o Tj@sDUdZddlmZddlZddlmZiZded<Gdd d Z dS) a&Small sentence embedder shared by the RAG index, the retriever, and the semantic cache. We use one tiny model (default: BAAI/bge-small-en-v1.5) for everything so the memory/energy footprint stays low. bge/e5 models expect a short query prefix for retrieval; documents are embedded without it. ) annotationsN)Configzdict[str, object] _MODEL_CACHEc@s2eZdZdddZddZdd d ZdddZdS)EmbeddercfgrcCs4|dd|_|dd|_|dd|_d|_dS)Nzembeddings.model_idzBAAI/bge-small-en-v1.5zembeddings.query_prefixzembeddings.deviceZcpu)getmodel_id query_prefixdevice_model)selfrrsrc/agent/embeddings.py__init__s zEmbedder.__init__cCsZ|jdur|jS|jd|j}|tvr%ddlm}||j|jdt|<t||_|jS)N:r)SentenceTransformer)r )r r r rZsentence_transformersr)rkeyrrrr_loads   zEmbedder._loadtexts list[str]return np.ndarraycCs$|}|j|dddd}|dSNTF)Znormalize_embeddingsZconvert_to_numpyZshow_progress_barZfloat32)rencodeastype)rrmodelZvecsrrrembed_documents#s  zEmbedder.embed_documentstextstrcCs*|}|j|j|dddd}|dSr)rrr r)rrrZvecrrr embed_query*s zEmbedder.embed_queryN)rr)rrrr)rr rr)__name__ __module__ __qualname__rrrr!rrrrrs   r) __doc__Z __future__rZnumpyZnpZsettingsrr__annotations__rrrrrs