From 910d26add380dedbfeec2fcc1370a13a61f0d1fb Mon Sep 17 00:00:00 2001
From: Jose Henrique <jose.henrique.ivan@gmail.com>
Date: Fri, 3 Apr 2026 11:48:17 -0300
Subject: [PATCH] fixes!

---
 packages/backend/src/agents/retrieval.ts         |  1 +
 packages/backend/src/pipelines/recommendation.ts | 13 +++++++------
 2 files changed, 8 insertions(+), 6 deletions(-)

diff --git a/packages/backend/src/agents/retrieval.ts b/packages/backend/src/agents/retrieval.ts
index bc6da7b..f292104 100644
--- a/packages/backend/src/agents/retrieval.ts
+++ b/packages/backend/src/agents/retrieval.ts
@@ -37,6 +37,7 @@ Rules:
 - Each "reason" should briefly explain why the ${mediaLabel} matches the preferences
 - Avoid duplicates
 - Include ${mediaLabelPlural} from different decades, countries${mediaType === 'tv_show' ? ', and networks' : ', and directors'}
+- The "title" field must contain ONLY the exact title name — no years, descriptions, network names, episode counts, or parenthetical notes. ✗ Bad: "Breaking Bad (2008–2013, AMC)" ✓ Good: "Breaking Bad"
 - Aim for ${brainstormCount} candidates minimum${previousFullMatches.length > 0 ? '\n- Do NOT suggest titles already in the Previous Full Matches list — generate NEW candidates inspired by what made those successful' : ''}${hardRequirements ? '\n\nIMPORTANT: Strictly follow ALL requirements. Exclude any candidate that does not meet every stated requirement.' : ''}`,
     input: `Structured preferences:
 Liked ${mediaLabelPlural}: ${input.liked.join(', ') || '(none)'}
diff --git a/packages/backend/src/pipelines/recommendation.ts b/packages/backend/src/pipelines/recommendation.ts
index 2afded1..f67d227 100644
--- a/packages/backend/src/pipelines/recommendation.ts
+++ b/packages/backend/src/pipelines/recommendation.ts
@@ -121,12 +121,13 @@ async function runSubPipeline(ctx: SubPipelineCtx): Promise<CuratorOutput[]> {
       runRanking(interpreterOutput, { candidates: bucket }, mediaType, useHardRequirements)
     )
   );
+  const dedupTitles = (titles: string[]) => [...new Map(titles.map((t) => [t.toLowerCase(), t])).values()];
   const rankingOutput: RankingOutput = {
-    full_match: rankingBuckets.flatMap((r) => r.full_match),
-    definitely_like: rankingBuckets.flatMap((r) => r.definitely_like),
-    might_like: rankingBuckets.flatMap((r) => r.might_like),
-    questionable: rankingBuckets.flatMap((r) => r.questionable),
-    will_not_like: rankingBuckets.flatMap((r) => r.will_not_like),
+    full_match: dedupTitles(rankingBuckets.flatMap((r) => r.full_match)),
+    definitely_like: dedupTitles(rankingBuckets.flatMap((r) => r.definitely_like)),
+    might_like: dedupTitles(rankingBuckets.flatMap((r) => r.might_like)),
+    questionable: dedupTitles(rankingBuckets.flatMap((r) => r.questionable)),
+    will_not_like: dedupTitles(rankingBuckets.flatMap((r) => r.will_not_like)),
   };
   log(recId, `${stagePrefix}Ranking: done (${Date.now() - t2}ms) — ${rankBucketCount} buckets`, {
     full_match: rankingOutput.full_match.length,
@@ -163,7 +164,7 @@ async function runSubPipeline(ctx: SubPipelineCtx): Promise<CuratorOutput[]> {
       runCurator(ranking, interpreterOutput, mediaType, useWebSearch)
     )
   );
-  const curatorOutput = curatorBucketOutputs.flat();
+  const curatorOutput = curatorBucketOutputs.reduce((acc, bucket) => mergeCuratorOutputs(acc, bucket), [] as CuratorOutput[]);
   log(recId, `${stagePrefix}Curator: done (${Date.now() - t3}ms) — ${curatorOutput.length} items curated (${curatorBucketCount} buckets)`);
   sseWrite({ stage: p('curator'), status: 'done', data: curatorOutput });