(0,s.jsx)(e2.Z,{sx:{fontSize:12}})}),(0,s.jsxs)("span",{children:[n+1,"/",t.length]}),(0,s.jsx)(eb.Z,{onClick:()=>{a(e=>e<t.length-1?e+1:0)},children:(0,s.jsx)(e3.Z,{sx:{fontSize:12}})})]}),t[n].trim().split("\n").map((e,t)=>(0,s.jsxs)(l.Fragment,{children:[e,(0,s.jsx)("br",{})]},t))]})},e4=e=>{let{open:t,contextText:n,commentText:l,onClose:a,onSave:r,onChange:i}=e,o="dark"===(0,eI.Z)().palette.mode;return(0,s.jsxs)(Y.Z,{open:t,onClose:a,fullWidth:!0,maxWidth:"sm",children:[(0,s.jsx)(Q.Z,{children:"Edit Comment"}),(0,s.jsxs)($.Z,{children:[(0,s.jsx)(P.Z,{sx:{backgroundColor:o?"#1e1e1e":"#f0f0f0",padding:2,marginBottom:2},children:n}),(0,s.jsx)(J.Z,{autoFocus:!0,margin:"dense",type:"text",fullWidth:!0,multiline:!0,rows:4,value:l,onChange:e=>i(e.target.value)})]}),(0,s.jsxs)(q.Z,{children:[(0,s.jsx)(F.Z,{onClick:r,color:"primary",variant:"contained",children:"Save"}),(0,s.jsx)(F.Z,{onClick:a,color:"primary",children:"Cancel"})]})]})};let e8=()=>{let e=(0,l.useContext)(c);if(void 0===e)throw Error("useShiftKey must be used within a ShiftKeyProvider");return e};var e7=n(68192);let e9=l.memo(function(e){var t,n,a,r,i,o,c,d,u,h,p,m,x;let v,g,j,f,b,y,{output:Z,maxTextLength:w,rowIndex:C,promptIndex:S,onRating:k,firstOutput:E,showDiffs:R,searchText:I,showStats:N}=e,{renderMarkdown:P,prettifyJson:F,showPrompts:T,showPassFail:D}=ex(),[M,O]=l.useState(!1),[L,A]=l.useState(!1),[U,W]=l.useState(null),z=e=>{W(e||null),A(!L)},[V,_]=l.useState(!1),[B,J]=l.useState((null===(t=Z.gradingResult)||void 0===t?void 0:t.comment)||""),K=()=>{_(!0)},X="string"==typeof Z.text?Z.text:JSON.stringify(Z.text),G=[];if(!Z.pass&&X.includes("---")&&(G=((null===(h=Z.gradingResult)||void 0===h?void 0:h.componentResults)||[]).filter(e=>!!e&&!e.pass).map(e=>e.reason),X=X.split("---").slice(1).join("---")),R&&E){let e,t="string"==typeof E.text?E.text:JSON.stringify(E.text);t.includes("---")&&(t=t.split("---").slice(1).join("---"));try{JSON.parse(t),JSON.parse(X),e=(0,e7.CT)(t,X)}catch(n){e=t.includes(". ")&&X.includes(". ")?(0,e7.SY)(t,X):(0,e7.NV)(t,X)}v=(0,s.jsx)(s.Fragment,{children:e.map((e,t)=>e.added?(0,s.jsx)("ins",{children:e.value},t):e.removed?(0,s.jsx)("del",{children:e.value},t):(0,s.jsx)("span",{children:e.value},t))})}if(I)try{let e;let t=RegExp(I,"gi"),n=[];for(;null!==(e=t.exec(X));)n.push({start:e.index,end:t.lastIndex});v=(0,s.jsx)(s.Fragment,{children:n.length>0?(0,s.jsxs)(s.Fragment,{children:[(0,s.jsx)("span",{children:X.substring(0,n[0].start)},"text-before"),n.map((e,t)=>(0,s.jsxs)(s.Fragment,{children:[(0,s.jsx)("span",{className:"search-highlight",children:X.substring(e.start,e.end)},"match-"+t),(0,s.jsx)("span",{children:X.substring(e.end,n[t+1]?n[t+1].start:X.length)},"text-after-"+t)]}))]}):(0,s.jsx)("span",{children:X},"no-match")})}catch(e){console.error("Invalid regular expression:",e.message)}else if(P&&!R)v=(0,s.jsx)(eU.U,{remarkPlugins:[e1.Z],components:{img:e=>{let{src:t,alt:n}=e;return(0,s.jsx)("img",{loading:"lazy",src:t,alt:n,onClick:()=>z(t),style:{cursor:"pointer"}})}},children:X});else if(F)try{v=(0,s.jsx)("pre",{children:JSON.stringify(JSON.parse(X),null,2)})}catch(e){}let Y=l.useCallback(e=>{var t;k(e,void 0,null===(t=Z.gradingResult)||void 0===t?void 0:t.comment)},[k,null===(n=Z.gradingResult)||void 0===n?void 0:n.comment]),q=l.useCallback(()=>{let e=prompt("Set test score (0.0 - 1.0):",String(Z.score));if(null!==e){let n=parseFloat(e);if(!isNaN(n)&&n>=0&&n<=1){var t;k(void 0,n,null===(t=Z.gradingResult)||void 0===t?void 0:t.comment)}else alert("Invalid score. Please enter a value between 0.0 and 1.0.")}},[k,Z.score,null===(a=Z.gradingResult)||void 0===a?void 0:a.comment]),[$,Q]=l.useState(!1),ee=l.useCallback(()=>{navigator.clipboard.writeText(Z.text),Q(!0)},[Z.text]);if(Z.latencyMs&&(j=(0,s.jsxs)("span",{children:[Intl.NumberFormat(void 0,{maximumFractionDigits:0}).format(Z.latencyMs)," ms"]})),null===(r=Z.tokenUsage)||void 0===r?void 0:r.completion){let e=Z.tokenUsage.completion/(Z.latencyMs/1e3);f=(0,s.jsx)("span",{children:Intl.NumberFormat(void 0,{maximumFractionDigits:0}).format(e)})}if(Z.cost&&(b=(0,s.jsxs)("span",{children:["$",Z.cost.toPrecision(2)]})),null===(i=Z.tokenUsage)||void 0===i?void 0:i.cached)g=(0,s.jsxs)("span",{children:[Intl.NumberFormat(void 0,{maximumFractionDigits:0}).format(Z.tokenUsage.cached)," ","(cached)"]});else if(null===(o=Z.tokenUsage)||void 0===o?void 0:o.total){let e=Intl.NumberFormat(void 0,{maximumFractionDigits:0}).format(null!==(p=Z.tokenUsage.prompt)&&void 0!==p?p:0),t=Intl.NumberFormat(void 0,{maximumFractionDigits:0}).format(null!==(m=Z.tokenUsage.completion)&&void 0!==m?m:0),n=Intl.NumberFormat(void 0,{maximumFractionDigits:0}).format(Z.tokenUsage.total);g=(0,s.jsx)(H.Z,{title:"".concat(e," prompt tokens + ").concat(t," completion tokens = ").concat(n," total"),children:(0,s.jsxs)("span",{children:[n,("0"!==e||"0"!==t)&&" (".concat(e,"+").concat(t,")")]})})}let et=(null===(c=Z.gradingResult)||void 0===c?void 0:c.comment)&&"!highlight"!==Z.gradingResult.comment?(0,s.jsx)("div",{className:"comment",onClick:K,children:Z.gradingResult.comment}):null,en=N?(0,s.jsxs)("div",{className:"cell-detail",children:[g&&(0,s.jsxs)("div",{className:"stat-item",children:[(0,s.jsx)("strong",{children:"Tokens:"})," ",g]}),j&&(0,s.jsxs)("div",{className:"stat-item",children:[(0,s.jsx)("strong",{children:"Latency:"})," ",j]}),f&&(0,s.jsxs)("div",{className:"stat-item",children:[(0,s.jsx)("strong",{children:"Tokens/Sec:"})," ",f]}),b&&(0,s.jsxs)("div",{className:"stat-item",children:[(0,s.jsx)("strong",{children:"Cost:"})," ",b]})]}):null,es=e8(),el=(0,s.jsxs)("div",{className:"cell-actions",children:[es&&(0,s.jsxs)(s.Fragment,{children:[(0,s.jsx)("span",{className:"action",onClick:ee,onMouseDown:e=>e.preventDefault(),children:(0,s.jsx)(H.Z,{title:"Copy output to clipboard",children:(0,s.jsx)("span",{children:$?"✅":"\uD83D\uDCCB"})})}),(0,s.jsx)("span",{className:"action",onClick:()=>{let e;B.startsWith("!highlight")?k(void 0,void 0,e=B.slice(10).trim()):k(void 0,void 0,e=("!highlight "+B).trim()),J(e)},onMouseDown:e=>e.preventDefault(),children:(0,s.jsx)(H.Z,{title:"Toggle test highlight",children:(0,s.jsx)("span",{children:"\uD83C\uDF1F"})})})]}),Z.prompt&&(0,s.jsxs)(s.Fragment,{children:[(0,s.jsx)("span",{className:"action",onClick:()=>{O(!0)},children:(0,s.jsx)(H.Z,{title:"View output and test details",children:(0,s.jsx)("span",{children:"\uD83D\uDD0E"})})}),(0,s.jsx)(eJ,{open:M,onClose:()=>{O(!1)},prompt:Z.prompt,provider:Z.provider,gradingResults:null===(d=Z.gradingResult)||void 0===d?void 0:d.componentResults,output:X,metadata:Z.metadata})]}),(0,s.jsx)("span",{className:"action",onClick:()=>Y(!0),children:(0,s.jsx)(H.Z,{title:"Mark test passed (score 1.0)",children:(0,s.jsx)("span",{children:"\uD83D\uDC4D"})})}),(0,s.jsx)("span",{className:"action",onClick:()=>Y(!1),children:(0,s.jsx)(H.Z,{title:"Mark test failed (score 0.0)",children:(0,s.jsx)("span",{children:"\uD83D\uDC4E"})})}),(0,s.jsx)("span",{className:"action",onClick:q,children:(0,s.jsx)(H.Z,{title:"Set test score",children:(0,s.jsx)("span",{children:"\uD83D\uDD22"})})}),(0,s.jsx)("span",{className:"action",onClick:K,children:(0,s.jsx)(H.Z,{title:"Edit comment",children:(0,s.jsx)("span",{children:"✏️"})})})]}),ea={};(null===(u=Z.gradingResult)||void 0===u?void 0:u.comment)==="!highlight"&&(ea.backgroundColor="#ffffeb");let er=0,ei=0,eo=Z.gradingResult;if(eo?eo.componentResults?eo.componentResults.forEach(e=>{(null==e?void 0:e.pass)?er++:ei++}):(er=eo.pass?1:0,ei=eo.pass?0:1):Z.pass?er=1:Z.pass||(ei=1),1===ei&&1===er)y=(0,s.jsxs)(s.Fragment,{children:["".concat(ei," FAIL")," ","".concat(er," PASS")]});else{let e="";ei>1||er>1&&ei>0?e="".concat(ei," FAIL"):1===ei&&(e="FAIL");let t="";er>1||ei>1&&er>0?t="".concat(er," PASS"):1===er&&0===ei&&(t="PASS");let n=e&&t?" ":"";y=(0,s.jsxs)(s.Fragment,{children:[e,n,t]})}let ec=null===(x=Z.score)||0===x||1===x?"":"(".concat(x.toFixed(2),")");return(0,s.jsxs)("div",{className:"cell",style:ea,children:[D&&(0,s.jsx)(s.Fragment,{children:Z.pass?(0,s.jsx)(s.Fragment,{children:(0,s.jsxs)("div",{className:"status pass",children:[(0,s.jsxs)("div",{className:"pill",children:[y,ec&&(0,s.jsxs)("span",{className:"score",children:[" ",ec]})]}),(0,s.jsx)(eW,{lookup:Z.namedScores})]})}):(0,s.jsx)(s.Fragment,{children:(0,s.jsxs)("div",{className:"status fail",children:[(0,s.jsxs)("div",{className:"pill",children:[y,ec&&(0,s.jsxs)("span",{className:"score",children:[" ",ec]})]}),(0,s.jsx)(eW,{lookup:Z.namedScores}),(0,s.jsx)("span",{className:"fail-reason",children:(0,s.jsx)(e5,{failReasons:G})})]})})}),T&&E.prompt&&(0,s.jsxs)("div",{className:"prompt",children:[(0,s.jsx)("span",{className:"pill",children:"Prompt"}),Z.prompt]}),(0,s.jsx)(eY,{text:v||X,maxLength:w}),et,en,el,L&&U&&(0,s.jsx)("div",{className:"lightbox",onClick:()=>z(),children:(0,s.jsx)("img",{src:U,alt:"Lightbox"})}),(0,s.jsx)(e4,{open:V,contextText:Z.text,commentText:B,onClose:()=>{_(!1)},onSave:()=>{k(void 0,void 0,B),_(!1)},onChange:J})]})});function e6(e){let{text:t,maxLength:n,expandedText:a,resourceId:r,className:i}=e,[o,c]=l.useState(!1);return(0,s.jsxs)("div",{className:"".concat(i||""),children:[(0,s.jsx)(eY,{text:t,maxLength:n}),a&&(0,s.jsxs)(s.Fragment,{children:[(0,s.jsx)(H.Z,{title:"View prompt",children:(0,s.jsx)("span",{className:"action",onClick:()=>{c(!0)},children:"\uD83D\uDD0E"})}),(0,s.jsx)(eJ,{open:o,onClose:()=>{c(!1)},prompt:a}),r&&(0,s.jsx)(H.Z,{title:"View other evals and datasets for this prompt",children:(0,s.jsx)("span",{className:"action",children:(0,s.jsx)(e0(),{href:"/prompts/?id=".concat(r),target:"_blank",children:(0,s.jsx)(eq.Z,{fontSize:"small"})})})})]})]})}n(93207);var te=l.memo(function(e){let{maxTextLength:t,columnVisibility:n,wordBreak:a,filterMode:r,failureFilter:i,searchText:c,showStats:d,onFailureFilterToggle:u,onSearchTextChange:h}=e,{evalId:p,table:x,setTable:v,config:g,inComparisonMode:j}=ex(),{showToast:f}=m();(0,y.Z)(x,"Table should be defined");let{head:b,body:Z}=x,w=l.useCallback(async(e,t,n,s,l)=>{var a,r;let i=[...Z],c={...i[e]},d=[...c.outputs],u=null!