# semantic_ranker.py
from typing import Optional
import numpy as np
import pandas as pd
from sentence_transformers import SentenceTransformer, util

_model = None

def _lazy_model():
    global _model
    if _model is None:
        _model = SentenceTransformer("sentence-transformers/all-MiniLM-L6-v2")  # CPU-fast
    return _model

def score_courses(df: pd.DataFrame, query: str, text_cols=("name","subject")) -> pd.DataFrame:
    """Add 'sem_score' column based on cosine similarity to query; higher is better."""
    if not query or not query.strip():
        df["sem_score"] = 0.0
        return df
    model = _lazy_model()
    corpus = (df[list(text_cols)]
              .fillna("")
              .agg(" - ".join, axis=1)
              .tolist())
    q_emb = model.encode([query], convert_to_tensor=True, normalize_embeddings=True)
    c_emb = model.encode(corpus, convert_to_tensor=True, normalize_embeddings=True)
    sims = util.cos_sim(q_emb, c_emb).cpu().numpy().ravel()
    df = df.copy()
    df["sem_score"] = sims
    # sort with semantic score first, then keep original order for stability
    df = df.sort_values(by="sem_score", ascending=False)
    return df