Concepts

How candding turns a text into a vector, from templates and pooling to devices, dtypes and the files a model needs.

A model's vectors match its reference only when the template, the tokenizer's padding and truncation, the pooling and the normalization all match. A ModelDescriptor records those choices, from the registry, from you or from a repository's own sentence-transformers files, and every family runs the same pipeline around its encoder on every device.

Page	Contents
Pipeline	every stage from text to vector, with the tensor shapes between them
Pooling and templates	the pooling modes, query and passage templates, instructions and Matryoshka truncation
Devices and dtypes	device selection, default dtypes, the one attention branch and the feature flags
Model files and cache	the cache layout, the files candding downloads and reads, gated repositories and local directories
Sparse output	the six sparse heads, what each indexes and how two rows are scored
Multi-vector output	the four checkpoints' head conventions and late-interaction scoring
Rerank output	one score per query and document, what each checkpoint's score means, how long inputs are cut and each checkpoint's status