=n?n:d[t].pass,h=void 0===s?n?1:0:s||0;d[t].pass=u,d[t].score=h;let m=(null===(a=d[t].gradingResult)||void 0===a?void 0:a.componentResults)||[];if(void 0!==n){let e=m.findIndex(e=>{var t;return(null===(t=e.assertion)||void 0===t?void 0:t.type)==="human"}),t={pass:u,score:h,reason:"Manual result (overrides all other grading results)",comment:l,assertion:{type:"human"}};-1!==e?m[e]=t:m.push(t)}let x={...d[t].gradingResult||{},pass:u,score:h,reason:"Manual result (overrides all other grading results)",comment:l,assertion:(null===(r=d[t].gradingResult)||void 0===r?void 0:r.assertion)||null,componentResults:m};d[t].gradingResult=x,c.outputs=d,i[e]=c;let g={head:b,body:i};if(v(g),j)f("Ratings are not saved in comparison mode","warning");else try{let e=await fetch("".concat(await (0,o.b)(),"/api/eval/").concat(p),{method:"PATCH",headers:{"Content-Type":"application/json"},body:JSON.stringify({table:g})});if(!e.ok)throw Error("Network response was not ok")}catch(e){console.error("Failed to update table:",e)}},[Z,b,v,p,j,f]),C=Object.keys(n).length>0,S=l.useMemo(()=>{try{return RegExp(c,"i")}catch(e){return console.error("Invalid regular expression:",e.message),null}},[c]),k=l.useMemo(()=>{try{return Z.map((e,t)=>({...e,outputs:e.outputs.map((e,n)=>({...e,originalRowIndex:t,originalPromptIndex:n}))})).filter(e=>{let t=!0;return"failures"===r?t=e.outputs.some((e,t)=>{let s="Prompt ".concat(t+1);return i[s]&&!e.pass&&(!C||n[s])}):"different"===r?t=!e.outputs.every(t=>t.text===e.outputs[0].text):"highlights"===r&&(console.log(e.outputs[0].text),t=e.outputs.some(e=>{var t,n;return null===(n=e.gradingResult)||void 0===n?void 0:null===(t=n.comment)||void 0===t?void 0:t.startsWith("!highlight")})),!!t&&(!c||!S||e.outputs.some(t=>{var n,s;let l=e.vars.map(e=>"var=".concat(e)).join(" "),a="".concat(t.text," ").concat(Object.keys(t.namedScores).map(e=>"metric=".concat(e,":").concat(t.namedScores[e])).join(" ")," ").concat((null===(n=t.gradingResult)||void 0===n?void 0:n.reason)||""," ").concat((null===(s=t.gradingResult)||void 0===s?void 0:s.comment)||""),r="".concat(l," ").concat(a);return S.test(r)}))})}catch(e){return console.error("Invalid regular expression:",e.message),Z}},[Z,i,r,c,n,C,S]),[E,R]=l.useState({pageIndex:0,pageSize:50});l.useEffect(()=>{R(e=>({...e,pageIndex:0}))},[i,r,c]);let I=l.useMemo(()=>b.prompts.map((e,t)=>Z.reduce((e,n)=>e+(n.outputs[t].pass?1:0),0)),[b.prompts,Z]),N=l.useMemo(()=>b.prompts.map((e,t)=>Z.reduce((e,n)=>{var s,l;return e+((null===(l=n.outputs[t].gradingResult)||void 0===l?void 0:null===(s=l.componentResults)||void 0===s?void 0:s.length)||0)},0)),[b.prompts,Z]),D=l.useMemo(()=>b.prompts.map((e,t)=>Z.reduce((e,n)=>{var s;let l=null===(s=n.outputs[t].gradingResult)||void 0===s?void 0:s.componentResults;return e+(l?l.filter(e=>null==e?void 0:e.pass).length:0)},0)),[b.prompts,Z]),M=l.useMemo(()=>I.reduce((e,t,n,s)=>t>s[e]?n:e,0),[I]),O=I[M],L=l.useMemo(()=>(0,eL.Cl)(),[]),{renderMarkdown:A}=ex(),U=l.useMemo(()=>b.vars.length>0?[L.group({id:"vars",header:()=>(0,s.jsx)("span",{className:"font-bold",children:"Variables"}),columns:b.vars.map((e,n)=>L.accessor(e=>e.vars[n],{id:"Variable ".concat(n+1),header:()=>(0,s.jsx)(e6,{text:e,maxLength:t,className:"font-bold"}),cell:e=>{let n=e.getValue();return(0,s.jsx)("div",{className:"cell",children:A?(0,s.jsx)(eU.U,{remarkPlugins:[e1.Z],children:n}):(0,s.jsx)(eY,{text:n,maxLength:t})})},size:50}))})]:[],[L,b.vars,t,A]),W=l.useCallback((e,t)=>k[e].outputs[t],[k]),V=l.useCallback(e=>k[e].outputs[0],[k]),B=l.useMemo(()=>{let e={};return null==x||x.body.forEach(t=>{var n;null===(n=t.test.assert)||void 0===n||n.forEach(t=>{t.metric&&(e[t.metric]=(e[t.metric]||0)+1),"assert"in t&&Array.isArray(t.assert)&&t.assert.forEach(t=>{"metric"in t&&t.metric&&(e[t.metric]=(e[t.metric]||0)+1)})})}),e},[x]),K=l.useMemo(()=>[L.group({id:"prompts",header:()=>(0,s.jsx)("span",{className:"font-bold",children:"Outputs"}),columns:b.prompts.map((e,n)=>L.accessor(e=>(function(e){if("string"==typeof e){let t=e.startsWith("[PASS]"),n=e;return e.startsWith("[PASS]")?n=n.slice(6):e.startsWith("[FAIL]")&&(n=n.slice(6)),{text:n,pass:t,score:t?1:0}}return e})(e.outputs[n]),{id:"Prompt ".concat(n+1),header:()=>{var l,a,o,c,p,m,x,v;let j=I[n]&&Z.length?(I[n]/Z.length*100).toFixed(2):"0.00",f=I[n]===O&&0!==O,b="Prompt ".concat(n+1),y=i[b]||!1,w=d?(0,s.jsxs)("div",{className:"prompt-detail",children:[N[n]?(0,s.jsxs)("div",{children:[(0,s.jsx)("strong",{children:"Asserts:"})," ",D[n],"/",N[n]," passed"]}):null,(null===(l=e.metrics)||void 0===l?void 0:l.totalLatencyMs)?(0,s.jsxs)("div",{children:[(0,s.jsx)("strong",{children:"Avg Latency:"})," ",Intl.NumberFormat(void 0,{maximumFractionDigits:0}).format(e.metrics.totalLatencyMs/Z.length)," ","ms"]}):null,(null===(o=e.metrics)||void 0===o?void 0:null===(a=o.tokenUsage)||void 0===a?void 0:a.total)?(0,s.jsxs)("div",{children:[(0,s.jsx)("strong",{children:"Avg Tokens:"})," ",Intl.NumberFormat(void 0,{maximumFractionDigits:0}).format(e.metrics.tokenUsage.total/Z.length)]}):null,(null===(c=e.metrics)||void 0===c?void 0:c.totalLatencyMs)&&(null===(m=e.metrics)||void 0===m?void 0:null===(p=m.tokenUsage)||void 0===p?void 0:p.completion)?(0,s.jsxs)("div",{children:[(0,s.jsx)("strong",{children:"Tokens/Sec:"})," ",Intl.NumberFormat(void 0,{maximumFractionDigits:0}).format(e.metrics.tokenUsage.completion/(e.metrics.totalLatencyMs/1e3))]}):null,(null===(x=e.metrics)||void 0===x?void 0:x.cost)?(0,s.jsxs)("div",{children:[(0,s.jsx)("strong",{children:"Cost:"})," $",e.metrics.cost.toPrecision(2)]}):null]}):null,C=Array.isArray(null==g?void 0:g.providers)?g.providers[n]:void 0,S=e.provider?e.provider.split(":"):[],k=(0,s.jsx)(H.Z,{title:C?(0,s.jsx)("pre",{children:ew.default.dump(C)}):"",children:S.length>1?(0,s.jsxs)(s.Fragment,{children:[S[0],":",(0,s.jsx)("strong",{children:S.slice(1).join(":")})]}):(0,s.jsx)("strong",{children:e.provider})});return(0,s.jsxs)("div",{className:"output-header",children:[(0,s.jsxs)("div",{className:"pills",children:[e.provider?(0,s.jsx)("div",{className:"provider",children:k}):null,(0,s.jsx)("div",{className:"summary",children:(0,s.jsxs)("div",{className:"highlight ".concat(f?"success":""),children:[(0,s.jsxs)("strong",{children:[j,"% passing"]})," (",I[n],"/",Z.length," cases)"]})}),(null===(v=e.metrics)||void 0===v?void 0:v.namedScores)&&Object.keys(e.metrics.namedScores).length>0?(0,s.jsx)(eW,{lookup:e.metrics.namedScores,metricTotals:B,onSearchTextChange:h}):null]}),(0,s.jsx)(e6,{className:"prompt-container",text:e.label||e.display||e.raw,expandedText:e.raw,maxLength:t,resourceId:e.id}),w,"failures"===r&&(0,s.jsx)(e$.Z,{sx:{"& .MuiFormControlLabel-label":{fontSize:"0.75rem"}},control:(0,s.jsx)(T.Z,{checked:y,onChange:e=>u(b,e.target.checked)}),label:"Show failures"})]})},cell:e=>{var l,a;let i=W(e.row.index,n);return(0,s.jsx)(e9,{output:i,maxTextLength:t,rowIndex:e.row.index,promptIndex:n,onRating:w.bind(null,null!==(l=i.originalRowIndex)&&void 0!==l?l:e.row.index,null!==(a=i.originalPromptIndex)&&void 0!==a?a:n),firstOutput:V(e.row.index),showDiffs:"different"===r,searchText:c,showStats:d})}}))})],[Z.length,null==g?void 0:g.providers,L,i,r,V,W,w,b.prompts,O,t,B,N,D,I,u,h,c,d]),X=l.useMemo(()=>{let e=Z.some(e=>e.description);return e?{accessorFn:e=>e.description||"",id:"description",header:()=>(0,s.jsx)("span",{className:"font-bold",children:"Description"}),cell:e=>(0,s.jsx)("div",{className:"cell",children:(0,s.jsx)(eY,{text:String(e.getValue()),maxLength:t})}),size:50}:null},[Z,t]),G=l.useMemo(()=>{let e=[];return X&&e.push(X),e.push(...U,...K),e},[X,U,K]),Y=(0,eA.b7)({data:k,columns:G,columnResizeMode:"onChange",getCoreRowModel:(0,eL.sC)(),getPaginationRowModel:(0,eL.G_)(),state:{columnVisibility:n,pagination:E}});return(0,s.jsxs)("div",{children:[(0,s.jsxs)("table",{className:"results-table firefox-fix ".concat(t<=25?"compact":""),style:{wordBreak:a},children:[(0,s.jsx)("thead",{children:Y.getHeaderGroups().map(e=>(0,s.jsx)("tr",{className:"header",children:e.headers.map(e=>(0,s.jsxs)("th",{colSpan:e.colSpan,style:{width:e.getSize()},children:[e.isPlaceholder?null:(0,eA.ie)(e.column.columnDef.header,e.getContext()),(0,s.jsx)("div",{onMouseDown:e.getResizeHandler(),onTouchStart:e.getResizeHandler(),className:"resizer ".concat(e.column.getIsResizing()?"isResizing":"")})]},e.id))},e.id))}),(0,s.jsx)("tbody",{children:Y.getRowModel().rows.map((e,t)=>{let n=!1;return(0,s.jsx)("tr",{children:e.getVisibleCells().map(e=>{let l=e.column.id.startsWith("Variable")||"description"===e.column.id,a=!l&&!n;return a&&(n=!0),(0,s.jsx)("td",{style:{width:e.column.getSize()},className:"".concat(l?"variable":""," ").concat(0!