# Embedding architecture

Semantic retrieval is an optional lane beside lexical, exact-signal, tag, and
trust-aware retrieval. The local provider reads a verified model directory and
never sends memory or query text to Hugging Face. SQLite stores settings,
profile identity, installation metadata, jobs, and rebuildable vector rows.

Setup downloads allowlisted files into private staging, verifies size and
SHA-256, runs an offline probe, and atomically installs the revision. Database
activation and model I/O are separate phases. A profile identity includes the
model revision, artifact manifest, runtime version, dtype, pooling,
normalization, token limit, dimensions, and E5 input contract.