Two-Tower and FAISS: AI Recommendation System for Personalization

A new user enters a marketplace — the system doesn't know their tastes. They see popular products, but conversion drops. This is a typical cold-start. The Two-Tower + FAISS architecture is a proven production-ready approach that delivers CTR of 8-15% and additional revenue of 12-25%. We have implemented dozens of such projects, from startups to enterprise.

Let's consider the key problems: cold-start, scaling to millions of transactions, balancing accuracy and latency. Each requires a separate solution. Production-ready development of an AI recommendation system is built on a Two-Tower neural network and a FAISS index.

How Two-Tower solves the cold start problem?

Cold start for new users is a typical pain point. Without purchase history, the system displays random items, conversion drops. Our approach uses semantic embeddings of items and demographic features to provide relevant recommendations from the first visit. Matrix Factorization works, but only for users with history — for cold start, content features are needed.

We use a hybrid approach: content-based filtering based on text embeddings (BERT, 768-dim) for new users, and after accumulating 3-5 interactions, we switch to collaborative filtering. Recall@10 for cold users reaches 20% vs 5% for pure MF.

Why is Two-Tower architecture more effective than content-based?

Two-tower allows learning user and item embeddings in a single space, training on implicit feedback (clicks, views). This yields CTR of 8-15% vs 2-4% for rule-based systems. On large volumes (>5M transactions), early fusion does not scale — two-tower with in-batch negatives trains in 2-3 hours on 8 V100.

Comparison of metrics before and after implementation

As a result of implementation, additional revenue increases by 12-25%, which for a large e-commerce platform can amount to millions of rubles per month.

Technical implementation: Two-Tower, training and inference

Two-Tower Architecture

import torch
import torch.nn as nn
import numpy as np
from torch.utils.data import DataLoader, Dataset

class UserTower(nn.Module):
    """User encoder"""
    def __init__(self, n_users: int, n_categories: int,
                 embedding_dim: int = 64, hidden_dim: int = 128):
        super().__init__()
        self.user_emb = nn.Embedding(n_users + 1, embedding_dim, padding_idx=0)
        self.category_emb = nn.Embedding(n_categories + 1, 16, padding_idx=0)

        self.mlp = nn.Sequential(
            nn.Linear(embedding_dim + 16 * 5 + 10, hidden_dim),
            nn.ReLU(),
            nn.Dropout(0.3),
            nn.Linear(hidden_dim, 64),
            nn.LayerNorm(64)
        )

    def forward(self, user_id, top_categories, behavior_features):
        user_vec = self.user_emb(user_id)
        cat_vecs = self.category_emb(top_categories).view(top_categories.shape[0], -1)
        combined = torch.cat([user_vec, cat_vecs, behavior_features], dim=1)
        return self.mlp(combined)


class ItemTower(nn.Module):
    """Item encoder"""
    def __init__(self, n_items: int, n_categories: int,
                 embedding_dim: int = 64, text_dim: int = 128):
        super().__init__()
        self.item_emb = nn.Embedding(n_items + 1, embedding_dim, padding_idx=0)
        self.category_emb = nn.Embedding(n_categories + 1, 16, padding_idx=0)

        self.mlp = nn.Sequential(
            nn.Linear(embedding_dim + 16 + text_dim + 5, 128),
            nn.ReLU(),
            nn.Dropout(0.3),
            nn.Linear(128, 64),
            nn.LayerNorm(64)
        )

    def forward(self, item_id, category_id, text_features, item_features):
        item_vec = self.item_emb(item_id)
        cat_vec = self.category_emb(category_id)
        combined = torch.cat([item_vec, cat_vec, text_features, item_features], dim=1)
        return self.mlp(combined)


class TwoTowerModel(nn.Module):
    def __init__(self, n_users, n_items, n_categories, text_dim=128):
        super().__init__()
        self.user_tower = UserTower(n_users, n_categories)
        self.item_tower = ItemTower(n_items, n_categories, text_dim=text_dim)
        self.temperature = nn.Parameter(torch.ones(1) * 0.05)

    def forward(self, user_inputs, item_inputs):
        user_emb = self.user_tower(**user_inputs)
        item_emb = self.item_tower(**item_inputs)

        # Cosine similarity
        user_norm = nn.functional.normalize(user_emb, dim=1)
        item_norm = nn.functional.normalize(item_emb, dim=1)
        scores = torch.sum(user_norm * item_norm, dim=1) / self.temperature
        return scores

    def get_user_embedding(self, user_inputs) -> torch.Tensor:
        with torch.no_grad():
            return nn.functional.normalize(self.user_tower(**user_inputs), dim=1)

    def get_item_embedding(self, item_inputs) -> torch.Tensor:
        with torch.no_grad():
            return nn.functional.normalize(self.item_tower(**item_inputs), dim=1)