==t||l?"":"first-prompt-row"," ").concat(a?"first-prompt-col":""),children:(0,eA.ie)(e.column.columnDef.cell,e.getContext())},e.id)})},e.id)})})]}),Y.getPageCount()>1&&(0,s.jsxs)(P.Z,{className:"pagination",mx:1,sx:{display:"flex",alignItems:"center",gap:2},children:[(0,s.jsx)(F.Z,{onClick:()=>{R(e=>({...e,pageIndex:Math.max(e.pageIndex-1,0)})),window.scrollTo(0,0)},disabled:0===Y.getState().pagination.pageIndex,variant:"contained",children:"Previous"}),(0,s.jsxs)(ei.Z,{component:"span",sx:{display:"flex",alignItems:"center",gap:1},children:["Page",(0,s.jsx)(J.Z,{size:"small",type:"number",value:Y.getState().pagination.pageIndex+1,onChange:e=>{let t=e.target.value?Number(e.target.value)-1:0;R(e=>({...e,pageIndex:Math.min(Math.max(t,0),Y.getPageCount()-1)}))},InputProps:{style:{width:"60px",textAlign:"center"}},variant:"outlined"}),(0,s.jsxs)("span",{children:["of ",Y.getPageCount()]})]}),(0,s.jsx)(F.Z,{onClick:()=>{R(e=>({...e,pageIndex:Math.min(e.pageIndex+1,Y.getPageCount()-1)})),window.scrollTo(0,0)},disabled:Y.getState().pagination.pageIndex+1>=Y.getPageCount(),variant:"contained",children:"Next"}),(0,s.jsxs)(ei.Z,{component:"span",sx:{display:"flex",alignItems:"center",gap:1},children:[(0,s.jsxs)(_.Z,{value:E.pageSize,onChange:e=>{R({pageIndex:0,pageSize:Number(e.target.value)}),window.scrollTo(0,0)},displayEmpty:!0,inputProps:{"aria-label":"Results per page"},size:"small",sx:{m:1,minWidth:80},children:[(0,s.jsx)(z.Z,{value:10,children:"10"}),(0,s.jsx)(z.Z,{value:50,children:"50"}),(0,s.jsx)(z.Z,{value:100,children:"100"}),(0,s.jsx)(z.Z,{value:500,children:"500"}),(0,s.jsx)(z.Z,{value:1e3,children:"1000"})]}),(0,s.jsx)("span",{children:"results per page"})]})]}),(0,s.jsx)(eX,{})]})}),tt=n(4451),tn=e=>{let{open:t,onClose:n}=e,{maxTextLength:l,setMaxTextLength:a,wordBreak:r,setWordBreak:i,showInferenceDetails:o,setShowInferenceDetails:c,renderMarkdown:d,setRenderMarkdown:u,prettifyJson:h,setPrettifyJson:p,showPrompts:m,setShowPrompts:x,showPassFail:v,setShowPassFail:g}=ex();return(0,s.jsxs)(Y.Z,{open:t,onClose:n,fullWidth:!0,maxWidth:"sm",children:[(0,s.jsx)(Q.Z,{children:"Table View Settings"}),(0,s.jsxs)($.Z,{children:[(0,s.jsx)(P.Z,{children:(0,s.jsx)(H.Z,{title:"Forcing line breaks makes it easier to adjust column widths to your liking",placement:"right",children:(0,s.jsx)(e$.Z,{control:(0,s.jsx)(T.Z,{checked:"break-all"===r,onChange:e=>i(e.target.checked?"break-all":"break-word")}),label:"Force line breaks"})})}),(0,s.jsx)(P.Z,{children:(0,s.jsx)(e$.Z,{control:(0,s.jsx)(T.Z,{checked:d,onChange:e=>u(e.target.checked)}),label:"Render model outputs as Markdown"})}),(0,s.jsx)(P.Z,{children:(0,s.jsx)(e$.Z,{control:(0,s.jsx)(T.Z,{checked:h,onChange:e=>p(e.target.checked)}),label:"Prettify JSON outputs"})}),(0,s.jsx)(P.Z,{children:(0,s.jsx)(H.Z,{title:"Show the final prompt that produced the output in each cell.",placement:"right",children:(0,s.jsx)(e$.Z,{control:(0,s.jsx)(T.Z,{checked:m,onChange:e=>x(e.target.checked)}),label:"Show full prompt in output cell"})})}),(0,s.jsx)(P.Z,{children:(0,s.jsx)(H.Z,{title:"Show pass/fail status for each output.",placement:"right",children:(0,s.jsx)(e$.Z,{control:(0,s.jsx)(T.Z,{checked:v,onChange:e=>g(e.target.checked)}),label:"Show pass/fail status"})})}),(0,s.jsx)(P.Z,{children:(0,s.jsx)(H.Z,{title:"Show detailed inference statistics such as latency, tokens used, cost, etc.",placement:"right",children:(0,s.jsx)(e$.Z,{control:(0,s.jsx)(T.Z,{checked:o,onChange:e=>c(e.target.checked)}),label:"Show inference details"})})}),(0,s.jsxs)(P.Z,{maxWidth:"sm",children:[(0,s.jsxs)(ei.Z,{mt:2,children:["Max text length: ",l]}),(0,s.jsx)(tt.ZP,{min:25,max:1e3,value:l,onChange:(e,t)=>a(t)})]})]}),(0,s.jsx)(q.Z,{children:(0,s.jsx)(F.Z,{onClick:n,children:"Close"})})]})},ts=e=>{let{open:t,onClose:n,shareUrl:a}=e,r=(0,l.useRef)(null),[i,o]=(0,l.useState)(!1),c=()=>{n(),o(!1)};return(0,s.jsxs)(Y.Z,{open:t,onClose:c,PaperProps:{style:{minWidth:"min(660px, 100%)"}},children:[(0,s.jsx)(Q.Z,{children:"Your eval is ready to share"}),(0,s.jsxs)($.Z,{children:[(0,s.jsx)(J.Z,{inputRef:r,value:a,fullWidth:!0,InputProps:{readOnly:!0,endAdornment:(0,s.jsx)(eb.Z,{onClick:()=>{r.current&&(r.current.select(),document.execCommand("copy"),o(!0))},children:i?(0,s.jsx)(eg.Z,{}):(0,s.jsx)(ef.Z,{})})}}),(0,s.jsx)(eK.Z,{sx:{fontSize:"0.75rem"},children:"Shared URLs are deleted after 2 weeks."})]}),(0,s.jsx)(q.Z,{children:(0,s.jsx)(F.Z,{onClick:c,color:"primary",children:"Close"})})]})};n(58022);let tl=(0,K.Z)(B.Z)(e=>{let{theme:t}=e;return{maxWidth:"100%",flexWrap:"wrap",[t.breakpoints.down("sm")]:{flexDirection:"column"}}});function ta(e){var t;let{recentEvals:n,onRecentEvalSelected:a,defaultEvalId:c}=e,d=(0,f.useRouter)(),h=(0,f.useSearchParams)(),{author:p,table:m,setTable:x,config:v,setConfig:j,maxTextLength:b,wordBreak:B,showInferenceDetails:K,evalId:G,setInComparisonMode:Y,columnStates:q,setColumnState:$}=ex(),{setStateFromConfig:Q}=(0,Z.o)(),[ee,et]=l.useState((null==h?void 0:h.get("search"))||""),[en]=(0,X.Nr)(ee,1e3),es=e=>{et(e)},[el,ea]=l.useState({}),er=l.useCallback((e,t)=>{ea(n=>({...n,[e]:t}))},[ea]);(0,y.Z)(m,"Table data must be loaded before rendering ResultsView");let{head:ei}=m,[eo,ec]=l.useState("all"),[eu,eh]=l.useState(!1),[ep,em]=l.useState(""),[eg,ej]=l.useState(!1),[ef,eb]=l.useState(null),eZ=async()=>{ej(!0);try{let e=await fetch("".concat(r,"/api/eval"),{method:"POST",headers:{"Content-Type":"application/json"},body:JSON.stringify({data:{version:2,createdAt:new Date().toISOString(),results:{table:m},config:v}})}),{id:t}=await e.json(),n="".concat(i,"/eval/").concat(t);em(n),eh(!0)}catch(e){alert("Sorry, something went wrong.")}finally{ej(!1)}},ew=async e=>{eb(null);try{var t;let n=await fetch("".concat(await (0,o.b)(),"/api/results/").concat(e),{cache:"no-store"}),s=await n.json(),l=s.data.results.table,a={head:{prompts:[...m.head.prompts.map(e=>({...e,label:"[".concat(G||c||"Eval A","] ").concat(e.label||"")})),...l.head.prompts.map(t=>({...t,label:"[".concat(e,"] ").concat(t.label||"")}))],vars:m.head.vars},body:m.body.map((e,t)=>{var n;return{...e,outputs:[...e.outputs,...(null===(n=l.body[t])||void 0===n?void 0:n.outputs)||[]]}})};x(a),j({...v,...s.data.config,description:'Combined: "'.concat((null==v?void 0:v.description)||"Eval A",'" and "').concat((null===(t=s.data.config)||void 0===t?void 0:t.description)||"Eval B",'"')}),Y(!0)}catch(e){console.error("Error fetching comparison eval:",e),alert("Failed to load comparison eval. Please try again.")}},eS=l.useMemo(()=>m.body.some(e=>e.description),[m.body]),eE=ei.prompts.map((e,t)=>{let n=e.label||e.display||e.raw;return{value:"Prompt ".concat(t+1),label:"Prompt ".concat(t+1,": ").concat(n&&n.length>100?n.slice(0,100)+"...":n||""),group:"Prompts"}}),eR=l.useMemo(()=>[...eS?[{value:"description",label:"Description"}]:[],...ei.vars.map((e,t)=>({value:"Variable ".concat(t+1),label:"Var ".concat(t+1,": ").concat(ei.vars[t].length>100?ei.vars[t].slice(0,97)+"...":ei.vars[t]),group:"Variables"})),...eE],[ei.vars,eE,eS]),[eI,eN]=l.useState(!1),[eP,eF]=l.useState(!1),eT=l.useMemo(()=>[...eS?["description"]:[],...ei.vars.map((e,t)=>"Variable ".concat(t+1)),...ei.prompts.map((e,t)=>"Prompt ".concat(t+1))],[eS,ei.vars,ei.prompts]),eD=G||c||"default",eM=q[eD]||{selectedColumns:[],columnVisibility:{}},eL=l.useCallback(e=>{let t={};eT.forEach(n=>{t[n]=e.includes(n)}),$(eD,{selectedColumns:e,columnVisibility:t})},[eT,$,eD]);l.useEffect(()=>{0!==eM.selectedColumns.length&&eM.selectedColumns.every(e=>eT.includes(e))||eL(eT)},[eT,eM.selectedColumns,eL]);let eA=l.useCallback(e=>{let t=Array.isArray(e.target.value)?e.target.value:e.target.value.split(",");eL(t)},[eL]),eU=async()=>{(0,y.Z)(v,"Config must be loaded before clicking its description");let e=window.prompt("Enter new description:",v.description);if(null!==e&&e!==v.description){let t={...v,description:e};try{let e=await fetch("".concat(await (0,o.b)(),"/api/eval/").concat(G),{method:"PATCH",headers:{"Content-Type":"application/json"},body:JSON.stringify({config:t})});if(!e.ok)throw Error("Network response was not ok");j(t)}catch(e){console.error("Failed to update table:",e)}}},eW=async()=>{if(window.confirm("Are you sure you want to delete this evaluation?"))try{let e=await fetch("".concat(await (0,o.b)(),"/api/eval/").concat(G),{method:"DELETE"});if(!e.ok)throw Error("Network response was not ok");d.push("/")}catch(e){console.error("Failed to delete evaluation:",e),alert("Failed to delete evaluation")}},[ez,eV]=l.useState(!1),[e_,eB]=l.useState(!1),eJ=async()=>{G&&(await navigator.clipboard.writeText(G),eV(!0),setTimeout(()=>{eV(!1)},1e3))};return(0,s.jsxs)("div",{style:{marginLeft:"1rem",marginRight:"1rem"},children:[(0,s.jsxs)(tl,{direction:"row",mb:3,spacing:1,alignItems:"center",className:"eval-header",children:[(0,s.jsxs)(P.Z,{sx:{display:"flex",alignItems:"center",width:"100%",maxWidth:250},children:[(0,s.jsx)(J.Z,{variant:"outlined",size:"small",fullWidth:!0,value:(null==v?void 0:v.description)||G||"",InputProps:{readOnly:!0,startAdornment:(0,s.jsx)(O.Z,{position:"start",children:(0,s.jsx)(I.Z,{})}),endAdornment:(0,s.jsx)(O.Z,{position:"end",children:(0,s.jsx)(w.Z,{})})},onClick:()=>eB(!0),placeholder:"Search or select an eval...",sx:{cursor:"pointer"}}),(0,s.jsx)(ed,{open:e_,onClose:()=>eB(!1),recentEvals:n,onRecentEvalSelected:a,title:"Select an Eval"})]}),(null==v?void 0:v.description)&&G&&(0,s.jsxs)(s.Fragment,{children:[(0,s.jsx)(H.Z,{title:"Click to copy",children:(0,s.jsx)(D.Z,{size:"small",label:(0,s.jsxs)(s.Fragment,{children:[(0,s.jsx)("strong",{children:"ID:"})," ",G]}),sx:{opacity:.7,cursor:"pointer"},onClick:eJ})}),(0,s.jsx)(u.Z,{open:ez,autoHideDuration:1e3,onClose:()=>eV(!1),message:"Eval id copied to clipboard"})]}),(0,s.jsx)(H.Z,{title:p?"":"Set eval author with `promptfoo config set email`",children:(0,s.jsx)(D.Z,{size:"small",label:(0,s.jsxs)(s.Fragment,{children:[(0,s.jsx)("strong",{children:"Author:"})," ",p||"Unknown"]}),sx:{opacity:.7}})})]}),(0,s.jsxs)(tl,{direction:"row",spacing:1,alignItems:"center",children:[(0,s.jsx)(P.Z,{children:(0,s.jsxs)(M.Z,{sx:{minWidth:200,maxWidth:350},size:"small",children:[(0,s.jsx)(L.Z,{id:"visible-columns-label",children:"Columns"}),(0,s.jsx)(_.Z,{labelId:"visible-columns-label",id:"visible-columns",multiple:!0,value:eM.selectedColumns,onChange:eA,input:(0,s.jsx)(V.Z,{label:"Visible columns"}),renderValue:e=>e.join(", "),children:eR.map(e=>(0,s.jsxs)(z.Z,{dense:!0,value:e.value,children:[(0,s.jsx)(T.Z,{checked:eM.selectedColumns.includes(e.value)}),(0,s.jsx)(U.Z,{primary:e.label})]},e.value))})]})}),(0,s.jsx)(P.Z,{children:(0,s.jsxs)(M.Z,{sx:{minWidth:180},size:"small",children:[(0,s.jsx)(L.Z,{id:"failure-filter-mode-label",children:"Display"}),(0,s.jsxs)(_.Z,{labelId:"filter-mode-label",id:"filter-mode",value:eo,onChange:e=>{let