Feat: better evaluation results

packages/backend/src/api/v1/evaluations/index.ts

@@ -49,15 +49,20 @@ evaluations.post(
       checks: [], // TODO: remove this legacy col from DB,
     }
 
-    const [insertedEvaluation] =
+    const [evaluation] =
       await sql`insert into evaluation ${sql(evaluationToInsert)} returning *`
 
-    const evaluation = await getEvaluation(insertedEvaluation.id)
+    const prompts = await sql`
+      select * from dataset_prompt where dataset_id = ${datasetId}
+    `
 
     let count = 0
 
-    for (const prompt of evaluation.dataset.prompts) {
-      for (const variation of prompt.variations) {
+    for (const prompt of prompts) {
+      const variations = await sql`
+        select * from dataset_prompt_variation where prompt_id = ${prompt.id}
+      `
+      for (const variation of variations) {
         for (const provider of evaluation.providers) {
           count++
           queue.add(() =>
@@ -66,7 +71,7 @@ evaluations.post(
               promptId: prompt.id,
               variation,
               provider,
-              prompt: prompt.content,
+              prompt: prompt.messages,
               checklistId,
             }),
           )
@@ -121,9 +126,7 @@ evaluations.get(
     const results = await sql`
   select 
     *,
-    p.id as prompt_id,
-    p.messages as prompt_content
-    --p.extra as prompt_extra
+    p.id as prompt_id
   from 
     evaluation_result er 
     left join dataset_prompt p on p.id = er.prompt_id

packages/backend/src/api/v1/evaluations/utils.ts

@@ -133,71 +133,3 @@ export async function runEval({
     console.error(error)
   }
 }
-
-export async function getEvaluation(evaluationId: string) {
-  const rows = await sql`
-    select
-      e.id as id,
-      e.created_at as created_at,
-      e.name as name,
-      e.project_id as project_id,
-      e.owner_id as owner_id,
-      e.providers as providers,
-      e.checks as checks,
-      d.id as dataset_id,
-      d.slug as dataset_slug,
-      p.id as prompt_id,
-      p.messages as prompt_messages,
-      pv.id as variation_id,
-      pv.variables,
-      pv.context,
-      pv.ideal_output
-    from
-      evaluation e
-      left join dataset d on e.dataset_id = d.id 
-      left join dataset_prompt p on d.id = p.dataset_id
-      left join dataset_prompt_variation pv on pv.prompt_id = p.id
-    where 
-      e.id = ${evaluationId}
-    `
-
-  const {
-    id,
-    createdAt,
-    name,
-    ownerId,
-    projectId,
-    providers,
-    checks,
-    datasetId,
-    datasetSlug,
-  } = rows[0]
-
-  const evaluation = {
-    id,
-    createdAt,
-    name,
-    projectId,
-    ownerId,
-    providers,
-    checks,
-    dataset: {
-      id: datasetId,
-      slug: datasetSlug,
-      prompts: rows.map(({ promptId, promptMessages }) => ({
-        id: promptId,
-        content: promptMessages,
-        variations: rows
-          .filter((row) => row.promptId === promptId)
-          .map(({ variationId, variables, context, idealOutput }) => ({
-            id: variationId,
-            variables,
-            context,
-            idealOutput,
-          })),
-      })),
-    },
-  }
-
-  return evaluation
-}

packages/db/0009.sql

@@ -0,0 +1,8 @@
+alter table evaluation add column if not exists checklist_id uuid;
+alter table evaluation DROP CONSTRAINT IF EXISTS evaluation_checklist_id_fkey;
+alter table evaluation add constraint evaluation_checklist_id_fkey foreign key (checklist_id) references checklist(id) on delete set null;
+
+drop table if exists evaluation_prompt cascade;
+drop table if exists evaluation_prompt_variation cascade;
+
+alter table evaluation_result add constraint "fk_evaluation_result_prompt_id" foreign key (prompt_id) references dataset_prompt(id) on delete cascade;

packages/frontend/components/blocks/Feedback.tsx

@@ -34,9 +34,12 @@ export default function Feedback({
   })
 
   return (
-    <Tooltip label="Feedback from parent run" position="bottom">
+    <Tooltip
+      label="Feedback from parent run"
+      position="bottom"
+      disabled={!isFromParent}
+    >
       <Indicator inline disabled={!isFromParent} color="red" size={10}>
-        {/* <Tooltip disabled={!isFromParent} label={"Feedback from parent run"}> */}
         <Group
           style={{
             padding: isFromParent ? "3px 6px" : "",