In-Batch Negative Sampling Training

class RecommendationTrainer:
    def __init__(self, model: TwoTowerModel, lr: float = 1e-3):
        self.model = model
        self.optimizer = torch.optim.Adam(model.parameters(), lr=lr)
        self.scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(
            self.optimizer, T_max=10
        )

    def train_epoch(self, dataloader: DataLoader) -> float:
        self.model.train()
        total_loss = 0

        for batch in dataloader:
            user_inputs, pos_item_inputs = batch['user'], batch['positive_item']

            # User embeddings: [batch_size, dim]
            user_embs = nn.functional.normalize(
                self.model.user_tower(**user_inputs), dim=1
            )
            # Item embeddings: [batch_size, dim]
            item_embs = nn.functional.normalize(
                self.model.item_tower(**pos_item_inputs), dim=1
            )

            # In-batch negatives: similarity matrix [batch_size x batch_size]
            scores = torch.matmul(user_embs, item_embs.T) / self.model.temperature

            # Diagonal = positive examples
            labels = torch.arange(len(user_embs)).to(scores.device)
            loss = nn.functional.cross_entropy(scores, labels)

            self.optimizer.zero_grad()
            loss.backward()
            torch.nn.utils.clip_grad_norm_(self.model.parameters(), 1.0)
            self.optimizer.step()

            total_loss += loss.item()

        self.scheduler.step()
        return total_loss / len(dataloader)

ANN Index for Real-Time

import faiss

class RecommendationIndex:
    """Index for fast top-K recommendation search"""

    def __init__(self, dim: int = 64, n_items: int = 100000):
        # IVF-PQ index: speed/accuracy tradeoff
        quantizer = faiss.IndexFlatIP(dim)
        n_lists = min(int(np.sqrt(n_items)), 4096)
        self.index = faiss.IndexIVFPQ(quantizer, dim, n_lists, 8, 8)
        self.index.nprobe = 32  # Number of clusters to search
        self.item_ids = []

    def build(self, item_embeddings: np.ndarray, item_ids: list):
        """Build index from embeddings"""
        embeddings_norm = item_embeddings / np.linalg.norm(
            item_embeddings, axis=1, keepdims=True
        )
        self.index.train(embeddings_norm.astype(np.float32))
        self.index.add(embeddings_norm.astype(np.float32))
        self.item_ids = item_ids

    def recommend(self, user_embedding: np.ndarray,
                   k: int = 50,
                   exclude_ids: list = None) -> list[tuple]:
        """Top-K recommendations for a user"""
        user_norm = user_embedding / np.linalg.norm(user_embedding)
        scores, indices = self.index.search(
            user_norm.reshape(1, -1).astype(np.float32),
            k + (len(exclude_ids) if exclude_ids else 0) + 10
        )

        results = []
        for score, idx in zip(scores[0], indices[0]):
            if idx < 0:
                continue
            item_id = self.item_ids[idx]
            if exclude_ids and item_id in exclude_ids:
                continue
            results.append((item_id, float(score)))
            if len(results) >= k:
                break

        return results

Re-ranking and Business Rules

class RecommendationReranker:
    """Apply business logic to candidates from retrieval model"""

    def rerank(self, candidates: list[tuple],
                user_context: dict,
                business_rules: dict) -> list[tuple]:
        """
        candidates: [(item_id, base_score)]
        business_rules: {'boost_new_items': 1.2, 'boost_on_sale': 1.15, 'max_same_category': 3}
        """
        scored_candidates = []

        category_count = {}
        for item_id, base_score in candidates:
            item = self._get_item_metadata(item_id)
            if item is None:
                continue

            # Apply boosts
            score = base_score
            if business_rules.get('boost_new_items') and item.get('is_new'):
                score *= business_rules['boost_new_items']
            if business_rules.get('boost_on_sale') and item.get('on_sale'):
                score *= business_rules['boost_on_sale']

            # Category limit
            cat = item.get('category')
            max_per_cat = business_rules.get('max_same_category', 5)
            if category_count.get(cat, 0) >= max_per_cat:
                continue
            category_count[cat] = category_count.get(cat, 0) + 1

            scored_candidates.append((item_id, score))

        return sorted(scored_candidates, key=lambda x: x[1], reverse=True)

Typical metrics after implementing a recommendation system: CTR of recommendations 8-15% (vs 2-4% for popular items), conversion from recommendations 3-7%, additional revenue 12-25% of total turnover. Cold start time for new users without history: 3-5 interactions to receive personalized recommendations.

Process, scope, and cost

1. Analytics: data audit, user profiling, defining business goals.
2. Design: choosing architecture (content-based, collaborative, two-tower), training pipelines.
3. Development: implementing models (PyTorch), training on historical data, setting up ANN index (FAISS).
4. Integration: REST API, gRPC, containerization (Docker), deployment in Kubernetes.
5. A/B testing: metrics (Recall, NDCG, revenue share), monitoring data drift.
6. Support: 12 months warranty, model updates, business rules refinement.

