Simpler evaluations — part 2.4

latitude-dev · Nov 4, 2024 · 6c67335 · 6c67335
1 parent 203455e
commit 6c67335
Show file tree

Hide file tree

Showing 29 changed files with 403 additions and 224 deletions.
diff --git a/apps/web/src/actions/copilot/refinePrompt.ts b/apps/web/src/actions/copilot/refinePrompt.ts
@@ -8,6 +8,7 @@ import {
   EvaluationsRepository,
 } from '@latitude-data/core/repositories'
 import { serialize as serializeEvaluationResult } from '@latitude-data/core/services/evaluationResults/serialize'
+import { getEvaluationPrompt } from '@latitude-data/core/services/evaluations/index'
 import { env } from '@latitude-data/env'
 import { createSdk } from '$/app/(private)/_lib/createSdk'
 import { z } from 'zod'
@@ -76,11 +77,16 @@ export const refinePromptAction = authProcedure
       projectId: env.COPILOT_PROJECT_ID,
     }).then((r) => r.unwrap())
 
+    const evaluationPrompt = await getEvaluationPrompt({
+      workspace: ctx.workspace,
+      evaluation,
+    }).then((r) => r.unwrap())
+
     const result = await sdk.run(env.COPILOT_REFINE_PROMPT_PATH, {
       stream: false,
       parameters: {
         prompt: document.content,
-        evaluation: evaluation.metadata.prompt,
+        evaluation: evaluationPrompt,
         results: serializedEvaluationResults,
       },
     })

diff --git a/apps/web/src/app/(private)/_data-access/index.ts b/apps/web/src/app/(private)/_data-access/index.ts
@@ -212,13 +212,20 @@ export const getApiKeysCached = cache(async () => {
   return result.unwrap()
 })
 
-export const getProviderApiKeyCached = cache(async (name: string) => {
+export const getProviderApiKeyByNameCached = cache(async (name: string) => {
   const { workspace } = await getCurrentUser()
   const scope = new ProviderApiKeysRepository(workspace.id)
   const result = await scope.findByName(name)
   return result.unwrap()
 })
 
+export const getProviderApiKeyByIdCached = cache(async (id: number) => {
+  const { workspace } = await getCurrentUser()
+  const scope = new ProviderApiKeysRepository(workspace.id)
+  const result = await scope.find(id)
+  return result.unwrap()
+})
+
 export const getProviderApiKeysCached = cache(async () => {
   const { workspace } = await getCurrentUser()
   const scope = new ProviderApiKeysRepository(workspace.id)

diff --git a/...vate)/evaluations/(evaluation)/[evaluationUuid]/dashboard/_components/EvaluationStats.tsx b/...vate)/evaluations/(evaluation)/[evaluationUuid]/dashboard/_components/EvaluationStats.tsx
@@ -3,7 +3,10 @@
 import { useEffect, useState } from 'react'
 
 import { readMetadata } from '@latitude-data/compiler'
-import { EvaluationDto } from '@latitude-data/core/browser'
+import {
+  EvaluationDto,
+  EvaluationMetadataType,
+} from '@latitude-data/core/browser'
 import { ConnectedDocumentWithMetadata } from '@latitude-data/core/repositories'
 import { Skeleton, Text } from '@latitude-data/web-ui'
 import { formatCostInMillicents } from '$/app/_lib/formatUtils'
@@ -34,10 +37,14 @@ export default function EvaluationStats({
     useConnectedDocuments({ evaluation })
 
   useEffect(() => {
-    readMetadata({ prompt: evaluation.metadata.prompt }).then((metadata) => {
-      const metadataModel = (metadata.config['model'] as string) ?? 'Unknown'
-      setModel(metadataModel)
-    })
+    if (evaluation.metadataType === EvaluationMetadataType.LlmAsJudgeAdvanced) {
+      readMetadata({ prompt: evaluation.metadata.prompt }).then((metadata) => {
+        const metadataModel = (metadata.config['model'] as string) ?? 'Unknown'
+        setModel(metadataModel)
+      })
+    } else {
+      setModel(evaluation.metadata.model)
+    }
   }, [evaluation.metadata])
 
   return (

diff --git a/...on)/[evaluationUuid]/editor/_components/EvaluationEditor/Editor/Playground/Chat/index.tsx b/...on)/[evaluationUuid]/editor/_components/EvaluationEditor/Editor/Playground/Chat/index.tsx
@@ -7,6 +7,7 @@ import {
 import {
   ChainEventTypes,
   EvaluationDto,
+  EvaluationMetadataLlmAsJudgeAdvanced,
   StreamEventTypes,
 } from '@latitude-data/core/browser'
 import {
@@ -52,6 +53,10 @@ export default function Chat({
   const [responseStream, setResponseStream] = useState<string | undefined>()
   const [isStreaming, setIsStreaming] = useState(false)
 
