Create embed.py

embed.py

@@ -0,0 +1,50 @@
+# embed.py
+# Purpose: turn text-based columns (e.g., city, notes) or entire customer row into dense embeddings using
+# a sentence-transformers model from Hugging Face.
+
+
+import pandas as pd
+import numpy as np
+from sentence_transformers import SentenceTransformer
+import joblib
+
+
+MODEL_NAME = 'all-MiniLM-L6-v2' # lightweight sentence-transformer
+
+
+
+
+def build_text_for_embedding(df: pd.DataFrame, text_cols=None):
+# combine useful text columns into one string per customer
+if text_cols is None:
+# try to pick common textual columns
+cand = [c for c in df.columns if df[c].dtype == 'object']
+text_cols = cand[:3] # at most 3
+texts = (df[text_cols].fillna('').astype(str).agg(' | '.join, axis=1))
+return texts.tolist()
+
+
+
+
+def embed_texts(texts, model_name=MODEL_NAME, device='cpu'):
+model = SentenceTransformer(model_name)
+embs = model.encode(texts, show_progress_bar=True)
+return embs
+
+
+
+
+if __name__ == '__main__':
+import argparse
+parser = argparse.ArgumentParser()
+parser.add_argument('--features', default='data/features.parquet')
+parser.add_argument('--out_emb', default='data/embeddings.npy')
+parser.add_argument('--text_cols', nargs='*', default=None)
+args = parser.parse_args()
+
+
+df = pd.read_parquet(args.features)
+texts = build_text_for_embedding(df, text_cols=args.text_cols)
+embs = embed_texts(texts)
+np.save(args.out_emb, embs)
+print('Saved embeddings to', args.out_emb)