What is included in the work

• Architectural document (approach selection, justification)
• Source code for models and pipelines (PyTorch, FAISS)
• Docker images and Kubernetes manifests
• REST API / gRPC for real-time inference
• Integration with your CRM/website
• Documentation and team training
• A/B testing and monitoring setup (Prometheus + Grafana)
• 12 months of support and updates

Timeline: from 30 to 60 working days depending on data volume and integration complexity. Cost is calculated individually — we will evaluate your project in 1 day after reviewing the task.

We have dozens of implemented projects — from startups to enterprise. Additional revenue after implementation amounts to millions of rubles per month. Order the development of an AI recommendation system and get a consultation on your project architecture.

Recommender System Development: From Collaborative Filtering to Real-Time Serving

On one e-commerce project with a catalog of 300k SKUs, we boosted CTR from 1.8% to 4.4% — a 2.4x increase. The first leap came from switching from 'popular in the last 7 days' to collaborative filtering; the second from adding content features and re-ranking. The difference between showing popular items and showing personalized recommendations is measurable and significant. Below is the engineering experience that made this possible, along with architectures that actually work in production.

Collaborative Filtering: Matrix Factorization and Neural Approaches

Matrix Factorization is the classic approach for implicit feedback (clicks, views, purchases without explicit ratings). ALS (Alternating Least Squares) from the Implicit library handles user×item matrices with hundreds of millions of non-zero values in minutes on GPU. Latent factors 64–256, regularization λ=0.01–0.1 are starting parameters. Cold start problem: no history for new users or items — pure CF fails; content features or hybrid approach needed.

Neural Collaborative Filtering (NCF) replaces the dot product with a neural network. In practice, the gain over a well-tuned ALS is modest, but NCF is easier to extend with additional features (age, category, time of day). Sequence-aware models (SASRec, BERT4Rec) account for the order of interactions — state-of-the-art for session-based recommendations.

How to Choose Recommender System Architecture?

The answer depends on data, load, and cold start requirements. Below are three main approaches with selection criteria.

Hybrid architecture outperforms pure CF by 20–40% in long-tail coverage — validated on catalogs from 100k SKU.

Content-Based Filtering: When Interaction History is Scarce

Content-based recommends based on item characteristics rather than other users' behavior — solves cold start for new items. Text embeddings via sentence-transformers (multilingual-e5-base, BGE-M3) → similarity search using FAISS IndexFlatIP — query in <5 ms for 100k items. Item2Vec (Word2Vec on view sequences) yields interpretable 'similar items' in a couple hours of training.

Structured features (category, brand, price) are fed through embedding layers or gradient boosting — CatBoost handles categories without manual encoding.

Why Hybrid Models Work Better?

Production systems are almost always two-level. Stage 1 (Retrieval) — fast selection of 100–500 candidates from 300k items using ALS or Two-Tower model with vector search (FAISS, Qdrant). Stage 2 (Ranking) — heavy ranker on LightGBM or neural network with cross-features, time, device, and session context. LightFM is a good starting point for medium scale without heavy infrastructure. Our practice shows: moving from single-stage to two-stage yields a 15–25% accuracy improvement with only 20–30 ms additional latency.

Real-Time Serving: Architecture Under Load

Latency SLA — 50–100 ms at thousands of requests per second. Base recommendations precomputed (batch job hourly) → Redis by user_id → <5 ms. Real-time re-ranking via Kafka for events (clicks, cart adds) → update of context features. Feature serving — Redis with TTL (views in 24 hours, last clicked item). At 10k req/s, we deploy Redis Cluster with replication.

A/B testing is the only reliable way to measure improvements. Offline metrics do not always correlate with online. Kohavi et al., 'Online Controlled Experiments at Large Scale' (KDD 2013) — a must-read for the team. Test on 5–10% of traffic, monitor CTR, conversion, revenue per session. One of our client systems after hybridization increased revenue by 18% over a month of A/B.

Recommender System Development Timeline

The stages and typical time frames are in the table below. Costs are calculated individually based on catalog scale and latency requirements.

What's Included in Turnkey Development

1. Data audit — user×item matrix density (typically <0.1%), activity distribution, temporal patterns, cold start statistics.
2. Baseline — 'popular' as a simple threshold that is often hard to beat.
3. Iterative improvement — ALS → content features → two-stage → sequence-aware. Each step with A/B.
4. Serving infrastructure — batch precomputation, Redis, real-time re-ranking, Grafana monitoring.
5. Documentation — model card with metrics, deployment instructions, feature descriptions.
6. Team training — session on interpreting results and model fine-tuning.
7. Support — 1 month post-launch (incident fixes, pipeline tuning).

We are a team with 7+ years of experience in recommender systems, having delivered over 30 projects for e-commerce and media. We guarantee transparent A/B testing and documented metric improvements.

Want to assess the growth potential of your catalog? Contact us for a free data audit. Order recommender system development — first prototype within two weeks.

from implicit.als import AlternatingLeastSquares

model = AlternatingLeastSquares(
    factors=64,
    regularization=0.05,
    iterations=15,
    use_gpu=True
)
model.fit(user_item_matrix)

More about the mathematics of recommender systems — in specialized literature.