+  // TODO: Only advanced evaluations are available right now. Next PR will add saparate components for each evaluation type
+  const prompt = (evaluation.metadata as EvaluationMetadataLlmAsJudgeAdvanced)
+    .prompt
+
   const addMessageToConversation = useCallback(
     (message: ConversationMessage) => {
       let newConversation: Conversation
@@ -75,7 +80,7 @@ export default function Chat({
     let messagesCount = 0
 
     const [data, error] = await runPromptAction({
-      prompt: evaluation.metadata.prompt,
+      prompt,
       parameters,
     })
     if (error) {

diff --git a/...ations/(evaluation)/[evaluationUuid]/editor/_components/EvaluationEditor/Editor/index.tsx b/...ations/(evaluation)/[evaluationUuid]/editor/_components/EvaluationEditor/Editor/index.tsx
@@ -3,6 +3,7 @@
 import { Suspense, useCallback, useEffect, useMemo, useState } from 'react'
 
 import {
+  EvaluationMetadataLlmAsJudgeAdvanced,
   ProviderApiKey,
   SERIALIZED_DOCUMENT_LOG_FIELDS,
 } from '@latitude-data/core/browser'
@@ -70,6 +71,10 @@ export default function EvaluationEditor({
 
   if (!evaluation) return null
 
+  // TODO: Only advanced evaluations are available right now. Next PR will add saparate components for each evaluation type
+  const prompt = (evaluation.metadata as EvaluationMetadataLlmAsJudgeAdvanced)
+    .prompt
+
   return (
     <div className='flex flex-row w-full h-full gap-8'>
       <div className='flex flex-col flex-1 flex-grow flex-shrink gap-2 min-w-0'>
@@ -82,7 +87,7 @@ export default function EvaluationEditor({
           onChangePrompt={onChange}
           rightActions={
             <>
-              {value !== evaluation.metadata.prompt && (
+              {value !== prompt && (
                 <Button
                   fancy
                   disabled={isUpdating || isLoading}

diff --git a/...)/evaluations/(evaluation)/[evaluationUuid]/editor/_components/EvaluationEditorLayout.tsx b/...)/evaluations/(evaluation)/[evaluationUuid]/editor/_components/EvaluationEditorLayout.tsx
@@ -2,7 +2,11 @@
 
 import { ReactNode } from 'react'
 
-import { EvaluationDto, ProviderApiKey } from '@latitude-data/core/browser'
+import {
+  EvaluationDto,
+  EvaluationMetadataLlmAsJudgeAdvanced,
+  ProviderApiKey,
+} from '@latitude-data/core/browser'
 import { Button, Icon } from '@latitude-data/web-ui'
 import Link from 'next/link'
 import { useSearchParams } from 'next/navigation'
@@ -28,6 +32,10 @@ export default function EvaluationEditorLayout({
   const searchParams = useSearchParams()
   const backUrl = searchParams.get('back')
 
+  // TODO: Only advanced evaluations are available right now. Next PR will add saparate components for each evaluation type
+  const prompt = (evaluation.metadata as EvaluationMetadataLlmAsJudgeAdvanced)
+    .prompt
+
   return (
     <div className='h-full flex flex-col gap-y-4 p-6'>
       {backUrl && (
@@ -42,7 +50,7 @@ export default function EvaluationEditorLayout({
         <EvaluationEditor
           providerApiKeys={providerApiKeys}
           evaluationUuid={evaluationUuid}
-          defaultPrompt={evaluation.metadata.prompt}
+          defaultPrompt={prompt}
           freeRunsCount={freeRunsCount ? Number(freeRunsCount) : undefined}
         />
       </div>

diff --git a/...r/Editor/RefineModal/steps/2_SelectEvaluationResults/SelectableEvaluationResultsTable.tsx b/...r/Editor/RefineModal/steps/2_SelectEvaluationResults/SelectableEvaluationResultsTable.tsx
@@ -29,21 +29,17 @@ export const ResultCellContent = ({
   evaluation: EvaluationDto
   value: unknown
 }) => {
-  if (
-    evaluation.metadata.configuration.type === EvaluationResultableType.Boolean
-  ) {
+  if (evaluation.resultType === EvaluationResultableType.Boolean) {
     return (
       <Text.H4 color={(value as boolean) ? 'success' : 'destructive'}>
         {String(value)}
       </Text.H4>
     )
   }
 
