aeat.application.corpus_search._query_embed module

Runtime query embedder (the live-query half of the R3 semantic stack).

The corpus vectors are precomputed at build time and ship as data (_embed_build); the model is needed at runtime ONLY to embed the operator’s live query into the same vector space so a cosine search can run. This module owns that one live use of model2vec, behind the capability- gated aeat-cli[search] extra: absent the extra, search_extra_available() reports False and the retrieval layer degrades to lexical-only, while constructing/using a QueryEmbedder refuses with the install hint.

The model download is app-controlled: the cache directory is rooted under the Settings aeat_local_storage_root (<root>/search-models) rather than the user’s global Hugging Face cache, so a bundled Desktop Extension keeps its model state inside the one app state root. The model is loaded lazily on first embed and cached for the embedder’s lifetime, so repeated queries pay the load once.

search_extra_available()[source]

Return whether the semantic search extra (model2vec) is importable.

The retrieval layer calls this to decide between hybrid and lexical-only degraded mode WITHOUT triggering a model load or download.

Return type:

bool

search_model_cache_dir(settings=None)[source]

Return the app-controlled model cache directory under the storage root.

Return type:

Path

Parameters:

settings (Settings | None)

class QueryEmbedder(*, model_id='minishlab/potion-multilingual-128M', revision='73908c3438cf03b6a01bcb9611d62b23d0726f08', cache_dir=None, settings=None)[source]

Bases: object

Embed live queries with the pinned potion static model.

Construction records the model id, pinned revision, and app-controlled cache directory but does NOT load the model; the first embed_query() loads it (refusing with the install hint when the extra is absent) and caches it for reuse.

Parameters:

  • model_id (str)

  • revision (str)

  • cache_dir (Path | None)

  • settings (Settings | None)

property model_id: str
property revision: str
property cache_dir: Path
embed_query(text)[source]

Embed one query string into a 1-D float32 vector.

Parameters:

text (str) – The free-text query.

Return type:

ndarray

Returns:

A 1-D float32 numpy vector in the corpus embedding space.

Raises: