I want a recommendation system for a retail site that can do similar items, cart add-ons, and email cross-sells while excluding out-of-stock products.

Here’s a practical recommendation-system design for a retail site that supports:

• similar items
• cart add-ons
• email cross-sells
• excluding out-of-stock products

1) Core idea

Use a hybrid recommender:

• Candidate generation: pull likely products from multiple sources
• Filtering layer: remove out-of-stock and other business-rule violations
• Ranking layer: order the remaining items by relevance for the context

This gives you flexibility for different placements:

• PDP “similar items”
• Cart “add-ons”
• Email “you may also like”

---

2) Data you’ll need

Product catalog

• product_id
• title
• description
• category / subcategory
• brand
• attributes: color, size, material, etc.
• price
• margin
• image embeddings or text embeddings
• stock status / available inventory
• lifecycle status: active, discontinued, hidden

Behavioral data

• views, clicks, add-to-cart, purchases
• co-occurrence: items often bought together
• session sequences
• user affinities if available

Context data

• page type: PDP, cart, email
• current product(s)
• cart contents
• user segment / lifecycle stage
• inventory and shipping constraints

---

3) Recommendation types

A. Similar items

Best for product detail pages.

Signals:

• same category/subcategory
• similar embeddings from title/description/images
• similar price band
• similar brand or alternative brands
• high click similarity

Ranking examples:

• cosine similarity on item embeddings
• weighted similarity on structured attributes
• boost items with better availability or conversion rate

Good logic:

• “near substitutes” first
• “same category alternatives” next
• “premium/down-market variants” next

---

B. Cart add-ons

Best for cart or mini-cart.

Goal:

• increase basket size with complementary products

Signals:

• frequent co-purchase pairs
• category complements
• accessory relationships
• margin and attach rate
• delivery feasibility and stock

Examples:

• headphones → case, cable, adapter
• coffee machine → filters, descaler, pods
• phone → screen protector, charger

Ranking:

• co-purchase strength
• complement category match
• availability
• price compatibility with cart

---

C. Email cross-sells

Best for post-purchase or browse-abandon emails.

Goal:

• personalized cross-sell based on recent activity or purchase history

Signals:

• purchased items
• viewed items
• similar-category follow-ups
• lifecycle timing
• recipient segment

Examples:

• after buying a camera: memory card, bag, tripod
• after buying running shoes: socks, shorts, hydration belt

Ranking:

• historical purchase affinities
• user preferences
• category expansion
• margin and inventory

---

4) Excluding out-of-stock products

This should be enforced in the filtering layer, not just the ranking layer.

Hard filters

Exclude products that are:

• out of stock
• backordered, if you don’t want to show them
• inactive/discontinued
• not shippable to user location
• restricted by size/color availability

Optional availability scoring

If you want to be smarter than a binary filter:

• boost high-stock items
• deprioritize low-stock items
• avoid items likely to go out of stock soon

But the simplest rule is:

• never recommend items with zero available inventory

---

5) Suggested architecture

Step 1: Candidate generation

Generate 50–500 candidates per context from:

• embedding similarity search
• co-purchase graph
• category rules
• editorial/business lists

Step 2: Business-rule filtering

Remove products that are:

• out of stock
• inactive
• invalid for channel or region
• already in cart, if you don’t want duplicates

Step 3: Ranking model

Rank the filtered items using:

• context features
• item features
• user features
• interaction features

A learning-to-rank model like:

• XGBoost / LightGBM
• neural ranker
• two-tower retrieval + ranker

works well.

Step 4: Re-ranking

Apply final constraints:

• diversity across categories/brands
• price range distribution
• no duplicates
• business priorities

---

6) A good practical scoring formula

For a candidate item i in context c:

score(i, c) =
  w1 * similarity_score
+ w2 * co_purchase_score
+ w3 * complement_score
+ w4 * user_affinity_score
+ w5 * margin_score
+ w6 * availability_score
- w7 * duplicate_penalty

Then apply:

if stock <= 0: exclude
if inactive: exclude
if region_not_shippable: exclude

For email cross-sells, replace some of the context weights with:

• recency of purchase
• purchase category transitions
• lifecycle timing

---

7) Example rules by placement

Similar items on PDP

• Only same or closely related category
• Must be in stock
• Prefer same brand and adjacent price points
• Show 4–8 items

Cart add-ons

• Must be complementary, not substitutes
• Must be in stock
• Avoid items already in cart
• Prefer low-friction add-ons under a threshold price

Email cross-sells

• In stock at send time
• Personalized by recent purchase
• Exclude items already purchased recently
• Optionally create fallback categories if no good matches

---

8) How to keep recommendations fresh

• refresh inventory in near real-time or hourly
• retrain behavioral models regularly
• update co-purchase stats weekly/daily
• maintain fallback rules when model confidence is low

---

9) Metrics to track

Offline

• precision@k
• recall@k
• MAP / NDCG
• attach-rate prediction quality
• stock-filter pass rate

Online

• CTR
• add-to-cart rate
• revenue per session/email
• conversion rate
• average order value
• out-of-stock impression rate
• email revenue per recipient

A key guardrail:

• zero or near-zero out-of-stock impressions

---

10) Best implementation pattern

If you want a strong first version:

1. Build item embeddings from title, description, image, category
2. Build a co-purchase graph from order history
3. Create explicit complement rules for cart/email
4. Add a hard inventory filter
5. Rank with a simple weighted model
6. Test and refine with A/B experiments

---

11) Recommendation for your use case

I’d suggest this setup:

• Retrieval
  • item embeddings for similar items
  • co-purchase graph for add-ons and cross-sells
• Filtering
  • hard stock exclusion
  • region/shipping exclusion
• Ranking
  • context-aware learning-to-rank
• Re-ranking
  • diversity + business constraints

This will cover all three use cases cleanly.

If you want, I can also provide:

1. a system architecture diagram,
2. a database/schema design, or
3. sample pseudocode / SQL / Python for the recommender.