t=e.target.value;ec(t);let n={};ei.prompts.forEach((e,s)=>{n["Prompt ".concat(s+1)]="failures"===t}),ea(n)},label:"Filter",children:[(0,s.jsx)(z.Z,{value:"all",children:"Show all results"}),(0,s.jsx)(z.Z,{value:"failures",children:"Show failures only"}),(0,s.jsx)(z.Z,{value:"different",children:"Show different only"}),(0,s.jsx)(z.Z,{value:"highlights",children:"Show highlights only"})]})]})}),(0,s.jsx)(P.Z,{children:(0,s.jsx)(J.Z,{sx:{minWidth:180},size:"small",label:"Search",placeholder:"Text or regex",value:ee,onChange:e=>es(e.target.value)})}),(0,s.jsx)(P.Z,{flexGrow:1}),(0,s.jsx)(P.Z,{display:"flex",justifyContent:"flex-end",children:(0,s.jsxs)(tl,{direction:"row",spacing:2,children:[(0,s.jsx)(F.Z,{color:"primary",onClick:e=>{eb(e.currentTarget)},startIcon:(0,s.jsx)(w.Z,{}),children:"Eval actions"}),v&&(0,s.jsxs)(W.Z,{id:"eval-actions-menu",anchorEl:ef,keepMounted:!0,open:!!ef,onClose:()=>{eb(null)},children:[(0,s.jsx)(H.Z,{title:"Edit the name of this eval",placement:"left",children:(0,s.jsxs)(z.Z,{onClick:eU,children:[(0,s.jsx)(A.Z,{children:(0,s.jsx)(S.Z,{fontSize:"small"})}),"Edit name"]})}),(0,s.jsx)(H.Z,{title:"Edit this eval in the web UI",placement:"left",children:(0,s.jsxs)(z.Z,{onClick:()=>{Q(v),d.push("/setup/")},children:[(0,s.jsx)(A.Z,{children:(0,s.jsx)(k.Z,{fontSize:"small"})}),"Edit and re-run"]})}),(0,s.jsx)(ev,{initialEvals:n,onComparisonEvalSelected:ew}),(0,s.jsx)(H.Z,{title:"View the configuration that defines this eval",placement:"left",children:(0,s.jsxs)(z.Z,{onClick:()=>eN(!0),children:[(0,s.jsx)(A.Z,{children:(0,s.jsx)(N.Z,{fontSize:"small"})}),"View YAML"]})}),(0,s.jsx)(eC,{}),(null==v?void 0:v.sharing)&&(0,s.jsx)(H.Z,{title:"Generate a unique URL that others can access",placement:"left",children:(0,s.jsxs)(z.Z,{onClick:eZ,disabled:eg,children:[(0,s.jsx)(A.Z,{children:eg?(0,s.jsx)(g.Z,{size:16}):(0,s.jsx)(R.Z,{fontSize:"small"})}),"Share"]})}),(0,s.jsx)(H.Z,{title:"Delete this eval",placement:"left",children:(0,s.jsxs)(z.Z,{onClick:eW,children:[(0,s.jsx)(A.Z,{children:(0,s.jsx)(C.Z,{fontSize:"small"})}),"Delete"]})})]}),(0,s.jsx)(H.Z,{title:"Edit table view settings",placement:"bottom",children:(0,s.jsx)(F.Z,{color:"primary",onClick:()=>eF(!0),startIcon:(0,s.jsx)(E.Z,{}),children:"Table Settings"})}),((null==v?void 0:v.redteam)||(null==v?void 0:null===(t=v.metadata)||void 0===t?void 0:t.redteam))&&(0,s.jsx)(H.Z,{title:"View vulnerability scan report",placement:"bottom",children:(0,s.jsx)(F.Z,{color:"primary",variant:"contained",startIcon:(0,s.jsx)(N.Z,{}),onClick:()=>d.push("/report/?evalId=".concat(G||c)),children:"Vulnerability Report"})})]})})]}),(0,s.jsx)(eO,{columnVisibility:eM.columnVisibility}),(0,s.jsx)(te,{maxTextLength:b,columnVisibility:eM.columnVisibility,wordBreak:B,showStats:K,filterMode:eo,failureFilter:el,searchText:en,onFailureFilterToggle:er,onSearchTextChange:es}),(0,s.jsx)(ey,{open:eI,onClose:()=>eN(!1)}),(0,s.jsx)(ts,{open:eu,onClose:()=>eh(!1),shareUrl:ep}),(0,s.jsx)(tn,{open:eP,onClose:()=>eF(!1)}),(0,s.jsx)(ek,{recentEvals:n,onRecentEvalSelected:a})]})}async function tr(){let e=(0,j.createClientComponentClient)(),{data:{user:t}}=await e.auth.getUser();(0,y.Z)(t,"User not logged in");let{data:n}=await e.from("EvaluationResult").select("id, createdAt").eq("user_id",t.id).order("createdAt",{ascending:!1}).limit(100);return n||[]}async function ti(e){let t=(0,j.createClientComponentClient)(),{data:n}=await t.from("EvaluationResult").select("*").eq("id",e).single();return n}function to(e){var t;let{fetchId:n,preloadedData:a,recentEvals:i,defaultEvalId:c}=e,u=(0,f.useRouter)(),{table:h,setTable:p,config:m,setConfig:j,evalId:Z,setEvalId:w,setAuthor:C,setInComparisonMode:S}=ex(),[k,E]=l.useState(!1),[R,I]=l.useState(!1),[N,P]=l.useState(i||[]),F=async()=>{let e=await fetch("".concat(await (0,o.b)(),"/api/results"),{cache:"no-store"}),t=await e.json();return P(t.data),t.data},T=l.useCallback(async e=>{let t=await fetch("".concat(await (0,o.b)(),"/api/results/").concat(e),{cache:"no-store"}),n=await t.json();p(n.data.results.table),j(n.data.config),C(n.data.author),w(e)},[p,j,w,C]),D=async e=>{v.Ox?(E(!1),u.push("/eval/remote:".concat(encodeURIComponent(e)))):u.push("/eval/?evalId=".concat(encodeURIComponent(e)))},[M,O]=l.useState(c||(null===(t=N[0])||void 0===t?void 0:t.evalId)),L=(0,f.useSearchParams)(),A=L?L.get("evalId"):null;return(l.useEffect(()=>{if(A){console.log("Eval init: Fetching eval by id",A);let e=async()=>{await T(A),E(!0),O(A),F()};e()}else if(a){var e;console.log("Eval init: Using preloaded data"),p(null===(e=a.data.results)||void 0===e?void 0:e.table),j(a.data.config),C(a.data.author||null),E(!0)}else if(n){console.log("Eval init: Fetching eval from remote server",n);let e=async()=>{var e;let t="".concat(r,"/api/eval/").concat(n);console.log("Fetching eval from remote server",t);let s=await fetch(t);if(!s.ok){I(!0);return}let l=await s.json();p(null===(e=l.data.results)||void 0===e?void 0:e.table),j(l.data.config),C(l.data.author||null),E(!0)};e()}else if(v.T8)console.log("Eval init: Using local server websocket"),(0,o.b)().then(e=>{let t=(0,b.io)(e);return t.on("init",e=>{console.log("Initialized socket connection",e),E(!0),p(null==e?void 0:e.results.table),j(null==e?void 0:e.config),C((null==e?void 0:e.author)||null),F().then(e=>{var t,n,s;O(null===(t=e[0])||void 0===t?void 0:t.evalId),console.log("setting default eval id",null===(n=e[0])||void 0===n?void 0:n.evalId),w(null===(s=e[0])||void 0===s?void 0:s.evalId)})}),t.on("update",e=>{console.log("Received data update",e),p(e.results.table),j(e.config),C(e.author||null),F().then(e=>{var t;let n=null===(t=e[0])||void 0===t?void 0:t.evalId;n&&(O(n),w(n))})}),()=>{t.disconnect()}});else if(v.Ox)console.log("Eval init: Using Supabase"),tr().then(e=>{P(e.map(e=>({evalId:e.id,datasetId:null,label:e.createdAt,createdAt:new Date(e.createdAt).getTime(),description:"None",numTests:-1}))),e.length>0&&ti(e[0].id).then(t=>{(0,y.Z)(t,"Eval not found");let n=t.results,s=t.config;O(e[0].id),p(n.table),j(s),C(null),E(!0)})});else{console.log("Eval init: Fetching eval via recent");let e=async()=>{let e=await F();if(!(e.length>0))return(0,s.jsx)("div",{className:"notice",children:"No evals yet. Share some evals to this server and they will appear here."});{let t=await (0,o.b)(),n=e[0].evalId,s=await fetch("".concat(t,"/api/results/").concat(n)),l=await s.json();p(l.data.results.table),j(l.data.config),C(l.data.author||null),E(!0),O(n),w(n)}};e()}S(!1)},[n,p,j,C,w,T,a,O,A,S]),l.useEffect(()=>{document.title="".concat((null==m?void 0:m.description)||Z||"Eval"," | promptfoo")},[m,Z]),R)?(0,s.jsx)("div",{className:"notice",children:"404 Eval not found"}):k&&h?(0,s.jsx)(x,{children:(0,s.jsx)(d,{children:(0,s.jsx)(ta,{defaultEvalId:M,recentEvals:N,onRecentEvalSelected:D})})}):(0,s.jsxs)("div",{className:"notice",children:[(0,s.jsx)("div",{children:(0,s.jsx)(g.Z,{size:22})}),(0,s.jsx)("div",{children:"Waiting for eval data"})]})}n(94455)},52428:function(e,t,n){"use strict";n.d(t,{Ox:function(){return a},T8:function(){return l},eA:function(){return r}});var s=n(77580);let l=!s.env.NEXT_PUBLIC_PROMPTFOO_BUILD_STANDALONE_SERVER,a=!!s.env.NEXT_PUBLIC_PROMPTFOO_USE_SUPABASE,r=""},47887:function(e,t,n){"use strict";n.d(t,{o:function(){return a}});var s=n(94660),l=n(74810);let a=(0,s.Ue)()((0,l.tJ)((e,t)=>({env:{},testCases:[],description:"",providers:[],prompts:[],defaultTest:{},evaluateOptions:{},scenarios:[],setEnv:t=>e({env:t}),setTestCases:t=>e({testCases:t}),setDescription:t=>e({description:t}),setProviders:t=>e({providers:t}),setPrompts:t=>e({prompts:t}),setDefaultTest:t=>e({defaultTest:t}),setEvaluateOptions:t=>e({evaluateOptions:t}),setScenarios:t=>e({scenarios:t}),setStateFromConfig:t=>{let n={};t.description&&(n.description=t.description||""),t.tests&&(n.testCases=t.tests),t.providers&&(n.providers=t.providers),t.prompts&&("string"==typeof t.prompts?n.prompts=[t.prompts]:Array.isArray(t.prompts)?n.prompts=t.prompts.filter(e=>"string"==typeof e&&!e.endsWith(".txt")&&!e.endsWith(".json")&&!e.endsWith(".yaml")):console.warn("Invalid prompts config",t.prompts)),t.defaultTest&&(n.defaultTest=t.defaultTest),t.evaluateOptions&&(n.evaluateOptions=t.evaluateOptions),t.scenarios&&(n.scenarios=t.scenarios),e(n)},getTestSuite:()=>{let{description:e,testCases:n,providers:s,prompts:l,env:a,scenarios:r}=t();return{env:a,description:e,providers:s,prompts:l,tests:n,scenarios:r}}}),{name:"promptfoo",skipHydration:!0}))},55974:function(){},94455:function(){},16658:function(){},93207:function(){},58022:function(){}}]);

package/dist/src/web/nextui/_next/static/chunks/app/progress/page-e0bd2de55fe3a3b6.js

@@ -1 +0,0 @@
