Embedding

The mechanism underneath semantic search, retrieval, clustering, deduplication, classification and recommendation: convert text (or images, or audio) to a vector, and treat distance as similarity.

The operational facts that shape architecture:

The embedding model is a hard dependency of the index. Changing it invalidates every stored vector, because the new vectors live in a different space. Re-embedding a large corpus is a real project, so plan for it — versioned indexes and a dual-read migration path — before you need it.

Query and document must use the same model, and asymmetric models require using the correct side for each.

Chunking determines quality more than the model does. A chunk that splits a table from its header embeds to something that matches nothing useful.

Dense retrieval misses exact identifiers. Product codes, error numbers and policy references are found immediately by keyword search and poorly by vectors, which is why hybrid retrieval is consistently the largest single quality improvement available.

Dimensionality is a cost lever: larger vectors are more expressive and cost more to store and search, and many models now support truncation to a shorter prefix with modest quality loss.