class OnDeviceRecommender { private let userProfileKey = "user_embedding_v2" private var itemIndex: [(id: String, embedding: [Float])] = [] func loadItemIndex(from url: URL) { let data = try! Data(contentsOf: url) itemIndex = try! JSONDecoder().decode([(id: String, embedding: [Float])].self, from: data) } func getRecommendations(count: Int) -> [String] { guard let profileData = UserDefaults.standard.data(forKey: userProfileKey), let profile = try? JSONDecoder().decode([Float].self, from: profileData) else { return popularItemIds(count: count) } return itemIndex .map { item in (item.id, cosineSimilarity(profile, item.embedding)) } .sorted { $0.1 > $1.1 } .prefix(count) .map { $0.0 } } private func cosineSimilarity(_ a: [Float], _ b: [Float]) -> Float { zip(a, b).map(*).reduce(0, +) } } 

Embeddings in the index are updated on start or on a schedule. We use precomputed vectors from the server, which minimizes device load.

Core system: TF-IDF and text embeddings

For articles, descriptions, news — two approaches: TF-IDF for speed, sentence embeddings for quality. In practice, we use sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2 — 278 MB, supports Russian. Each item becomes a 384-dimensional vector. According to TF-IDF, word frequency and inverse document frequency provide a simple but effective relevance measure.

from sentence_transformers import SentenceTransformer model = SentenceTransformer('paraphrase-multilingual-MiniLM-L12-v2') def embed_article(article: Article) -> np.ndarray: text = f"{article.title}. {article.description}. {' '.join(article.tags)}" return model.encode(text, normalize_embeddings=True) def similarity(v1: np.ndarray, v2: np.ndarray) -> float: return float(np.dot(v1, v2)) 

User profile — moving average

The user profile is a weighted average of the embeddings of content they interacted with. Recent interactions weigh more (exponential decay with coefficient 0.9):

def update_user_profile(profile: np.ndarray, new_item_embedding: np.ndarray, interaction_weight: float, decay: float = 0.9) -> np.ndarray: updated = decay * profile + (1 - decay) * interaction_weight * new_item_embedding return updated / np.linalg.norm(updated) 

Structured metadata: not just text

For product catalogs, text embeddings are supplemented with categorical features: category, brand, price range, color. The final vector is a concatenation of a normalized text embedding and one-hot/ordinal features with weights 0.7 and 0.3 respectively:

def build_item_vector(item: Product) -> np.ndarray: text_emb = embed_text(f"{item.name} {item.description}") cat_features = encode_categorical({ 'category': item.category_id, 'brand': item.brand_id, 'price_range': bucket_price(item.price) }) return np.concatenate([text_emb * 0.7, cat_features * 0.3]) 

Comparison: Content-Based vs Collaborative

How to implement on-device CB on Swift: step-by-step

1. Content indexing. On the server, compute embeddings for each item using sentence-transformers. Save to a JSON file with fields id and embedding.
2. Index loading. On app launch, load the JSON into the OnDeviceRecommender array. For speed, use a binary format.
3. Collect interactions. Each time a user opens or likes content, save the item embedding in UserDefaults with a timestamp.
4. Update profile. On each interaction, recompute the moving average with exponential decay.
5. Generate recommendations. Call getRecommendations when displaying the feed or in the background.

What is included in the work

We provide:

• Analysis of content structure and available metadata
• Selection of the optimal embedding model for language and domain
• Building the index and profile update mechanism
• Implementation of on-device or server-side solution
• Integration with existing backend
• Documentation and team training

Our experience: over 10 years in mobile development, 30+ projects with AI recommendations. Guaranteed support after deployment. Contact us for a project evaluation — we will select the optimal architecture for your case. Order the development of a recommendation system for your project.

Time estimates

Get a consultation: tell us about your content and user scenarios, and we will propose a solution.