-(self.webpackChunk_N_E=self.webpackChunk_N_E||[]).push([[92],{64738:function(e,t,i){Promise.resolve().then(i.bind(i,19259))},19259:function(e,t,i){"use strict";i.r(t),i.d(t,{default:function(){return j}});var a=i(24004),r=i(14978),n=i(52481),s=i(26972),o=i(79245),l=i(49050),d=i(33932),c=i(39114),u=i(7575),p=i(73701),v=i(39279),m=i(30666),h=i(15795),x=i(66988),Z=i(98489),f=i(95781),g=i(21975),C=i(38640),I=i.n(C);function j(){let[e,t]=(0,r.useState)([]),[i,C]=(0,r.useState)(null),[j,b]=(0,r.useState)("asc"),[y,P]=(0,r.useState)(null),[w,k]=r.useState(1),[M,F]=(0,r.useState)({evalId:"",datasetId:"",provider:"",promptId:""}),S=!!y;(0,r.useEffect)(()=>{(async()=>{let e=await fetch("/api/progress"),i=await e.json();i&&i.data&&t(i.data)})()},[]);let $=e=>{let t=i===e&&"asc"===j;C(e),b(t?"desc":"asc")},O=e=>(null==e?void 0:e.testPassCount)!=null&&(null==e?void 0:e.testFailCount)!=null?(e.testPassCount/(e.testPassCount+e.testFailCount)*100).toFixed(2):"-",R=e=>{var t,i,a,r,n;let s=e.map(e=>{var s,o,l,d,c,u;return[null!==(t=e.evalId)&&void 0!==t?t:"",null!==(i=null===(s=e.datasetId)||void 0===s?void 0:s.slice(0,6))&&void 0!==i?i:"",null!==(a=e.provider)&&void 0!==a?a:"",(null!==(r=null===(o=e.promptId)||void 0===o?void 0:o.slice(0,6))&&void 0!==r?r:"")+" "+(null!==(n=e.raw)&&void 0!==n?n:""),O(e.metrics),(null===(l=e.metrics)||void 0===l?void 0:l.testPassCount)==null?"-":"".concat(e.metrics.testPassCount),(null===(d=e.metrics)||void 0===d?void 0:d.testFailCount)==null?"-":"".concat(e.metrics.testFailCount),(null===(c=e.metrics)||void 0===c?void 0:c.score)==null?"-":null===(u=e.metrics.score)||void 0===u?void 0:u.toFixed(2)]});return[["Eval","Dataset","Provider","Prompt","Pass Rate %","Pass Count","Fail Count","Raw score"]].concat(s).map(e=>e.map(e=>null!=e?e:"").join(",")).join("\n")},L=t=>{let i="json"===t?JSON.stringify(e):R(e),a=new Blob([i],{type:"text/".concat(t,";charset=utf-8;")}),r=document.createElement("a");r.href=URL.createObjectURL(a),r.download="cols_export.".concat(t),r.click(),URL.revokeObjectURL(r.href),P(null)},D=r.useMemo(()=>e.filter(e=>{var t,i,a,r;return(!M.evalId||(null===(t=e.evalId)||void 0===t?void 0:t.includes(M.evalId)))&&(!M.datasetId||(null===(i=e.datasetId)||void 0===i?void 0:i.startsWith(M.datasetId)))&&(!M.provider||(null===(a=e.provider)||void 0===a?void 0:a.includes(M.provider)))&&(!M.promptId||(null===(r=e.promptId)||void 0===r?void 0:r.startsWith(M.promptId)))}),[e,M]),E=r.useMemo(()=>D.sort((e,t)=>{if(!i)return 0;if("passRate"===i){let i=parseFloat(O(e.metrics)),a=parseFloat(O(t.metrics));return"asc"===j?i-a:a-i}if(i in e&&i in t){let a=e[i]||"",r=t[i]||"";return"asc"===j?a.toString().localeCompare(r.toString()):r.toString().localeCompare(a.toString())}return 0}),[D,i,j]),V=r.useMemo(()=>Array.from(new Set(e.map(e=>e.evalId))),[e]),z=r.useMemo(()=>Array.from(new Set(e.map(e=>e.datasetId))),[e]),N=r.useMemo(()=>Array.from(new Set(e.map(e=>e.provider))),[e]),T=r.useMemo(()=>Array.from(new Set(e.map(e=>e.promptId))),[e]);return(0,a.jsxs)(o.Z,{paddingX:2,children:[(0,a.jsxs)(o.Z,{display:"flex",justifyContent:"space-between",alignItems:"center",children:[(0,a.jsx)("h2",{children:"Progress summary"}),(0,a.jsxs)("div",{children:[(0,a.jsx)(l.Z,{id:"export-button","aria-controls":S?"export-menu":void 0,"aria-haspopup":"true","aria-expanded":S?"true":void 0,onClick:e=>{P(e.currentTarget)},startIcon:(0,a.jsx)(n.Z,{}),children:"Export"}),(0,a.jsxs)(d.Z,{id:"export-menu",anchorEl:y,open:S,onClose:()=>{P(null)},MenuListProps:{"aria-labelledby":"export-button"},children:[(0,a.jsx)(c.Z,{onClick:()=>L("csv"),children:"CSV"}),(0,a.jsx)(c.Z,{onClick:()=>L("json"),children:"JSON"})]})]})]}),(0,a.jsx)(o.Z,{children:"This page shows performance metrics for recent evals."}),(0,a.jsxs)(o.Z,{display:"flex",flexDirection:"row",gap:2,my:2,children:[(0,a.jsx)(s.Z,{options:V,value:M.evalId,onChange:(e,t)=>{F({...M,evalId:t||""})},renderInput:e=>(0,a.jsx)(g.Z,{...e,label:"Eval ID",variant:"outlined",size:"small",fullWidth:!0}),sx:{width:220}}),(0,a.jsx)(s.Z,{options:z,value:M.datasetId,onChange:(e,t)=>{F({...M,datasetId:t||""})},renderInput:e=>(0,a.jsx)(g.Z,{...e,label:"Dataset ID",variant:"outlined",size:"small",fullWidth:!0}),sx:{width:220}}),(0,a.jsx)(s.Z,{options:N,value:M.provider,onChange:(e,t)=>{F({...M,provider:t||""})},renderInput:e=>(0,a.jsx)(g.Z,{...e,label:"Provider",variant:"outlined",size:"small",fullWidth:!0}),sx:{width:220}}),(0,a.jsx)(s.Z,{options:T,value:M.promptId,onChange:(e,t)=>{F({...M,promptId:t||""})},renderInput:e=>(0,a.jsx)(g.Z,{...e,label:"Prompt ID",variant:"outlined",size:"small",fullWidth:!0}),sx:{width:220}})]}),(0,a.jsxs)(h.Z,{children:[(0,a.jsxs)(p.Z,{children:[(0,a.jsx)(x.Z,{children:(0,a.jsxs)(Z.Z,{children:[(0,a.jsx)(m.Z,{children:(0,a.jsx)(f.Z,{active:"evalId"===i,direction:"evalId"===i?j:"asc",onClick:()=>$("evalId"),children:"Eval"})}),(0,a.jsx)(m.Z,{children:"Dataset"}),(0,a.jsx)(m.Z,{children:"Provider"}),(0,a.jsx)(m.Z,{children:"Prompt"}),(0,a.jsx)(m.Z,{children:(0,a.jsx)(f.Z,{active:"passRate"===i,direction:"passRate"===i?j:"asc",onClick:()=>$("passRate"),children:"Pass Rate %"})}),(0,a.jsx)(m.Z,{children:(0,a.jsx)(f.Z,{active:"testPassCount"===i,direction:"testPassCount"===i?j:"asc",onClick:()=>$("testPassCount"),children:"Pass Count"})}),(0,a.jsx)(m.Z,{children:(0,a.jsx)(f.Z,{active:"testFailCount"===i,direction:"testFailCount"===i?j:"asc",onClick:()=>$("testFailCount"),children:"Fail Count"})}),(0,a.jsx)(m.Z,{children:(0,a.jsx)(f.Z,{active:"score"===i,direction:"score"===i?j:"asc",onClick:()=>$("score"),children:"Raw score"})})]})}),(0,a.jsx)(v.Z,{children:E.slice((w-1)*25,25*w).map((e,t)=>{var i,r,n,s,o,l;return(0,a.jsxs)(Z.Z,{hover:!0,onClick:()=>F({...M,evalId:e.evalId,datasetId:e.datasetId||"",promptId:e.promptId||"",provider:e.provider}),children:[(0,a.jsx)(m.Z,{children:(0,a.jsx)(I(),{href:"/eval?evalId=".concat(e.evalId),onClick:e=>e.stopPropagation(),children:e.evalId})}),(0,a.jsx)(m.Z,{children:(0,a.jsx)(I(),{href:"/datasets?id=".concat(e.datasetId),onClick:e=>e.stopPropagation(),children:null===(i=e.datasetId)||void 0===i?void 0:i.slice(0,6)})}),(0,a.jsx)(m.Z,{children:e.provider}),(0,a.jsxs)(m.Z,{children:[(0,a.jsxs)(I(),{href:"/prompts?id=".concat(e.promptId),onClick:e=>e.stopPropagation(),children:["[",null===(r=e.promptId)||void 0===r?void 0:r.slice(0,6),"]"]})," ",e.raw]}),(0,a.jsx)(m.Z,{children:O(e.metrics)}),(0,a.jsx)(m.Z,{children:(null===(n=e.metrics)||void 0===n?void 0:n.testPassCount)==null?"-":"".concat(e.metrics.testPassCount)}),(0,a.jsx)(m.Z,{children:(null===(s=e.metrics)||void 0===s?void 0:s.testFailCount)==null?"-":"".concat(e.metrics.testFailCount)}),(0,a.jsx)(m.Z,{children:(null===(o=e.metrics)||void 0===o?void 0:o.score)==null?"-":null===(l=e.metrics.score)||void 0===l?void 0:l.toFixed(2)})]},t)})})]}),Math.ceil(D.length/25)>1&&(0,a.jsx)(u.Z,{count:Math.ceil(E.length/25),page:w,onChange:(e,t)=>k(t),sx:{pt:2,pb:4,display:"flex",justifyContent:"center"}})]})]})}},52481:function(e,t,i){"use strict";var a=i(71851);t.Z=void 0;var r=a(i(80984)),n=i(24004);t.Z=(0,r.default)((0,n.jsx)("path",{d:"M5 20h14v-2H5zM19 9h-4V3H9v6H5l7 7z"}),"Download")},78342:function(e,t,i){"use strict";i.d(t,{f:function(){return n}});var a=i(26520),r=i(25702);function n(e){return(0,r.ZP)("MuiListItemIcon",e)}let s=(0,a.Z)("MuiListItemIcon",["root","alignItemsFlexStart"]);t.Z=s},69660:function(e,t,i){"use strict";i.d(t,{L:function(){return n}});var a=i(26520),r=i(25702);function n(e){return(0,r.ZP)("MuiListItemText",e)}let s=(0,a.Z)("MuiListItemText",["root","multiline","dense","inset","primary","secondary"]);t.Z=s},39114:function(e,t,i){"use strict";i.d(t,{Z:function(){return M}});var a=i(94312),r=i(79042),n=i(14978),s=i(53468),o=i(95600),l=i(24871),d=i(35843),c=i(22630),u=i(19996),p=i(77820),v=i(84008),m=i(88519),h=i(37663),x=i(26520);let Z=(0,x.Z)("MuiDivider",["root","absolute","fullWidth","inset","middle","flexItem","light","vertical","withChildren","withChildrenVertical","textAlignRight","textAlignLeft","wrapper","wrapperVertical"]);var f=i(78342),g=i(69660),C=i(25702);function I(e){return(0,C.ZP)("MuiMenuItem",e)}let j=(0,x.Z)("MuiMenuItem",["root","focusVisible","dense","disabled","divider","gutters","selected"]);var b=i(24004);let y=["autoFocus","component","dense","divider","disableGutters","focusVisibleClassName","role","tabIndex","className"],P=e=>{let{disabled:t,dense:i,divider:a,disableGutters:n,selected:s,classes:l}=e,d=(0,o.Z)({root:["root",i&&"dense",t&&"disabled",!n&&"gutters",a&&"divider",s&&"selected"]},I,l);return(0,r.Z)({},l,d)},w=(0,d.ZP)(v.Z,{shouldForwardProp:e=>(0,c.Z)(e)||"classes"===e,name:"MuiMenuItem",slot:"Root",overridesResolver:(e,t)=>{let{ownerState:i}=e;return[t.root,i.dense&&t.dense,i.divider&&t.divider,!i.disableGutters&&t.gutters]}})(({theme:e,ownerState:t})=>(0,r.Z)({},e.typography.body1,{display:"flex",justifyContent:"flex-start",alignItems:"center",position:"relative",textDecoration:"none",minHeight:48,paddingTop:6,paddingBottom:6,boxSizing:"border-box",whiteSpace:"nowrap"},!t.disableGutters&&{paddingLeft:16,paddingRight:16},t.divider&&{borderBottom:`1px solid ${(e.vars||e).palette.divider}`,backgroundClip:"padding-box"},{"&:hover":{textDecoration:"none",backgroundColor:(e.vars||e).palette.action.hover,"@media (hover: none)":{backgroundColor:"transparent"}},[`&.${j.selected}`]:{backgroundColor:e.vars?`rgba(${e.vars.palette.primary.mainChannel} / ${e.vars.palette.action.selectedOpacity})`:(0,l.Fq)(e.palette.primary.main,e.palette.action.selectedOpacity),[`&.${j.focusVisible}`]:{backgroundColor:e.vars?`rgba(${e.vars.palette.primary.mainChannel} / calc(${e.vars.palette.action.selectedOpacity} + ${e.vars.palette.action.focusOpacity}))`:(0,l.Fq)(e.palette.primary.main,e.palette.action.selectedOpacity+e.palette.action.focusOpacity)}},[`&.${j.selected}:hover`]:{backgroundColor:e.vars?`rgba(${e.vars.palette.primary.mainChannel} / calc(${e.vars.palette.action.selectedOpacity} + ${e.vars.palette.action.hoverOpacity}))`:(0,l.Fq)(e.palette.primary.main,e.palette.action.selectedOpacity+e.palette.action.hoverOpacity),"@media (hover: none)":{backgroundColor:e.vars?`rgba(${e.vars.palette.primary.mainChannel} / ${e.vars.palette.action.selectedOpacity})`:(0,l.Fq)(e.palette.primary.main,e.palette.action.selectedOpacity)}},[`&.${j.focusVisible}`]:{backgroundColor:(e.vars||e).palette.action.focus},[`&.${j.disabled}`]:{opacity:(e.vars||e).palette.action.disabledOpacity},[`& + .${Z.root}`]:{marginTop:e.spacing(1),marginBottom:e.spacing(1)},[`& + .${Z.inset}`]:{marginLeft:52},[`& .${g.Z.root}`]:{marginTop:0,marginBottom:0},[`& .${g.Z.inset}`]:{paddingLeft:36},[`& .${f.Z.root}`]:{minWidth:36}},!t.dense&&{[e.breakpoints.up("sm")]:{minHeight:"auto"}},t.dense&&(0,r.Z)({minHeight:32,paddingTop:4,paddingBottom:4},e.typography.body2,{[`& .${f.Z.root} svg`]:{fontSize:"1.25rem"}}))),k=n.forwardRef(function(e,t){let i;let o=(0,u.i)({props:e,name:"MuiMenuItem"}),{autoFocus:l=!1,component:d="li",dense:c=!1,divider:v=!1,disableGutters:x=!1,focusVisibleClassName:Z,role:f="menuitem",tabIndex:g,className:C}=o,I=(0,a.Z)(o,y),j=n.useContext(p.Z),k=n.useMemo(()=>({dense:c||j.dense||!1,disableGutters:x}),[j.dense,c,x]),M=n.useRef(null);(0,m.Z)(()=>{l&&M.current&&M.current.focus()},[l]);let F=(0,r.Z)({},o,{dense:k.dense,divider:v,disableGutters:x}),S=P(o),$=(0,h.Z)(M,t);return o.disabled||(i=void 0!==g?g:-1),(0,b.jsx)(p.Z.Provider,{value:k,children:(0,b.jsx)(w,(0,r.Z)({ref:$,role:f,tabIndex:i,component:d,focusVisibleClassName:(0,s.Z)(S.focusVisible,Z),className:(0,s.Z)(S.root,C)},I,{ownerState:F,classes:S}))})});var M=k},83691:function(e,t,i){"use strict";i(14978);var a=i(59782),r=i(24004);t.Z=(0,a.Z)((0,r.jsx)("path",{d:"M19 6.41L17.59 5 12 10.59 6.41 5 5 6.41 10.59 12 5 17.59 6.41 19 12 13.41 17.59 19 19 17.59 13.41 12z"}),"Close")}},function(e){e.O(0,[148,640,193,391,654,975,874,854,972,470,730,744],function(){return e(e.s=64738)}),_N_E=e.O()}]);

