Use Case: AI Recommendations

Scenario: Netflix/Spotify style "Recommended for You". Traditional: Collaborative Filtering (Matrix Factorization). Hard to implement. AI Way: Vector Similarity.

Concept: User Embeddings

1. Item Embedding: Turn every movie into a vector (based on description/genre).
2. User History: "User watched Matrix, Inception, Interstellar".
3. User Embedding: Average the vectors of the movies they watched. UserVector = (Matrix + Inception + Interstellar) / 3

Implementation

1. Update User Profile

Every time they "Like" an item:

async function onLike(userId, movieId) {
  const movieVector = await getVector(movieId);
  const userVector = await getUserVector(userId);
  
  // Moving Average update
  const newVector = (userVector * 0.9) + (movieVector * 0.1);
  
  await saveUserVector(userId, newVector);
}

2. Generate Feed

Query the Vector DB using the User Vector.

const recommendations = await pinecone.query({
  vector: userVector,
  topK: 10,
  filter: { id: { $nin: watchedMovieIds } } // Exclude watched
});

Cold Start Problem

What if the user is new? Solution: Ask 3 questions during onboarding. "What genres do you like?" -> Generate initial vector from that text.