-  if (
-    evaluation.metadata.configuration.type === EvaluationResultableType.Number
-  ) {
-    const minValue = evaluation.metadata.configuration.detail?.range.from ?? 0
-    const maxValue = evaluation.metadata.configuration.detail?.range.to ?? 10
+  if (evaluation.resultType === EvaluationResultableType.Number) {
+    const minValue = evaluation.resultConfiguration.minValue
+    const maxValue = evaluation.resultConfiguration.maxValue
 
     return (
       <RangeBadge

diff --git a/...Uuid]/evaluations/[evaluationId]/_components/EvaluationResults/EvaluationResultsTable.tsx b/...Uuid]/evaluations/[evaluationId]/_components/EvaluationResults/EvaluationResultsTable.tsx
@@ -39,21 +39,17 @@ export const ResultCellContent = ({
   evaluation: EvaluationDto
   value: unknown
 }) => {
-  if (
-    evaluation.metadata.configuration.type === EvaluationResultableType.Boolean
-  ) {
+  if (evaluation.resultType === EvaluationResultableType.Boolean) {
     return (
       <Badge variant={value === 'true' ? 'success' : 'destructive'}>
         {String(value)}
       </Badge>
     )
   }
 
-  if (
-    evaluation.metadata.configuration.type === EvaluationResultableType.Number
-  ) {
-    const minValue = evaluation.metadata.configuration.detail?.range.from ?? 0
-    const maxValue = evaluation.metadata.configuration.detail?.range.to ?? 10
+  if (evaluation.resultType === EvaluationResultableType.Number) {
+    const minValue = evaluation.resultConfiguration.minValue
+    const maxValue = evaluation.resultConfiguration.maxValue
 
     return (
       <RangeBadge

diff --git a/...ations/[evaluationId]/_components/MetricsSummary/BigNumberPanels/MeanValuePanel/index.tsx b/...ations/[evaluationId]/_components/MetricsSummary/BigNumberPanels/MeanValuePanel/index.tsx
@@ -2,7 +2,10 @@
 
 import { useCallback } from 'react'
 
-import { EvaluationDto } from '@latitude-data/core/browser'
+import {
+  EvaluationConfigurationNumerical,
+  EvaluationDto,
+} from '@latitude-data/core/browser'
 import { RangeBadge } from '@latitude-data/web-ui'
 import useEvaluationResultsMeanValue from '$/stores/evaluationResultCharts/evaluationResultsMeanValue'
 import { useDebouncedCallback } from 'use-debounce'
@@ -40,9 +43,10 @@ export default function MeanValuePanel({
     documentUuid,
     onStatusChange,
   })
-  const config = evaluation.metadata.configuration.detail!
-  const defaultMinValue = config.range.from
-  const defaultMaxValue = config.range.to
+  const config =
+    evaluation.resultConfiguration as EvaluationConfigurationNumerical
+  const defaultMinValue = config.minValue
+  const defaultMaxValue = config.maxValue
   return (
     <Panel
       label='Current average'

diff --git a/...mentUuid]/evaluations/[evaluationId]/_components/MetricsSummary/BigNumberPanels/index.tsx b/...mentUuid]/evaluations/[evaluationId]/_components/MetricsSummary/BigNumberPanels/index.tsx
@@ -1,16 +1,18 @@
-import { Commit, EvaluationDto } from '@latitude-data/core/browser'
+import {
+  Commit,
+  EvaluationDto,
+  EvaluationResultableType,
+} from '@latitude-data/core/browser'
 
 import MeanValuePanel from './MeanValuePanel'
 import ModalValuePanel from './ModalValuePanel'
 import TotalsPanels from './TotalsPanels'
 
-export function BigNumberPanels<T extends boolean>({
+export function BigNumberPanels({
   commit,
   evaluation,
   documentUuid,
-  isNumeric,
 }: {
-  isNumeric: T
   commit: Commit
   evaluation: EvaluationDto
   documentUuid: string
@@ -23,15 +25,13 @@ export function BigNumberPanels<T extends boolean>({
         evaluationId={evaluation.id}
       />
 
-      {isNumeric && (
+      {evaluation.resultType == EvaluationResultableType.Number ? (
         <MeanValuePanel
           evaluation={evaluation}
           commitUuid={commit.uuid}
           documentUuid={documentUuid}
         />
-      )}
-
-      {!isNumeric && (
+      ) : (
         <ModalValuePanel
           evaluationId={evaluation.id}
           commitUuid={commit.uuid}

diff --git a/...valuations/[evaluationId]/_components/MetricsSummary/Charts/Numerical/CostOverResults.tsx b/...valuations/[evaluationId]/_components/MetricsSummary/Charts/Numerical/CostOverResults.tsx
@@ -2,7 +2,10 @@
 
 import { useCallback, useMemo } from 'react'
 