package/dist/src/web/nextui/_next/static/chunks/app/setup/page-0203187deebc18e8.js

@@ -1 +0,0 @@
-(self.webpackChunk_N_E=self.webpackChunk_N_E||[]).push([[413],{8458:function(e,t,n){Promise.resolve().then(n.bind(n,83013))},83013:function(e,t,n){"use strict";n.r(t),n.d(t,{default:function(){return eu}});var a=n(24004),r=n(14978),i=n(69179),l=n(47887),s=n(79245),o=n(49050),c=n(88938),d=n(89394),p=n(42834),u=n(26337),h=n(64173),x=n(91797),m=n(13457),f=n(43226),j=n(93295),v=n(82057),g=n(58768),Z=n(15873),b=n(21975),y=()=>{let{env:e,setEnv:t}=(0,l.o)(),[n,i]=(0,r.useState)(!1),[s,c]=(0,r.useState)(e),h=()=>{i(!1)};return(0,a.jsxs)(a.Fragment,{children:[(0,a.jsx)(o.Z,{variant:"outlined",startIcon:(0,a.jsx)(j.Z,{}),onClick:()=>{i(!0)},children:"API keys"}),(0,a.jsxs)(d.Z,{open:n,onClose:h,fullWidth:!0,maxWidth:"md",children:[(0,a.jsx)(x.Z,{children:"Provider settings"}),(0,a.jsxs)(u.Z,{children:[(0,a.jsxs)(v.Z,{defaultExpanded:!0,children:[(0,a.jsx)(Z.Z,{children:"OpenAI"}),(0,a.jsxs)(g.Z,{children:[(0,a.jsx)(b.Z,{label:"OpenAI API key",fullWidth:!0,margin:"normal",value:s.OPENAI_API_KEY,onChange:e=>c({...s,OPENAI_API_KEY:e.target.value})}),(0,a.jsx)(b.Z,{label:"OpenAI API host",fullWidth:!0,margin:"normal",value:s.OPENAI_API_HOST,onChange:e=>c({...s,OPENAI_API_HOST:e.target.value})}),(0,a.jsx)(b.Z,{label:"OpenAI organization",fullWidth:!0,margin:"normal",value:s.OPENAI_ORGANIZATION,onChange:e=>c({...s,OPENAI_ORGANIZATION:e.target.value})})]})]}),(0,a.jsxs)(v.Z,{children:[(0,a.jsx)(Z.Z,{children:"Azure"}),(0,a.jsx)(g.Z,{children:(0,a.jsx)(b.Z,{label:"Azure API key",fullWidth:!0,margin:"normal",value:s.AZURE_OPENAI_API_KEY,onChange:e=>c({...s,AZURE_OPENAI_API_KEY:e.target.value})})})]}),(0,a.jsxs)(v.Z,{children:[(0,a.jsx)(Z.Z,{children:"Amazon Bedrock"}),(0,a.jsx)(g.Z,{children:(0,a.jsx)(b.Z,{label:"Bedrock Region",fullWidth:!0,margin:"normal",value:s.AWS_BEDROCK_REGION,onChange:e=>c({...s,AWS_BEDROCK_REGION:e.target.value})})})]}),(0,a.jsxs)(v.Z,{children:[(0,a.jsx)(Z.Z,{children:"Anthropic"}),(0,a.jsx)(g.Z,{children:(0,a.jsx)(b.Z,{label:"Anthropic API key",fullWidth:!0,margin:"normal",value:s.ANTHROPIC_API_KEY,onChange:e=>c({...s,ANTHROPIC_API_KEY:e.target.value})})})]}),(0,a.jsxs)(v.Z,{children:[(0,a.jsx)(Z.Z,{children:"Google Vertex AI"}),(0,a.jsx)(g.Z,{children:(0,a.jsx)(b.Z,{label:"Vertex API Key",fullWidth:!0,margin:"normal",value:s.VERTEX_API_KEY,onChange:e=>c({...s,VERTEX_API_KEY:e.target.value})})}),(0,a.jsx)(g.Z,{children:(0,a.jsx)(b.Z,{label:"Vertex Project ID",fullWidth:!0,margin:"normal",value:s.VERTEX_PROJECT_ID,onChange:e=>c({...s,VERTEX_PROJECT_ID:e.target.value})})}),(0,a.jsx)(g.Z,{children:(0,a.jsx)(b.Z,{label:"Vertex Region",fullWidth:!0,margin:"normal",value:s.VERTEX_REGION,onChange:e=>c({...s,VERTEX_REGION:e.target.value})})})]}),(0,a.jsxs)(v.Z,{children:[(0,a.jsx)(Z.Z,{children:"Replicate"}),(0,a.jsx)(g.Z,{children:(0,a.jsx)(b.Z,{label:"Replicate API key",fullWidth:!0,margin:"normal",value:s.REPLICATE_API_KEY,onChange:e=>c({...s,REPLICATE_API_KEY:e.target.value})})})]})]}),(0,a.jsxs)(p.Z,{children:[(0,a.jsx)(o.Z,{onClick:h,color:"primary",children:"Cancel"}),(0,a.jsx)(o.Z,{onClick:()=>{t(s),h()},color:"primary",variant:"contained",children:"Save"})]})]})]})},_=n(39329),C=n(46446),A=n(13391),k=n(71280),E=n(52653),I=n(73701),O=n(39279),S=n(30666),P=n(15795),T=n(98489),w=n(78276),R=e=>{let{open:t,prompt:n,index:i,onAdd:l,onCancel:s}=e,[c,h]=r.useState(n),m=r.useRef(null);r.useEffect(()=>{h(n)},[n]);let f=e=>{l(c),h(""),e?s():m.current&&m.current.focus()};return(0,a.jsxs)(d.Z,{open:t,onClose:s,fullWidth:!0,maxWidth:"md",children:[(0,a.jsx)(x.Z,{children:"Edit Prompt ".concat(i+1)}),(0,a.jsx)(u.Z,{children:(0,a.jsx)(b.Z,{value:c,onChange:e=>h(e.target.value),fullWidth:!0,margin:"normal",multiline:!0,placeholder:"The quick brown {{animal1}} jumps over the lazy {{animal2}}.",helperText:"Tip: use the {{varname}} syntax to add variables to your prompt.",inputRef:m})}),(0,a.jsxs)(p.Z,{children:[(0,a.jsx)(o.Z,{onClick:f.bind(null,!0),color:"primary",variant:"contained",disabled:!c.length,children:"Add"}),(0,a.jsx)(o.Z,{onClick:f.bind(null,!1),color:"primary",variant:"contained",disabled:!c.length,children:"Add Another"}),(0,a.jsx)(o.Z,{onClick:s,color:"secondary",children:"Cancel"})]})]})};n(94499);var W=()=>{let[e,t]=(0,r.useState)(!1),[n,i]=(0,r.useState)(null),{prompts:s,setPrompts:c}=(0,l.o)(),d=(0,r.useRef)(null);(0,r.useEffect)(()=>{null!==n&&n>0&&d.current&&d.current.focus()},[n]);let p=e=>{i(e),t(!0)},u=(e,t)=>{e.stopPropagation();let n=s[t];c([...s,n])},h=(e,t)=>{c(s.map((n,a)=>a===e?t:n))},x=(e,t)=>{e.stopPropagation(),confirm("Are you sure you want to remove this prompt?")&&c(s.filter((e,n)=>n!==t))};return(0,a.jsxs)("div",{children:[(0,a.jsxs)(m.Z,{direction:"row",spacing:2,mb:2,justifyContent:"space-between",children:[(0,a.jsx)(f.Z,{variant:"h5",children:"Prompts"}),(0,a.jsxs)("div",{children:[(0,a.jsx)("label",{htmlFor:"file-input-add-prompt",children:(0,a.jsx)(w.Z,{title:"Upload prompt from file",children:(0,a.jsxs)("span",{children:[(0,a.jsx)(E.Z,{component:"span",children:(0,a.jsx)(k.Z,{})}),(0,a.jsx)("input",{id:"file-input-add-prompt",type:"file",accept:".txt,.md",onChange:e=>{var t;e.stopPropagation(),e.preventDefault();let n=null===(t=e.target.files)||void 0===t?void 0:t[0];if(n){let e=new FileReader;e.onload=e=>{var t,n;let a=null===(n=e.target)||void 0===n?void 0:null===(t=n.result)||void 0===t?void 0:t.toString();a&&c([...s,a])},e.readAsText(n)}},style:{display:"none"}})]})})}),(0,a.jsx)(o.Z,{color:"primary",onClick:()=>{t(!0)},variant:"contained",children:"Add Prompt"})]})]}),(0,a.jsx)(P.Z,{children:(0,a.jsx)(I.Z,{children:(0,a.jsx)(O.Z,{children:0===s.length?(0,a.jsx)(T.Z,{children:(0,a.jsx)(S.Z,{colSpan:2,align:"center",children:"No prompts added yet."})}):s.map((e,t)=>(0,a.jsxs)(T.Z,{sx:{"&:hover":{backgroundColor:"rgba(0, 0, 0, 0.04)",cursor:"pointer"}},onClick:()=>p(t),children:[(0,a.jsx)(S.Z,{children:(0,a.jsxs)(f.Z,{variant:"body2",children:["Prompt #".concat(t+1,": "),(e.length>250?e.slice(0,250)+" ...":e).split(/({{\w+}})/g).map((e,t)=>/{{\s*(\w+)\s*}}/g.test(e)?(0,a.jsx)("span",{className:"prompt-var-highlight",children:e},t):e)]})}),(0,a.jsxs)(S.Z,{align:"right",sx:{minWidth:150},children:[(0,a.jsx)(E.Z,{onClick:()=>p(t),size:"small",children:(0,a.jsx)(A.Z,{})}),(0,a.jsx)(E.Z,{onClick:e=>u(e,t),size:"small",children:(0,a.jsx)(_.Z,{})}),(0,a.jsx)(E.Z,{onClick:e=>x(e,t),size:"small",children:(0,a.jsx)(C.Z,{})})]})]},t))})})}),(0,a.jsx)(R,{open:e,prompt:null!==n?s[n]:"",index:null!==n?n:0,onAdd:e=>{null!==n?h(n,e):c([...s,e]),i(null)},onCancel:()=>{i(null),t(!1)}})]})},N=n(26972),z=n(47827),F=e=>{let{onChange:t,...n}=e,[i,l]=r.useState(""),[s,o]=r.useState(!1);return(0,a.jsx)(b.Z,{...n,error:s,helperText:s?"Invalid JSON":"",value:i,onChange:e=>{let n=e.target.value;try{let e=JSON.parse(n);l(n),o(!1),t&&t(e)}catch(e){l(n),o(!0)}}})},V=e=>{let{open:t,providerId:n,config:i,onClose:l,onSave:c}=e,[h,m]=r.useState(i);return r.useEffect(()=>{m(i)},[i]),(0,a.jsxs)(d.Z,{open:t,onClose:l,fullWidth:!0,maxWidth:"md",children:[(0,a.jsxs)(x.Z,{children:["Edit ",n.length>50?n.slice(0,50)+"...":n]}),(0,a.jsx)(u.Z,{children:Object.keys(h).map(e=>{let t;let n=h[e];return"number"==typeof n||"boolean"==typeof n||"string"==typeof n?(t="number"==typeof n?t=>m({...h,[e]:parseFloat(t.target.value)}):"boolean"==typeof n?t=>m({...h,[e]:"true"===t.target.value}):t=>{let n=t.target.value.trim();if(n.startsWith("{")||n.startsWith("["))try{m({...h,[e]:JSON.parse(n)})}catch(t){m({...h,[e]:n})}else"null"===n?m({...h,[e]:null}):"undefined"===n?m({...h,[e]:void 0}):m({...h,[e]:n})},(0,a.jsx)(s.Z,{my:2,children:(0,a.jsx)(b.Z,{label:e,value:n,onChange:t,fullWidth:!0,InputLabelProps:{shrink:!0},type:"number"==typeof n?"number":"text"})},e)):(0,a.jsx)(s.Z,{my:2,children:(0,a.jsx)(F,{label:e,defaultValue:JSON.stringify(n),onChange:t=>{m({...h,[e]:t})},fullWidth:!0,multiline:!0,minRows:2,InputLabelProps:{shrink:!0}})},e)})}),(0,a.jsxs)(p.Z,{children:[(0,a.jsx)(o.Z,{onClick:l,children:"Cancel"}),(0,a.jsx)(o.Z,{onClick:()=>{c(h)},children:"Save"})]})]})};let L=[].concat(["openai:gpt-4o","openai:gpt-4o-2024-05-13","openai:gpt-4o-mini","openai:gpt-4o-mini-2024-07-18","openai:gpt-4-turbo","openai:gpt-4","openai:gpt-3.5-turbo"].map(e=>({id:e,config:{organization:"",temperature:.5,max_tokens:1024,top_p:1,frequency_penalty:0,presence_penalty:0,function_call:void 0,functions:void 0,stop:void 0}}))).concat(["anthropic:messages:claude-3-5-sonnet-20240620","anthropic:messages:claude-3-haiku-20240307","anthropic:messages:claude-3-sonnet-20240229","anthropic:messages:claude-3-opus-20240229"].map(e=>({id:e,config:{max_tokens:1024,temperature:.5}}))).concat(["bedrock:anthropic.claude-3-haiku-20240307-v1:0","bedrock:anthropic.claude-3-sonnet-20240229-v1:0","bedrock:anthropic.claude-3-opus-20240307-v1:0"].map(e=>({id:e,config:{max_tokens_to_sample:256,temperature:.5}}))).concat(["azureopenai:gpt-4o-2024-05-13","azureopenai:gpt-4-turbo-2024-04-09","azureopenai:gpt-4-32k-0613","azureopenai:gpt-4-0613","azureopenai:gpt-35-turbo-0125"].map(e=>({id:e,config:{temperature:.5,max_tokens:1024,top_p:1,frequency_penalty:0,presence_penalty:0,function_call:void 