new things
All checks were successful
Recommender Build and Deploy (internal) / Build Recommender Image (push) Successful in 4m4s
Recommender Build and Deploy (internal) / Deploy Recommender (internal) (push) Successful in 11s

This commit is contained in:
2026-04-02 19:24:58 -03:00
parent 91870f4046
commit ba38092784
19 changed files with 695 additions and 130 deletions

View File

@@ -3,14 +3,16 @@ import { db } from '../db.js';
import { recommendations } from '../db/schema.js';
import { runInterpreter } from '../agents/interpreter.js';
import { runRetrieval } from '../agents/retrieval.js';
import { runValidator } from '../agents/validator.js';
import { runRanking } from '../agents/ranking.js';
import { runCurator } from '../agents/curator.js';
import type { CuratorOutput, MediaType, RankingOutput, RetrievalCandidate, SSEEvent } from '../types/agents.js';
import type { CuratorOutput, InterpreterOutput, MediaType, RankingOutput, RetrievalCandidate, SSEEvent } from '../types/agents.js';
import { generateTitle } from '../agents/titleGenerator.js';
/* -- Agent pipeline --
[1] Interpreter -> gets user input, transforms into structured data
[2] Retrieval -> gets candidates from OpenAI (high temperature)
[2.5] Validator (optional) -> verifies candidates exist, removes trash
[3] Ranking -> ranks candidates based on user input
[4] Curator -> curates candidates based on user input
*/
@@ -24,8 +26,8 @@ function getBucketCount(count: number): number {
return 4;
}
function deduplicateCandidates(candidates: RetrievalCandidate[]): RetrievalCandidate[] {
const seen = new Set<string>();
function deduplicateCandidates(candidates: RetrievalCandidate[], seenTitles?: Set<string>): RetrievalCandidate[] {
const seen = seenTitles ?? new Set<string>();
return candidates.filter((c) => {
const key = c.title.toLowerCase();
if (seen.has(key)) return false;
@@ -40,6 +42,11 @@ function splitIntoBuckets<T>(items: T[], n: number): T[][] {
.filter((b) => b.length > 0);
}
function mergeCuratorOutputs(a: CuratorOutput[], b: CuratorOutput[]): CuratorOutput[] {
const seen = new Set(a.map((x) => x.title.toLowerCase()));
return [...a, ...b.filter((x) => !seen.has(x.title.toLowerCase()))];
}
function log(recId: string, msg: string, data?: unknown) {
const ts = new Date().toISOString();
if (data !== undefined) {
@@ -49,6 +56,123 @@ function log(recId: string, msg: string, data?: unknown) {
}
}
interface SubPipelineCtx {
recId: string;
interpreterOutput: InterpreterOutput;
mediaType: MediaType;
useWebSearch: boolean;
useValidator: boolean;
useHardRequirements: boolean;
brainstormCount: number;
previousFullMatches: string[];
allSeenTitles: Set<string>;
stagePrefix: string;
sseWrite: (event: SSEEvent) => void;
}
async function runSubPipeline(ctx: SubPipelineCtx): Promise<CuratorOutput[]> {
const {
recId, interpreterOutput, mediaType, useWebSearch, useValidator,
useHardRequirements, brainstormCount, previousFullMatches,
allSeenTitles, stagePrefix, sseWrite,
} = ctx;
const p = (stage: string) => (stagePrefix + stage) as SSEEvent['stage'];
// --- Retrieval (bucketed) ---
log(recId, `${stagePrefix}Retrieval: start`);
sseWrite({ stage: p('retrieval'), status: 'start' });
const t1 = Date.now();
const retrievalBucketCount = getBucketCount(brainstormCount);
const perBucketCount = Math.ceil(brainstormCount / retrievalBucketCount);
const retrievalBuckets = await Promise.all(
Array.from({ length: retrievalBucketCount }, () =>
runRetrieval(interpreterOutput, perBucketCount, mediaType, useWebSearch, useHardRequirements, previousFullMatches)
)
);
const allCandidates = retrievalBuckets.flatMap((r) => r.candidates);
const dedupedCandidates = deduplicateCandidates(allCandidates, allSeenTitles);
log(recId, `${stagePrefix}Retrieval: done (${Date.now() - t1}ms) — ${dedupedCandidates.length} candidates (${retrievalBucketCount} buckets, ${allCandidates.length} before dedup)`, {