-import { EvaluationDto } from '@latitude-data/core/browser'
+import {
+  EvaluationConfigurationNumerical,
+  EvaluationDto,
+} from '@latitude-data/core/browser'
 import {
   Badge,
   ScatterChart,
@@ -23,6 +26,8 @@ export function CostOverResultsChart({
   evaluation: EvaluationDto
   documentUuid: string
 }) {
+  const evaluationConfiguration =
+    evaluation.resultConfiguration as EvaluationConfigurationNumerical
   const { project } = useCurrentProject()
   const { commit } = useCurrentCommit()
   const { isLoading, error, data, refetch } =
@@ -75,8 +80,8 @@ export function CostOverResultsChart({
             xAxis: {
               label: 'Average result',
               type: 'number',
-              min: evaluation.metadata.configuration.detail!.range.from,
-              max: evaluation.metadata.configuration.detail!.range.to,
+              min: evaluationConfiguration.minValue,
+              max: evaluationConfiguration.maxValue,
             },
             yAxis: {
               label: 'Average cost',

diff --git a/...valuations/[evaluationId]/_components/MetricsSummary/Charts/Numerical/ResultsOverTime.tsx b/...valuations/[evaluationId]/_components/MetricsSummary/Charts/Numerical/ResultsOverTime.tsx
@@ -2,7 +2,10 @@
 
 import { useCallback, useMemo } from 'react'
 
-import { EvaluationDto } from '@latitude-data/core/browser'
+import {
+  EvaluationConfigurationNumerical,
+  EvaluationDto,
+} from '@latitude-data/core/browser'
 import {
   AreaChart,
   Text,
@@ -29,6 +32,8 @@ export function ResultOverTimeChart({
   evaluation: EvaluationDto
   documentUuid: string
 }) {
+  const evaluationConfiguration =
+    evaluation.resultConfiguration as EvaluationConfigurationNumerical
   const { project } = useCurrentProject()
   const { commit } = useCurrentCommit()
   const { isLoading, error, data, refetch } = useAverageResultOverTime(
@@ -97,8 +102,8 @@ export function ResultOverTimeChart({
             yAxis: {
               label: 'Average result',
               type: 'number',
-              min: evaluation.metadata.configuration.detail!.range.from,
-              max: evaluation.metadata.configuration.detail!.range.to,
+              min: evaluationConfiguration.minValue,
+              max: evaluationConfiguration.maxValue,
             },
             data: parsedData,
             tooltipLabel: (item) => {

diff --git a/...nts/[documentUuid]/evaluations/[evaluationId]/_components/MetricsSummary/Charts/index.tsx b/...nts/[documentUuid]/evaluations/[evaluationId]/_components/MetricsSummary/Charts/index.tsx
@@ -12,10 +12,6 @@ export function EvaluationResultsCharts({
   evaluation: EvaluationDto
   documentUuid: string
 }) {
-  const isNumerical =
-    evaluation.metadata.configuration.type === EvaluationResultableType.Number
-
-  if (!isNumerical) return null
-
+  if (evaluation.resultType != EvaluationResultableType.Number) return null
   return <NumericalCharts evaluation={evaluation} documentUuid={documentUuid} />
 }
diff --git a/.../documents/[documentUuid]/evaluations/[evaluationId]/_components/MetricsSummary/index.tsx b/.../documents/[documentUuid]/evaluations/[evaluationId]/_components/MetricsSummary/index.tsx
@@ -5,16 +5,14 @@ import { Commit, EvaluationDto } from '@latitude-data/core/browser'
 import { BigNumberPanels } from './BigNumberPanels'
 import { EvaluationResultsCharts } from './Charts'
 
-export function MetricsSummary<T extends boolean>({
+export function MetricsSummary({
   commit,
   evaluation,
   documentUuid,
-  isNumeric,
 }: {
   commit: Commit
   evaluation: EvaluationDto
   documentUuid: string
-  isNumeric: T
 }) {
   return (
     <div className='flex gap-6 flex-wrap'>
@@ -27,7 +25,6 @@ export function MetricsSummary<T extends boolean>({
           commit={commit}
           evaluation={evaluation}
           documentUuid={documentUuid}
-          isNumeric={isNumeric}
         />
       </div>
     </div>