0,functions:void 0,stop:void 0}}))).concat(["vertex:chat-bison@001","vertex:chat-bison","vertex:chat-bison-32k","vertex:chat-bison-32k@001"].map(e=>({id:e,config:{context:void 0,examples:void 0,temperature:0,maxOutputTokens:1024,topP:.95,topK:40,safetySettings:void 0,stopSequence:void 0}}))).concat(["replicate:replicate/flan-t5-small:69716ad8c34274043bf4a135b7315c7c569ec931d8f23d6826e249e1c142a264"].map(e=>({id:e,config:{temperature:.5,max_length:1024,repetition_penality:1}}))).concat(["replicate:replicate/codellama-7b-instruct:0103579e86fc75ba0d65912890fa19ef03c84a68554635319accf2e0ba93d3ae","replicate:replicate/codellama-13b-instruct:da5676342de1a5a335b848383af297f592b816b950a43d251a0a9edd0113604b","replicate:replicate/llama-2-70b-chat:2796ee9483c3fd7aa2e171d38f4ca12251a30609463dcfd4cd76703f22e96cdf"].map(e=>({id:e,config:{system_prompt:"",temperature:.75,top_p:.9,top_k:50,max_new_tokens:128,min_new_tokens:-1}}))).concat(["replicate:replicate/codellama-7b:6880b103613a9cd23950c5fd6c140197e519905bd0dd00e448c4858bdd06090a","replicate:replicate/codellama-13b-python:09b87c02dfa403e0c3289166dece62286b3bce49bae39a9c9204713cf94b8b7d","replicate:replicate/codellama-13b:1c914d844307b0588599b8393480a3ba917b660c7e9dfae681542b5325f228db","replicate:replicate/codellama-34b-python:9048743d22a7b19cd0abb018066809ea6af4f2b4717bef9aad3c5ae21ceac00d","replicate:replicate/codellama-34b:0666717e5ead8557dff55ee8f11924b5c0309f5f1ca52f64bb8eec405fdb38a7"].map(e=>({id:e,config:{temperature:.75,top_p:.9,top_k:50,max_new_tokens:128,min_new_tokens:-1}}))).concat(["replicate:a16z-infra/llama-2-7b-chat:7b0bfc9aff140d5b75bacbed23e91fd3c34b01a1e958d32132de6e0a19796e2c","replicate:a16z-infra/llama-2-13b-chat:2a7f981751ec7fdf87b5b91ad4db53683a98082e9ff7bfd12c8cd5ea85980a52"].map(e=>({id:e,config:{temperature:.95,top_p:.95,top_k:250,max_new_tokens:500,min_new_tokens:-1,repetition_penalty:1,system_prompt:""}}))).concat(["replicate:mistralai/mistral-7b-v0.1","replicate:mistralai/mistral-7b-instruct-v0.2","replicate:mistralai/mixtral-8x7b-instruct-v0.1"].map(e=>({id:e,config:{temperature:.7,top_p:.9,top_k:-1,max_new_tokens:128,min_new_tokens:-1,repetition_penalty:1.15,prompt_template:"{prompt}"}}))).sort((e,t)=>e.id.localeCompare(t.id)),U={anthropic:"Anthropic",bedrock:"Amazon Web Services",azureopenai:"Azure",openai:"OpenAI",replicate:"Replicate"};var q=e=>{let{providers:t,onChange:n}=e,[i,l]=r.useState(null),o=e=>"string"==typeof e?e:e.id||"Unknown provider",c=(e,t)=>"string"==typeof e?e:e.id||t,d=e=>{"string"==typeof e?alert("Cannot edit custom providers"):e.config?l(e):alert("There is no config for this provider")};return(0,a.jsxs)(s.Z,{mt:2,children:[(0,a.jsx)(N.Z,{multiple:!0,freeSolo:!0,options:L,value:t,groupBy:e=>(function(e){if(!e)return"Other";let t=e.split(":")[0];return U[t]||t})(e.id),onChange:(e,t)=>{n(t.map(e=>"string"==typeof e?{id:e}:e))},getOptionLabel:e=>{if(!e)return"";let t="";"string"==typeof e&&(t=e),e.id&&"string"==typeof e.id&&(t=e.id);let n=t.split(":");return n.length>1?n.length>2&&"anthropic"===n[0]?n[2]:n[1]:"Unknown provider"},renderTags:(e,t)=>e.map((e,n)=>{let a=o(e),i=c(e,n);return(0,r.createElement)(z.Z,{variant:"outlined",label:a,...t({index:n}),key:i,onClick:()=>d(e)})}),renderInput:e=>(0,a.jsx)(b.Z,{...e,variant:"outlined",placeholder:"Select LLM providers",helperText:t.length>0?"Click a provider to configure its settings.":null})}),i&&i.id&&(0,a.jsx)(V,{open:!!i,providerId:i.id,config:i.config,onClose:()=>l(null),onSave:e=>{if(i){let a=t.map(t=>t.id===i.id?{...t,config:e}:t);n(a),l(null)}}})]})},B=n(52428),K=n(6882),Y=n(28891),D=()=>{let e=(0,Y.useRouter)(),{defaultTest:t,description:n,env:i,evaluateOptions:s,prompts:c,providers:d,scenarios:p,testCases:u}=(0,l.o)(),[h,x]=(0,r.useState)(!1),[m,f]=(0,r.useState)(0),j=async()=>{x(!0);try{let a=await fetch("".concat(B.eA,"/api/eval/job/"),{method:"POST",headers:{"Content-Type":"application/json"},body:JSON.stringify({defaultTest:t,description:n,env:i,evaluateOptions:s,prompts:c,providers:d,scenarios:p,tests:u})});if(!a.ok)throw Error("HTTP error! status: ".concat(a.status));let r=await a.json(),l=setInterval(async()=>{let t=await fetch("".concat(B.eA,"/api/eval/job/").concat(r.id,"/"));if(!t.ok)throw clearInterval(l),Error("HTTP error! status: ".concat(t.status));let n=await t.json();if("complete"===n.status)clearInterval(l),x(!1),B.Ox?e.push("/eval/remote:".concat(encodeURIComponent(r.id))):e.push("/eval");else if("failed"===n.status)throw clearInterval(l),x(!1),Error("Job failed");else{let e=0===n.total?0:Math.round(n.progress/n.total*100);f(e)}},1e3)}catch(e){console.error(e),x(!1),alert("An error occurred: ".concat(e.message))}};return(0,a.jsx)(o.Z,{variant:"contained",color:"primary",onClick:j,disabled:h,children:h?(0,a.jsxs)(a.Fragment,{children:[(0,a.jsx)(K.Z,{size:24,sx:{marginRight:2}}),m.toFixed(0),"% complete"]}):"Run Evaluation"})},J=n(15320),M=n(66988);let X=["equals","contains","icontains","contains-all","contains-any","starts-with","regex","is-json","contains-json","is-sql","contains-sql","similar","llm-rubric","model-graded-closedqa","factuality","webhook","rouge-n","rouge-s","rouge-l","not-equals","not-contains","not-icontains","not-contains-all","not-contains-any","not-starts-with","not-regex","not-is-json","not-contains-json","not-similar","not-webhook","not-rouge-n","not-rouge-s","not-rouge-l","is-valid-openai-function-call","is-valid-openai-tools-call","latency","perplexity","perplexity-score","cost","answer-relevance","context-faithfulness","context-recall","context-relevance","select-best","moderation"];var G=e=>{let{onAdd:t,initialValues:n}=e,[i,l]=(0,r.useState)(n||[]),c=e=>{let n=i.filter((t,n)=>n!==e);l(n),t(n)};return(0,a.jsxs)(a.Fragment,{children:[(0,a.jsx)(f.Z,{variant:"h6",children:"Asserts"}),(0,a.jsx)(s.Z,{my:i.length>0?2:0,children:(0,a.jsx)(m.Z,{direction:"column",spacing:2,children:i.map((e,n)=>(0,a.jsxs)(m.Z,{direction:"row",spacing:2,alignItems:"center",children:[(0,a.jsx)(N.Z,{value:e.type,options:X,sx:{minWidth:200},onChange:(e,a)=>{let r=i.map((e,t)=>t===n?{...e,type:a}:e);l(r),t(r)},renderInput:e=>(0,a.jsx)(b.Z,{...e,label:"Type"})}),(0,a.jsx)(b.Z,{label:"Value",value:e.value,fullWidth:!0,onChange:e=>{let a=e.target.value,r=i.map((e,t)=>t===n?{...e,value:a}:e);l(r),t(r)}}),(0,a.jsx)(E.Z,{onClick:()=>c(n),size:"small",children:(0,a.jsx)(C.Z,{})})]},n))})}),(0,a.jsx)(o.Z,{color:"primary",onClick:()=>{let e=[...i,{type:"equals",value:""}];l(e),t(e)},children:"Add Assert"})]})},H=e=>{let{onAdd:t,varsList:n,initialValues:i}=e,[l,o]=r.useState(i||{});return(0,r.useEffect)(()=>{let e={};n.forEach(t=>{e[t]=(null==i?void 0:i[t])||""}),o(e)},[n,i]),(0,a.jsxs)(s.Z,{my:2,children:[(0,a.jsx)(f.Z,{variant:"h6",mb:2,children:"Vars"}),n.length>0?(0,a.jsx)(m.Z,{direction:"row",spacing:2,alignItems:"center",children:Object.keys(l).map((e,n)=>(0,a.jsx)(m.Z,{direction:"row",spacing:2,alignItems:"center",children:(0,a.jsx)(b.Z,{placeholder:e,label:e,value:l[e],fullWidth:!0,onChange:n=>{let a=n.target.value,r={...l,[e]:a};o(r),t(r)}})},n))}):(0,a.jsxs)(f.Z,{variant:"subtitle1",gutterBottom:!0,children:["Add variables to your prompt using the ","{{varname}}"," syntax."]})]})},$=e=>{let{open:t,onAdd:n,varsList:i,initialValues:l,onCancel:c}=e,[h,m]=(0,r.useState)((null==l?void 0:l.description)||""),[f,j]=(0,r.useState)((null==l?void 0:l.vars)||{}),[v,g]=(0,r.useState)((null==l?void 0:l.assert)||[]),[Z,b]=(0,r.useState)(0);r.useEffect(()=>{l?(m(l.description||""),j(l.vars||{}),g(l.assert||[])):(m(""),j({}),g([]))},[l]);let y=e=>{n({description:h,vars:f,assert:v},e),e&&c(),m(""),j({}),g([]),b(e=>e+1)};return(0,a.jsxs)(d.Z,{open:t,onClose:c,fullWidth:!0,maxWidth:"md",children:[(0,a.jsx)(x.Z,{children:l?"Edit Test Case":"Add Test Case"}),(0,a.jsx)(u.Z,{children:(0,a.jsxs)(s.Z,{children:[(0,a.jsx)(H,{onAdd:e=>j(e),varsList:i,initialValues:null==l?void 0:l.vars}),(0,a.jsx)(G,{onAdd:e=>g(e),initialValues:((null==l?void 0:l.assert)||[]).filter(e=>"assert-set"!==e.type)||[]},Z)]})}),(0,a.jsxs)(p.Z,{children:[(0,a.jsx)(o.Z,{onClick:y.bind(void 0,!0),color:"primary",variant:"contained",children:l?"Update Test Case":"Add Test Case"}),!l&&(0,a.jsx)(o.Z,{onClick:y.bind(void 0,!1),color:"primary",variant:"contained",children:"Add Another"}),(0,a.jsx)(o.Z,{onClick:c,color:"secondary",children:"Cancel"})]})]})},Q=e=>{let{varsList:t}=e,{testCases:i,setTestCases:s}=(0,l.o)(),[c,d]=r.useState(null),[p,u]=r.useState(!1),h=(e,t)=>{e.stopPropagation(),confirm("Are you sure you want to delete this test case?")&&s(i.filter((e,n)=>n!==t))},x=(e,t)=>{e.stopPropagation();let n=JSON.parse(JSON.stringify(i[t]));s([...i,n])};return(0,a.jsxs)(a.Fragment,{children:[(0,a.jsxs)(m.Z,{direction:"row",spacing:2,mb:2,justifyContent:"space-between",children:[(0,a.jsx)(f.Z,{variant:"h5",children:"Test Cases"}),(0,a.jsxs)("div",{children:[(0,a.jsx)("label",{htmlFor:"file-input-add-test-case",children:(0,a.jsx)(w.Z,{title:"Upload test cases from csv",children:(0,a.jsxs)("span",{children:[(0,a.jsx)(E.Z,{component:"span",children:(0,a.jsx)(k.Z,{})}),(0,a.jsx)("input",{id:"file-input-add-test-case",type:"file",accept:".csv",onChange:e=>{var t;e.stopPropagation(),e.preventDefault();let a=null===(t=e.target.files)||void 0===t?void 0:t[0];if(a){let e=new FileReader;e.onload=async e=>{var t,a;let r=null===(a=e.target)||void 0===a?void 0:null===(t=a.result)||void 0===t?void 0:t.toString();if(r){let{parse:e}=await Promise.all([n.e(631),n.e(887)]).then(n.bind(n,83887)),t=e(r,{columns:!0}),a=t.map(e=>(0,J.It)(e));s([...i,...a])}},e.readAsText(a)}},style:{display:"none"}})]})})}),(0,a.jsx)(o.Z,{color:"primary",onClick:()=>u(!0),variant:"contained",children:"Add Test Case"})]})]}),(0,a.jsx)(P.Z,{children:(0,a.jsxs)(I.Z,{children:[(0,a.jsx)(M.Z,{children:(0,a.jsxs)(T.Z,{children:[(0,a.jsx)(S.Z,{children:"Description"}),(0,a.jsx)(S.Z,{children:"Assertions"}),(0,a.jsx)(S.Z,{children:"Variables"}),(0,a.jsx)(S.Z,{align:"right"})]})}),(0,a.jsx)(O.Z,{children:0===i.length?