titles: dedupedCandidates.map((c) => c.title),
});
sseWrite({ stage: p('retrieval'), status: 'done', data: { candidates: dedupedCandidates } });
// --- Validator (optional) ---
let candidatesForRanking = dedupedCandidates;
if (useValidator) {
log(recId, `${stagePrefix}Validator: start`);
sseWrite({ stage: p('validator'), status: 'start' });
const tV = Date.now();
const validatorOutput = await runValidator(dedupedCandidates, mediaType);
const verified = validatorOutput.candidates.filter((c) => !c.isTrash);
const trashCount = validatorOutput.candidates.length - verified.length;
candidatesForRanking = verified.map(({ title, reason }) => ({ title, reason }));
log(recId, `${stagePrefix}Validator: done (${Date.now() - tV}ms) — removed ${trashCount} trash entries`);
sseWrite({ stage: p('validator'), status: 'done', data: { removed: trashCount } });
} else {
sseWrite({ stage: p('validator'), status: 'done', data: { skipped: true } });
}
// --- Ranking (bucketed) ---
log(recId, `${stagePrefix}Ranking: start`);
sseWrite({ stage: p('ranking'), status: 'start' });
const t2 = Date.now();
const rankBucketCount = getBucketCount(candidatesForRanking.length);
const candidateBuckets = splitIntoBuckets(candidatesForRanking, rankBucketCount);
const rankingBuckets = await Promise.all(
candidateBuckets.map((bucket) =>
runRanking(interpreterOutput, { candidates: bucket }, mediaType, useHardRequirements)
)
);
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),
};
log(recId, `${stagePrefix}Ranking: done (${Date.now() - t2}ms) — ${rankBucketCount} buckets`, {
full_match: rankingOutput.full_match.length,
definitely_like: rankingOutput.definitely_like.length,
might_like: rankingOutput.might_like.length,
questionable: rankingOutput.questionable.length,
will_not_like: rankingOutput.will_not_like.length,
});
sseWrite({ stage: p('ranking'), status: 'done', data: rankingOutput });
// --- Curator (bucketed) ---
log(recId, `${stagePrefix}Curator: start`);
sseWrite({ stage: p('curator'), status: 'start' });
const t3 = Date.now();
type CategorizedItem = { title: string; category: keyof RankingOutput };
const categorizedItems: CategorizedItem[] = [
...rankingOutput.full_match.map((t) => ({ title: t, category: 'full_match' as const })),
...rankingOutput.definitely_like.map((t) => ({ title: t, category: 'definitely_like' as const })),
...rankingOutput.might_like.map((t) => ({ title: t, category: 'might_like' as const })),
...rankingOutput.questionable.map((t) => ({ title: t, category: 'questionable' as const })),
...rankingOutput.will_not_like.map((t) => ({ title: t, category: 'will_not_like' as const })),
];
const curatorBucketCount = getBucketCount(categorizedItems.length);
const curatorItemBuckets = splitIntoBuckets(categorizedItems, curatorBucketCount);
const curatorBucketRankings: RankingOutput[] = curatorItemBuckets.map((bucket) => ({
full_match: bucket.filter((i) => i.category === 'full_match').map((i) => i.title),
definitely_like: bucket.filter((i) => i.category === 'definitely_like').map((i) => i.title),
might_like: bucket.filter((i) => i.category === 'might_like').map((i) => i.title),
questionable: bucket.filter((i) => i.category === 'questionable').map((i) => i.title),
will_not_like: bucket.filter((i) => i.category === 'will_not_like').map((i) => i.title),
}));
const curatorBucketOutputs = await Promise.all(
curatorBucketRankings.map((ranking) =>
runCurator(ranking, interpreterOutput, mediaType, useWebSearch)
)
);
const curatorOutput = curatorBucketOutputs.flat();
log(recId, `${stagePrefix}Curator: done (${Date.now() - t3}ms) — ${curatorOutput.length} items curated (${curatorBucketCount} buckets)`);
sseWrite({ stage: p('curator'), status: 'done', data: curatorOutput });
return curatorOutput;
}
export async function runPipeline(
rec: RecommendationRecord,
sseWrite: (event: SSEEvent) => void,
