Recommendation & search

Representation learning

item embeddings, image–text retrieval, similarity search

Learn a vector space where distance means similarity, to power retrieval, matching, and downstream models.

Linear algebra

Information theory

Optimisation

Foundations

Inner products, projections, and the geometry of high-dimensional spaces; InfoNCE is a mutual-information bound.

Frame the problem

Frame

Define what similar means and who consumes the space: retrieval, clustering, or a downstream model. That decides the loss and the evaluation.

Sourcing & signal

Labeling & data collection

Data & labels

Positive pairs come from co-occurrence, clicks, captions, or augmentation. Negatives are usually sampled, and how you sample them is the design.

Modalities → vectors

Tokenization

Embedding layers

word2vec & GloVebackground

Represent

One encoder per modality. Ids as embedding tables; text and images through pretrained encoders.

Split the data

Split

By entity. Hold out pairs whose entities appear on neither side of training if you need generalisation to new items.

Dumb baseline

Dimensionality reduction

Matrix factorization

Attention & transformersbackground

Contrastive training

Siamese or two-tower with in-batch negatives; hard negatives once the easy ones are solved.

Metric learning

Two-tower retrieval & ANN

Pretrained and multimodal

CLIP-style image–text spaces; sentence encoders from masked or contrastive pretraining.

Vision transformers & CLIP

Masked & contrastive pretraining

Model

PCA or SVD over co-occurrence, or matrix factorization, is the baseline embedding.

Loss functions

Training craft

Fine-tuning a pretrained model

Train

Triplet, InfoNCE, or sampled softmax. Temperature and batch size matter more than depth.

Embedding geometry & retrieval eval

Ranking & recsys metrics

Evaluate offline

Recall@k on held-out pairs. Check anisotropy and hubness; a space can score well and still collapse.

A/B testing & interleaving

Evaluate online

Measured through whatever consumes it: a retrieval A/B, or downstream model lift.

Serving & release

Inference optimisation

Ship & monitor

Version the space: a retrained encoder invalidates every stored vector. Plan the re-index.

React Flow