Step 1: QUIZ
  User answers 5 questions
  Claude Haiku processes Q&A pairs
  → Returns structured taste profile JSON

Step 2: EMBED
  Taste profile text → OpenAI text-embedding-3-small
  → 1536-dimensional float vector

Step 3: STORE
  Vector saved to taste_profiles.embedding (pgvector column)
  Upserted on every quiz completion

Step 4: SEARCH
  pgvector cosine similarity:
  SELECT ... ORDER BY embedding <=> $userVector LIMIT 20

Step 5: RE-RANK
  Claude Sonnet receives top-20 candidates + context:
  - Current stock levels
  - Season / time of year
  - User subscription tier
  - Order history (avoid repeats)
  → Returns ranked list of 4-8 product IDs

Step 6: FEEDBACK
  User interactions feed back into profile:
  - Click-through → mild positive signal
  - Cart add → strong positive signal
  - Purchase → strongest signal
  - Review rating → direct quality signal
  Weekly BullMQ cron refreshes embeddings for active users

You are a specialty coffee taste profile analyser for Brewhaus.
Given a user's quiz answers, return a JSON taste profile.
Respond ONLY with valid JSON — no markdown, no explanation.

Based on these coffee quiz answers, build a taste profile:

Q: How do you usually brew your coffee?
A: {answer}

Q: What flavours do you love most?
A: {answer}

Q: How do you take your coffee?
A: {answer}

Q: When do you drink your first coffee?
A: {answer}

Q: What's your adventure level with coffee?
A: {answer}

Return exactly:
{
  "flavorNotes":    ["note1", "note2", "note3"],
  "roastPreference": "light" | "medium" | "dark",
  "brewMethod":      "string",
  "adventureLevel":  "beginner" | "enthusiast" | "connoisseur",
  "summary":         "one sentence description"
}

interface TasteProfile {
  flavorNotes:    string[]
  roastPreference: "light" | "medium" | "dark"
  brewMethod:     string
  adventureLevel: "beginner" | "enthusiast" | "connoisseur"
  summary:        string
}

Write a personalised brew guide for {productName} from {origin}.

Product details:
- Flavor notes: {flavorNotes}
- Process: {process}
- Roast level: {roastLevel}
- Altitude: {altitude}

User's preferred brew method: {brewMethod}

Requirements:
- Specific water temperature, dose, yield, and timing
- One pro tip specific to this origin
- Max 200 words
- Friendly expert tone — like a barista friend

You are ranking coffee products for a customer based on their profile.

Customer profile: {JSON.stringify(tasteProfile)}
Customer tier: {subscriptionPlan}
Previous purchases: {previousProductIds}
Current date: {month} (affects seasonal recommendations)

Candidate products (ordered by vector similarity):
{candidates.map(p => `- ${p.id}: ${p.name} (${p.origin}, ${p.roastLevel}, ${p.flavorNotes})`).join('\n')}

Stock status: {stockMap}

Return a JSON array of 6 product IDs in your recommended order.
Exclude any out-of-stock products.
Avoid products the customer recently purchased.
For Collective subscribers, prioritise rare and exclusive lots.
Respond ONLY with a JSON array: ["id1", "id2", ...]

{name}. {description}. {origin}. {process}. {roastLevel}. 
{varietal}. Flavor notes: {flavorNotes.join(', ')}.

Flavor preferences: {flavorNotes.join(', ')}.
Roast preference: {roastPreference}.
Brew method: {brewMethod}.
Experience level: {adventureLevel}.
{summary}

-- <=> operator = cosine distance (0 = identical, 2 = opposite)
SELECT
  id,
  name,
  1 - (embedding <=> $1::vector) AS similarity
FROM products
WHERE
  in_stock = TRUE
  AND category = 'BEANS'
  AND embedding IS NOT NULL
ORDER BY embedding <=> $1::vector
LIMIT 20;

// Runs every Sunday at 2am
const embeddingRefreshQueue = new Queue("embedding-refresh", { connection });

embeddingRefreshQueue.add(
  "refresh-active-users",
  {},
  {
    repeat: { cron: "0 2 * * 0" },
  }
);

// Worker: re-embeds taste profiles for users active in last 30 days
// Picks up new reviews, purchase history, rating signals