Expand description
Product policy for the shared embedding engine: the retrieval model, identifier-aware inputs, lockdocs’ opt-out and existing cache location.
Structs§
- Query
Model - A query-side embedder that never loads the weight table: it reads the
vocabulary, the row scales and only the rows of the query’s own tokens
(about 512 bytes each) with positioned reads. It reproduces
Model::embedunderTokenization::Identifiersexactly (a test checks this); build the stored vectors with the fullModel. - Vec8
- An embedding stored as int8 values and one scale (value = q × s).
Constants§
Functions§
- available
- Embeddings are on and the model is installed (nothing is loaded).
- cosine
- Cosine similarity of a unit query with a stored unit vector.
- enabled
- Embeddings are on unless
LOCKDOCS_EMBED=0(or false/off). - ensure
- get
- Load existing model.q8/vocab.txt directly. The shared input policy keeps cached entry vectors and new query vectors in the same embedding space.
- installed
- quantize
- query_
model - A cheap embedder for query text, without loading the weight table.