@@ -58,8 +182,11 @@ export async function runPipeline(
const startTime = Date.now();
const mediaType = (rec.media_type ?? 'tv_show') as MediaType;
const useWebSearch = rec.use_web_search ?? false;
const useValidator = rec.use_validator ?? false;
const useHardRequirements = rec.hard_requirements ?? false;
const selfExpansive = rec.self_expansive ?? false;
log(rec.id, `Starting pipeline for "${rec.title}" [${mediaType}${useWebSearch ? ', web_search' : ''}]${feedbackContext ? ' (with feedback context)' : ''}`);
log(rec.id, `Starting pipeline for "${rec.title}" [${mediaType}${useWebSearch ? ', web_search' : ''}${useValidator ? ', validator' : ''}${useHardRequirements ? ', hard_req' : ''}${selfExpansive ? `, expansive×${rec.expansive_passes}(${rec.expansive_mode})` : ''}]${feedbackContext ? ' (with feedback context)' : ''}`);
try {
// Set status to running
@@ -91,84 +218,69 @@ export async function runPipeline(
});
sseWrite({ stage: 'interpreter', status: 'done', data: interpreterOutput });
// --- Retrieval (bucketed) ---
// --- Pass 1: Retrieval → [Validator?] → Ranking → Curator ---
currentStage = 'retrieval';
log(rec.id, 'Retrieval: start');
sseWrite({ stage: 'retrieval', status: 'start' });
const t1 = Date.now();
const retrievalBucketCount = getBucketCount(rec.brainstorm_count);
const perBucketCount = Math.ceil(rec.brainstorm_count / retrievalBucketCount);
const retrievalBuckets = await Promise.all(
Array.from({ length: retrievalBucketCount }, () =>
runRetrieval(interpreterOutput, perBucketCount, mediaType, useWebSearch)
)
);
const allCandidates = retrievalBuckets.flatMap((r) => r.candidates);
const dedupedCandidates = deduplicateCandidates(allCandidates);
const retrievalOutput = { candidates: dedupedCandidates };
log(rec.id, `Retrieval: done (${Date.now() - t1}ms) — ${dedupedCandidates.length} candidates (${retrievalBucketCount} buckets, ${allCandidates.length} before dedup)`, {
titles: dedupedCandidates.map((c) => c.title),
const allSeenTitles = new Set<string>();
const pass1Output = await runSubPipeline({
recId: rec.id,
interpreterOutput,
mediaType,
useWebSearch,
useValidator,
useHardRequirements,
brainstormCount: rec.brainstorm_count,
previousFullMatches: [],
allSeenTitles,
stagePrefix: '',
sseWrite: (event) => {
currentStage = event.stage;
sseWrite(event);
},
});
sseWrite({ stage: 'retrieval', status: 'done', data: retrievalOutput });
// --- Ranking (bucketed) ---
currentStage = 'ranking';
log(rec.id, 'Ranking: start');
sseWrite({ stage: 'ranking', status: 'start' });
const t2 = Date.now();
const rankBucketCount = getBucketCount(dedupedCandidates.length);
const candidateBuckets = splitIntoBuckets(dedupedCandidates, rankBucketCount);
const rankingBuckets = await Promise.all(
candidateBuckets.map((bucket) =>
runRanking(interpreterOutput, { candidates: bucket }, mediaType)
)
);
const rankingOutput: RankingOutput = {
full_match: rankingBuckets.flatMap((r) => r.full_match),
definitely_like: rankingBuckets.flatMap((r) => r.definitely_like),
might_like: rankingBuckets.flatMap((r) => r.might_like),
questionable: rankingBuckets.flatMap((r) => r.questionable),
will_not_like: rankingBuckets.flatMap((r) => r.will_not_like),
};
log(rec.id, `Ranking: done (${Date.now() - t2}ms) — ${rankBucketCount} buckets`, {
full_match: rankingOutput.full_match.length,
definitely_like: rankingOutput.definitely_like.length,
might_like: rankingOutput.might_like.length,
questionable: rankingOutput.questionable.length,
will_not_like: rankingOutput.will_not_like.length,
});
sseWrite({ stage: 'ranking', status: 'done', data: rankingOutput });
let mergedOutput = pass1Output;
// --- Curator (bucketed) ---
currentStage = 'curator';
log(rec.id, 'Curator: start');
sseWrite({ stage: 'curator', status: 'start' });
const t3 = Date.now();
type CategorizedItem = { title: string; category: keyof RankingOutput };
const categorizedItems: CategorizedItem[] = [