(0,a.jsx)(T.Z,{children:(0,a.jsx)(S.Z,{colSpan:4,align:"center",children:"No test cases added yet."})}):i.map((e,t)=>{var n;return(0,a.jsxs)(T.Z,{sx:{"&:hover":{backgroundColor:"rgba(0, 0, 0, 0.04)",cursor:"pointer"}},onClick:()=>{d(t),u(!0)},children:[(0,a.jsx)(S.Z,{children:(0,a.jsx)(f.Z,{variant:"body2",children:e.description||"Test Case #".concat(t+1)})}),(0,a.jsxs)(S.Z,{children:[(null===(n=e.assert)||void 0===n?void 0:n.length)||0," assertions"]}),(0,a.jsx)(S.Z,{children:Object.entries(e.vars||{}).map(e=>{let[t,n]=e;return t+"="+n}).join(", ")}),(0,a.jsxs)(S.Z,{align:"right",sx:{minWidth:150},children:[(0,a.jsx)(E.Z,{onClick:()=>{d(t),u(!0)},size:"small",children:(0,a.jsx)(A.Z,{})}),(0,a.jsx)(E.Z,{onClick:e=>x(e,t),size:"small",children:(0,a.jsx)(_.Z,{})}),(0,a.jsx)(E.Z,{onClick:e=>h(e,t),size:"small",children:(0,a.jsx)(C.Z,{})})]})]},t)})})]})}),(0,a.jsx)($,{open:p,onAdd:(e,t)=>{if(null===c)s([...i,e]);else{let t=i.map((t,n)=>n===c?e:t);s(t),d(null)}t&&u(!1)},varsList:t,initialValues:null!==c?i[c]:void 0,onCancel:()=>{d(null),u(!1)}})]})},ee=n(71911),et=n.n(ee),en=n(29963),ea=n(20505),er=n(41101),ei=n(34235),el=n(38640),es=n.n(el),eo=n(29270);n(54492),n(55096),n(93486);var ec=()=>{let e="dark"===(0,er.Z)().palette.mode,{defaultTest:t,setDefaultTest:n,description:i,setDescription:c,env:d,setEnv:p,evaluateOptions:u,setEvaluateOptions:h,prompts:x,setPrompts:m,providers:j,setProviders:v,scenarios:g,setScenarios:Z,testCases:b,setTestCases:y}=(0,l.o)(),[_,C]=r.useState(""),[k,E]=r.useState(!0),I=e=>{n(e.defaultTest||{}),c(e.description||""),p(e.env||{}),h(e.evaluateOptions||{}),m(e.prompts||[]),v(e.providers||[]),Z(e.scenarios||[]),y(e.tests||[])};return r.useEffect(()=>{C(ei.default.dump({defaultTest:t,description:i,env:d,evaluateOptions:u,prompts:x,providers:j,scenarios:g,tests:b}))},[t,i,d,u,x,j,g,b]),(0,a.jsxs)(s.Z,{mt:4,children:[(0,a.jsx)(f.Z,{variant:"h5",gutterBottom:!0,children:"Configuration"}),(0,a.jsxs)(f.Z,{variant:"body1",gutterBottom:!0,children:["This is the YAML config that defines the evaluation and is processed by promptfoo. See"," ",(0,a.jsx)(es(),{target:"_blank",href:"https://promptfoo.dev/docs/configuration/guide",children:"configuration docs"})," ","to learn more."]}),(0,a.jsxs)(s.Z,{display:"flex",gap:2,mb:2,children:[(0,a.jsx)(o.Z,{variant:"text",color:"primary",startIcon:k?(0,a.jsx)(A.Z,{}):(0,a.jsx)(en.Z,{}),onClick:()=>{if(!k)try{let e=ei.default.load(_,{json:!0});I(e)}catch(e){}E(!k)},children:k?"Edit YAML":"Save"}),(0,a.jsxs)(o.Z,{variant:"text",color:"primary",startIcon:(0,a.jsx)(ea.Z,{}),component:"label",children:["Upload YAML",(0,a.jsx)("input",{type:"file",hidden:!0,accept:".yaml,.yml",onChange:e=>{var t;let n=null===(t=e.target.files)||void 0===t?void 0:t[0];if(n){let e=new FileReader;e.onload=e=>{var t;let n=null===(t=e.target)||void 0===t?void 0:t.result;C(n);try{let e=ei.default.load(n,{json:!0});I(e)}catch(e){console.error("Error parsing uploaded YAML:",e)}},e.readAsText(n)}}})]})]}),(0,a.jsx)(et(),{autoCapitalize:"off",value:_,onValueChange:e=>{k||C(e)},highlight:e=>(0,eo.highlight)(e,eo.languages.yaml),padding:10,style:{backgroundColor:e?"#1e1e1e":"#fff",fontFamily:'"Fira code", "Fira Mono", monospace',fontSize:14},disabled:k,className:k?"":"glowing-border"})]})};n(7517);var ed=n(77580);function ep(e){let{error:t,resetErrorBoundary:n}=e;return(0,a.jsxs)("div",{role:"alert",children:[(0,a.jsx)("p",{children:"Something went wrong:"}),(0,a.jsx)("pre",{children:t.message}),(0,a.jsx)("button",{onClick:n,children:"Try again"})]})}var eu=()=>{let[e,t]=(0,r.useState)(!1),{setDescription:n,providers:j,setProviders:v,prompts:g,setPrompts:Z,setTestCases:b}=(0,l.o)();if((0,r.useEffect)(()=>{l.o.persist.rehydrate()},[]),ed.env.NEXT_PUBLIC_NO_BROWSING)return null;let _=(e=>{let t=/{{\s*(\w+)\s*}}/g,n=new Set;return e.forEach(e=>{let a;for(;null!==(a=t.exec(e));)n.add(a[1])}),Array.from(n)})(g);return(0,a.jsxs)(c.Z,{maxWidth:"lg",sx:{marginTop:"2rem"},children:[(0,a.jsxs)(m.Z,{direction:"row",spacing:2,justifyContent:"space-between",children:[(0,a.jsx)(f.Z,{variant:"h4",children:"Set up an evaluation"}),(0,a.jsxs)(m.Z,{direction:"row",spacing:2,children:[(0,a.jsx)(D,{}),(0,a.jsx)(y,{}),(0,a.jsx)(o.Z,{variant:"outlined",color:"primary",onClick:()=>t(!0),children:"Reset"})]})]}),(0,a.jsx)(s.Z,{mt:4}),(0,a.jsx)(s.Z,{mt:2,children:(0,a.jsx)(i.SV,{FallbackComponent:ep,onReset:()=>{v([])},children:(0,a.jsxs)(m.Z,{direction:"column",spacing:2,justifyContent:"space-between",children:[(0,a.jsx)(f.Z,{variant:"h5",children:"Providers"}),(0,a.jsx)(q,{providers:j,onChange:v})]})})}),(0,a.jsx)(s.Z,{mt:4}),(0,a.jsx)(i.SV,{FallbackComponent:ep,onReset:()=>{Z([])},children:(0,a.jsx)(W,{})}),(0,a.jsx)(s.Z,{mt:6}),(0,a.jsx)(i.SV,{FallbackComponent:ep,onReset:()=>{b([])},children:(0,a.jsx)(Q,{varsList:_})}),(0,a.jsx)(ec,{}),(0,a.jsxs)(d.Z,{open:e,onClose:()=>t(!1),"aria-labelledby":"alert-dialog-title","aria-describedby":"alert-dialog-description",children:[(0,a.jsx)(x.Z,{id:"alert-dialog-title",children:"Confirm Reset"}),(0,a.jsx)(u.Z,{children:(0,a.jsx)(h.Z,{id:"alert-dialog-description",children:"Are you sure you want to reset all the fields? This action cannot be undone."})}),(0,a.jsxs)(p.Z,{children:[(0,a.jsx)(o.Z,{onClick:()=>t(!1),children:"Cancel"}),(0,a.jsx)(o.Z,{onClick:()=>{n(""),v([]),Z([]),b([]),t(!1)},autoFocus:!0,children:"Reset"})]})]})]})}},52428:function(e,t,n){"use strict";n.d(t,{Ox:function(){return i},T8:function(){return r},eA:function(){return l}});var a=n(77580);let r=!a.env.NEXT_PUBLIC_PROMPTFOO_BUILD_STANDALONE_SERVER,i=!!a.env.NEXT_PUBLIC_PROMPTFOO_USE_SUPABASE,l=""},47887:function(e,t,n){"use strict";n.d(t,{o:function(){return i}});var a=n(94660),r=n(74810);let i=(0,a.Ue)()((0,r.tJ)((e,t)=>({env:{},testCases:[],description:"",providers:[],prompts:[],defaultTest:{},evaluateOptions:{},scenarios:[],setEnv:t=>e({env:t}),setTestCases:t=>e({testCases:t}),setDescription:t=>e({description:t}),setProviders:t=>e({providers:t}),setPrompts:t=>e({prompts:t}),setDefaultTest:t=>e({defaultTest:t}),setEvaluateOptions:t=>e({evaluateOptions:t}),setScenarios:t=>e({scenarios:t}),setStateFromConfig:t=>{let n={};t.description&&(n.description=t.description||""),t.tests&&(n.testCases=t.tests),t.providers&&(n.providers=t.providers),t.prompts&&("string"==typeof t.prompts?n.prompts=[t.prompts]:Array.isArray(t.prompts)?n.prompts=t.prompts.filter(e=>"string"==typeof e&&!e.endsWith(".txt")&&!e.endsWith(".json")&&!e.endsWith(".yaml")):console.warn("Invalid prompts config",t.prompts)),t.defaultTest&&(n.defaultTest=t.defaultTest),t.evaluateOptions&&(n.evaluateOptions=t.evaluateOptions),t.scenarios&&(n.scenarios=t.scenarios),e(n)},getTestSuite:()=>{let{description:e,testCases:n,providers:a,prompts:r,env:i,scenarios:l}=t();return{env:i,description:e,providers:a,prompts:r,tests:n,scenarios:l}}}),{name:"promptfoo",skipHydration:!0}))},94499:function(){},55096:function(){},7517:function(){},15320:function(e,t){"use strict";t.It=function(e){let t,n,a,r;let i={},l=[],s={};for(let[o,c]of Object.entries(e))o.startsWith("__expected")?""!==c.trim()&&l.push(function(e){if(e.startsWith("javascript:")||e.startsWith("fn:")||e.startsWith("eval:")){let t;e.startsWith("javascript:")&&(t=11),e.startsWith("fn:")&&(t=3),e.startsWith("eval:")&&(t=5);let n=e.slice(t).trim();return{type:"javascript",value:n}}if(e.startsWith("grade:")||e.startsWith("llm-rubric:"))return{type:"llm-rubric",value:e.slice(6)};if(e.startsWith("python:")){let t=e.slice(7).trim();return{type:"python",value:t}}let t=e.match(/^(not-)?(equals|contains-any|contains-all|icontains-any|icontains-all|contains-json|is-json|is-sql|regex|icontains|contains|webhook|rouge-n|similar|starts-with|levenshtein|classifier|model-graded-factuality|factuality|model-graded-closedqa|answer-relevance|context-recall|context-relevance|context-faithfulness|is-valid-openai-function-call|is-valid-openai-tools-call|latency|perplexity|perplexity-score|cost)(?:\((\d+(?:\.\d+)?)\))?(?::([\s\S]*))?$/);if(t){let[e,n,a,r,i]=t,l=n?`not-${a}`:a,s=parseFloat(r);return"contains-any"===a||"contains-all"===a||"icontains-any"===a||"icontains-all"===a?{type:l,value:i.split(",").map(e=>e.trim())}:"contains-json"===a||"is-json"===a?{type:l,value:i}:"rouge-n"===a||"similar"===a||"starts-with"===a||"levenshtein"===a||"classifier"===a||"answer-relevance"===a||"context-recall"===a||"context-relevance"===a||"context-faithfulness"===a||"latency"===a||"perplexity"===a||"perplexity-score"===a||"cost"===a?{type:l,value:i,threshold:s||("similar"===a?.8:.75)}:{type:l,value:i}}return{type:"equals",value:e}}(c)):"__prefix"===o?s.prefix=c:"__suffix"===o?s.suffix=c:"__description"===o?n=c:"__providerOutput"===o?t=c:"__metric"===o?a=c:"__threshold"===o?r=parseFloat(c):i[o]=c;for(let e of l)e.metric=a;return{vars:i,assert:l,options:s,...n?{description:n}:{},...t?{providerOutput:t}:{},...r?{threshold:r}:{}}}}},function(e){e.O(0,[148,640,193,391,654,65,975,276,874,235,264,734,972,325,470,730,744],function(){return e(e.s=8458)}),_N_E=e.O()}]);

package/dist/src/web/nextui/_next/static/chunks/main-app-0dc90092cd7acaaf.js

@@ -1 +0,0 @@
-(self.webpackChunk_N_E=self.webpackChunk_N_E||[]).push([[744],{71590:function(e,n,t){Promise.resolve().then(t.t.bind(t,6020,23)),Promise.resolve().then(t.t.bind(t,83507,23)),Promise.resolve().then(t.t.bind(t,27883,23)),Promise.resolve().then(t.t.bind(t,46785,23)),Promise.resolve().then(t.t.bind(t,27382,23))}},function(e){var n=function(n){return e(e.s=n)};e.O(0,[470,730],function(){return n(11409),n(71590)}),_N_E=e.O()}]);