untracked files on main: dceb101 feat(#63): IMP-34 R1 donor capacity measured bound (u1+u2)

tests/matching/embeddings.py

@@ -0,0 +1,37 @@
+"""ko-sroberta embeddings for template-fit content similarity.
+
+Model: jhgan/ko-sroberta-multitask (HuggingFace)
+용도: MDX summary ↔ frame description 코사인 유사도 계산
+32 frame 규모 — 벡터 DB 없이 numpy 만으로 충분
+"""
+import numpy as np
+
+_model = None
+
+
+def _get_model():
+    global _model
+    if _model is None:
+        from sentence_transformers import SentenceTransformer
+        _model = SentenceTransformer('jhgan/ko-sroberta-multitask')
+    return _model
+
+
+def embed_texts(texts):
+    """Batch embed. Returns numpy (N, 768)."""
+    model = _get_model()
+    return model.encode(list(texts), show_progress_bar=False, convert_to_numpy=True)
+
+
+def cosine(a, b):
+    """Single-vector cosine similarity."""
+    an = a / (np.linalg.norm(a) + 1e-12)
+    bn = b / (np.linalg.norm(b) + 1e-12)
+    return float(np.dot(an, bn))
+
+
+def embed_map(id_to_text):
+    """{id: text} → {id: vec(768,)}"""
+    ids = list(id_to_text.keys())
+    vecs = embed_texts([id_to_text[i] for i in ids])
+    return {ids[i]: vecs[i] for i in range(len(ids))}