Epic 6 — Discover Feed (AI Matching)

---

BE-DISCOVER-001 — AI Matching service (9-factor scoring)

• [x] Implemented

Files:

• Create: src/services/ai/matching.ts — full implementation replacing stub
• Edit: src/graphql/resolvers/job.ts — discoverFeed calls real matching service

9 scoring factors (per spec doc):

1. Salary alignment — compare job.paymentAmount range vs talent's expected rate
2. Work authorization — match job.requirements.workAuthorization vs talent.workAuthorization
3. Physical attributes — match age range, gender, height, body type if job specifies
4. Notable works relevance — semantic similarity (vector search) between talent's works and job genre/type
5. Production experience — match production type + subtype experience
6. Awards & nominations — presence and relevance
7. Dream gig alignment — semantic match of talent.dreamGigs vs job description
8. Language/accent proficiency — match required languages/accents
9. Availability — talent.availableFrom vs job.startDate

Dev mode: Falls back to weighted DB filter. Production: calls Qdrant (self-hosted, same Hetzner machine) for factors 4, 7.
Returns score (0–1) and matchReasons: string[] (human-readable explanation per matched factor).

---

BE-DISCOVER-002 — Vector embedding pipeline for talents

• [x] Implemented

Files:

• Create: src/services/ai/embeddings.ts — OpenAI text-embedding-3-small calls
• Create: src/services/ai/vectorStore.ts — Qdrant upsert/query abstraction
• Edit: src/services/talent/index.ts — call embeddings.indexTalent(profile) after profile updates

Description: When a talent profile is created or updated, generate a text embedding from: displayName + primaryCategory + skills + experienceEntries + notableWorks + dreamGigs. Upsert into Qdrant with talentId as the key. The discover feed uses these embeddings for factors 4 and 7.

Infrastructure: Qdrant runs as a Docker container on the same Hetzner machine as the backend. Requires ~200 MB RAM idle — fits comfortably on the existing CX22 (4 GB RAM) alongside the backend and face-comparison sidecar. No separate node needed.

Environment variables:

QDRANT_URL=http://localhost:6333   # internal — not exposed to the internet
OPENAI_API_KEY=<key>               # for text-embedding-3-small

Coolify setup: Add a Qdrant service to the existing Docker Compose stack:

qdrant:
  image: qdrant/qdrant:latest
  ports:
    - "6333:6333"
  volumes:
    - qdrant_data:/qdrant/storage
  restart: unless-stopped

---

BE-DISCOVER-003 — Extend AI matching to use new structured fields

• [x] Implemented

Files:

• Edit: src/services/ai/matching.ts — update factors 1, 2, 8 to use new field data
• Edit: src/__tests__/services/matching.test.ts

Factor updates:

Factor	Before	After
1 — Salary alignment	job.paymentAmount vs rough talent expectation	job.paymentAmount vs talent.salaryExpectation.{ min, max, currency, period } — normalise to same period before comparing
2 — Work authorization	job.requirements.workAuthorization[] vs talent.workAuthorization[]	+ talent.passportVisa.{ passports[], visas[], workPermits[] } — broader match surface
8 — Language/accent	language match only (accent field didn't exist)	+ talent.accentDialect[] matched against job.requirements.languages[].accent

No schema or API changes — purely service logic update.

---

BE-SEARCH-001 — Extend talent search filters (notable works, awards, dream job)

• [x] Implemented

Files:

• Edit: src/graphql/schema/index.ts — extend TalentSearchFilters input type
• Edit: src/services/talent/search.ts (or wherever searchTalents is resolved) — add filter logic
• Edit: src/__tests__/resolvers/search.test.ts

New filter fields:

input TalentSearchFilters {
  # ...existing fields...
  hasNotableWork:  String    # partial match on notableWorks[].title
  hasAward:        Boolean   # talent has at least one award entry
  dreamJobKeyword: String    # semantic match against dreamJobs[].title
}

Implementation:

• hasNotableWork: PostgreSQL JSON_ARRAY_ELEMENTS + ILIKE on notableWorks JSON array.
• hasAward: WHERE awards IS NOT NULL AND awards != '[]'.
• dreamJobKeyword: vector similarity search via Qdrant — embed the keyword and compare to talent dream-job embeddings. Falls back to JSON ILIKE in dev mode.

---

FE-DISCOVER-001 — Discover feed UI with swipe stack (producer)

• [x] Implemented

Files:

• Edit: apps/app/src/pages/discover/DiscoverPage.tsx — replace placeholder
• Create: apps/app/src/lib/queries/discover.ts — DISCOVER_FEED_QUERY
• Create: apps/app/src/hooks/useDiscoverFeed.ts
• Create: apps/app/src/hooks/useSwipeTalent.ts

Description: Replace "Home feed coming soon" with the real discover feed. Shows a stack of TalentCard components (from DS). Swipe right → shortlist, swipe left → skip, tap × → block. After each swipe, call swipeTalent mutation and advance to next card. When stack runs low (< 3 cards), pre-fetch more. Shows EmptyState when no more talents to show. Matches "Talento Fluxo" / "Produtor Fluxo" frames in Figma.

Note: This view is producer-only. Talent users see a different "home" (job listings or notifications feed). Route guard: redirect talents to /jobs.

---

FE-DISCOVER-002 — Talent home feed (job recommendations)

• [x] Implemented

Files:

• Create: apps/app/src/pages/discover/TalentHomePage.tsx
• Edit: apps/app/src/pages/discover/DiscoverPage.tsx — switch view by role

Description: Talent users' home view: personalized job recommendations shown as JobCard list, ranked by relevance to their profile. Uses jobs query with talent-profile-based pre-filters. Matches the talent's "Jornada" home screen in Figma.