Update skill_extraction.py

skill_extraction.py

@@ -14,11 +14,11 @@ except OSError:
 # Embedding model
 model = SentenceTransformer("sentence-transformers/all-MiniLM-L6-v2")
 
-# Load skills from JSON file
+# Load skills from JSON
 with open("skills_vocab.json", "r", encoding="utf-8") as f:
     SKILL_VOCAB = json.load(f)["skills"]
 
-# Precompute embeddings once
+# Precompute skill embeddings
 skill_embeddings = model.encode(SKILL_VOCAB, convert_to_numpy=True, normalize_embeddings=True)
 
 def extract_skills(text, threshold=0.50):
@@ -30,23 +30,16 @@ def extract_skills(text, threshold=0.50):
 
     sentence_embeddings = model.encode(sentences, convert_to_numpy=True, normalize_embeddings=True)
 
-    skill_confidences = {}
+    extracted = []
 
+    # Iterate skill-by-skill
     for j, skill_emb in enumerate(skill_embeddings):
-        sims = np.dot(sentence_embeddings, skill_emb)  # similarity to sentences
+        sims = np.dot(sentence_embeddings, skill_emb)
         max_sim = float(np.max(sims))
-
         if max_sim >= threshold:
-            skill_confidences[SKILL_VOCAB[j]] = round(max_sim, 2)
-
-    return sorted(skill_confidences.items(), key=lambda x: x[1], reverse=True)
+            best_sentence = sentences[int(np.argmax(sims))]
+            extracted.append((SKILL_VOCAB[j], round(max_sim, 2), best_sentence))
 
+    # Sort highest confidence first
+    return sorted(extracted, key=lambda x: x[1], reverse=True)
 
-def get_skill_embedding(skill_name):
-    """
-    Return embedding for a skill stored in SKILL_VOCAB
-    """
-    if skill_name not in SKILL_VOCAB:
-        return None
-    idx = SKILL_VOCAB.index(skill_name)
-    return skill_embeddings[idx]