from __future__ import annotations import logging from legora.performance import QueryCache, normalize_query from sentence_transformers import ( SentenceTransformer ) logger = logging.getLogger(__name__) class LegalEmbedder: def __init__( self, model_name: str = "BAAI/bge-large-en-v1.5" ): logger.info("Loading embedding model: %s", model_name) self.model_name = model_name self.query_cache = QueryCache() self.model = ( SentenceTransformer( model_name ) ) # ===================================================== # DOCUMENT EMBEDDINGS # ===================================================== def embed( self, texts: list[str] ): return self.model.encode( texts, normalize_embeddings=True, convert_to_numpy=True, batch_size=16, show_progress_bar=True ) # ===================================================== # QUERY EMBEDDING # ===================================================== def embed_query( self, query: str ): normalized = normalize_query(query) if self.query_cache.enabled else query return self.query_cache.get((self.model_name, normalized), lambda: self._embed_query(normalized)) def _embed_query(self, query: str): """Encode one query with the unchanged BGE retrieval instruction.""" query = ( "Represent this sentence " "for searching relevant " f"passages: {query}" ) return self.model.encode( query, normalize_embeddings=True, convert_to_numpy=True ) # ===================================================== # VECTOR SIZE # ===================================================== def vector_size( self ) -> int: return ( self.model .get_sentence_embedding_dimension() ) # ========================================================= # TEST # ========================================================= if __name__ == "__main__": embedder = ( LegalEmbedder() ) texts = [ "Section 52. Facts of which Court shall take judicial notice.", "Article 21. Protection of life and personal liberty." ] vectors = ( embedder.embed( texts ) ) logger.debug('%s %s', "Vectors:", len(vectors)) logger.debug('%s %s', "Dimension:", len(vectors[0])) query_vector = ( embedder.embed_query( "facts judicially noticed by court" ) ) logger.debug('%s %s', "Query Dimension:", len(query_vector)) logger.debug('%s %s', "Model Dimension:", embedder.vector_size())