Project 3: Building a Recommendation System

Search is about finding what you know you want. Recommendations are about finding what you don't know you want yet. In this project, you will build a "Discover" feature for an app, using vector proximity to find similar items based on user behavior.

---

1. Project Requirements

• Data: A CSV of items (Movies, Products, or Articles) with tags and descriptions.
• Goal: Implement "More Like This" and "Recommended for You" logic.

---

2. Architecture: Two Approaches

Path A: Item-to-Item (Similarity Search)

1. User is looking at "Product A".
2. Get the vector for "Product A".
3. Query the DB for the Top 5 most similar vectors (excluding Product A).
4. Display: "Customers who viewed this also liked..."

Path B: User-to-Item (Profile Embedding)

1. Look at the last 5 items a user liked.
2. Mean Vector: Average the vectors of those 5 items to create a "User Interest Vector."
3. Query the DB for items similar to the User Interest Vector.
4. Display: "Recommended for You".

---

3. Implementation (Python)

import numpy as np

def push_recommendations(user_likes_ids, collection):
    # 1. Fetch vectors for items the user liked
    liked_items = collection.get(ids=user_likes_ids, include=['embeddings'])
    vectors = liked_items['embeddings']
    
    # 2. Compute the Average (The User's 'Taste' Vector)
    user_taste_vec = np.mean(vectors, axis=0).tolist()
    
    # 3. Find items close to that average
    recommendations = collection.query(
        query_embeddings=[user_taste_vec],
        n_results=10,
        # Don't recommend things they've already seen!
        where={"id": {"$nin": user_likes_ids}}
    )
    
    return recommendations

---

4. Evaluation Criteria

• Diversity: Does the recommender show a variety of items, or just the same item 5 times?
• Relevance: Are the recommendations logically connected to the user's past likes?
• Latency: Does the "Taste Vector" calculation and search happen in under 150ms?

---

Deliverables

1. A recommendation script that takes a "User History" list and outputs suggestions.
2. A short paragraph on how you handled "Cold Start" (users with zero past likes).

---

Building recommendation engines is the 'Secret Sauce' of modern platforms. Let's get started.