...rankingOutput.full_match.map((t) => ({ title: t, category: 'full_match' as const })),
...rankingOutput.definitely_like.map((t) => ({ title: t, category: 'definitely_like' as const })),
...rankingOutput.might_like.map((t) => ({ title: t, category: 'might_like' as const })),
...rankingOutput.questionable.map((t) => ({ title: t, category: 'questionable' as const })),
...rankingOutput.will_not_like.map((t) => ({ title: t, category: 'will_not_like' as const })),
];
const curatorBucketCount = getBucketCount(categorizedItems.length);
const curatorItemBuckets = splitIntoBuckets(categorizedItems, curatorBucketCount);
const curatorBucketRankings: RankingOutput[] = curatorItemBuckets.map((bucket) => ({
full_match: bucket.filter((i) => i.category === 'full_match').map((i) => i.title),
definitely_like: bucket.filter((i) => i.category === 'definitely_like').map((i) => i.title),
might_like: bucket.filter((i) => i.category === 'might_like').map((i) => i.title),
questionable: bucket.filter((i) => i.category === 'questionable').map((i) => i.title),
will_not_like: bucket.filter((i) => i.category === 'will_not_like').map((i) => i.title),
}));
const curatorBucketOutputs = await Promise.all(
curatorBucketRankings.map((ranking) =>
runCurator(ranking, interpreterOutput, mediaType, useWebSearch)
)
);
const curatorOutput = curatorBucketOutputs.flat();
log(rec.id, `Curator: done (${Date.now() - t3}ms) — ${curatorOutput.length} items curated (${curatorBucketCount} buckets)`);
sseWrite({ stage: 'curator', status: 'done', data: curatorOutput });
// --- Self Expansive: extra passes ---
if (selfExpansive && rec.expansive_passes > 0) {
const allFullMatches = pass1Output
.filter((c) => c.category === 'Full Match')
.map((c) => c.title);
for (let i = 0; i < rec.expansive_passes; i++) {
const passNum = i + 2;
const passCount = rec.expansive_mode === 'extreme' ? rec.brainstorm_count : 60;
const passPrefix = `pass${passNum}:` as const;
log(rec.id, `Self Expansive Pass ${passNum}: start (${passCount} candidates, ${allFullMatches.length} full matches as context)`);
currentStage = `${passPrefix}retrieval` as SSEEvent['stage'];
const passOutput = await runSubPipeline({
recId: rec.id,
interpreterOutput,
mediaType,
useWebSearch,
useValidator,
useHardRequirements,
brainstormCount: passCount,
previousFullMatches: [...allFullMatches],
allSeenTitles,
stagePrefix: passPrefix,
sseWrite: (event) => {
currentStage = event.stage;
sseWrite(event);
},
});
mergedOutput = mergeCuratorOutputs(mergedOutput, passOutput);
const newFullMatches = passOutput
.filter((c) => c.category === 'Full Match')
.map((c) => c.title);
allFullMatches.push(...newFullMatches);
log(rec.id, `Self Expansive Pass ${passNum}: done — ${passOutput.length} new items, ${mergedOutput.length} total`);
}
}
// Generate AI title
let aiTitle: string = rec.title;
@@ -180,17 +292,27 @@ export async function runPipeline(
log(rec.id, `Title generation failed, keeping initial title: ${String(err)}`);
}
// Sort by category order before saving
const CATEGORY_ORDER: Record<string, number> = {
'Full Match': 0,
'Definitely Like': 1,
'Might Like': 2,
'Questionable': 3,
'Will Not Like': 4,
};
mergedOutput.sort((a, b) => (CATEGORY_ORDER[a.category] ?? 99) - (CATEGORY_ORDER[b.category] ?? 99));
// Save results to DB
log(rec.id, 'Saving results to DB');
await db
.update(recommendations)
.set({ recommendations: curatorOutput, status: 'done', title: aiTitle })
.set({ recommendations: mergedOutput, status: 'done', title: aiTitle })
.where(eq(recommendations.id, rec.id));
sseWrite({ stage: 'complete', status: 'done', data: { title: aiTitle } });
log(rec.id, `Pipeline complete (total: ${Date.now() - startTime}ms)`);
return curatorOutput;
return mergedOutput;
} catch (err) {
const message = err instanceof Error ? err.message : String(err);
log(rec.id, `Pipeline error at stage "${currentStage}": ${message}`);