Update app.py

app.py

@@ -17,7 +17,6 @@ sentence_model = SentenceTransformer('sentence-transformers/all-MiniLM-L6-v2')
 # Initialize Groq client
 client = Groq(api_key=os.environ.get("GROQ_API_KEY"))
 
-
 # System prompt for Groq
 system_prompt = {
     "role": "system",
@@ -56,7 +55,6 @@ def extract_text_from_image(filepath: str, languages: List[str]):
     lang_str = '+'.join(languages)  # Join languages for Tesseract
     return pytesseract.image_to_string(image=image, lang=lang_str)
 
-
 # Assign badges based on the grade
 def assign_badge(grade):
     if grade == 5:
@@ -68,7 +66,6 @@ def assign_badge(grade):
     else:
         return "Keep Improving Badge 💪"
 
-
 # Categorize feedback into clarity, completeness, and accuracy
 def detailed_feedback(similarity_score):
     if similarity_score >= 0.9:
@@ -93,8 +90,6 @@ def get_grade(similarity_score):
     else:
         return 1
 
-
-
 # Function to get BERT embeddings
 def get_bert_embedding(text):
     inputs = tokenizer(text, return_tensors='pt', truncation=True, padding=True)
@@ -116,13 +111,16 @@ def calculate_sentence_similarity(text1, text2):
 
 # Function to compare logic of student and teacher answers
 def compare_answers(student_answer, teacher_answer):
+    # Use primarily BERT similarity, and sentence embeddings as a secondary method
     bert_similarity = calculate_cosine_similarity(get_bert_embedding(student_answer), get_bert_embedding(teacher_answer))
+    
+    # Adjust weight: give more emphasis to BERT
     sentence_similarity = calculate_sentence_similarity(student_answer, teacher_answer)
-    # Combine scores with weights to emphasize one method over another if needed
-    final_similarity = (0.5 * bert_similarity + 0.5 * sentence_similarity)
+    
+    final_similarity = (0.7 * bert_similarity + 0.3 * sentence_similarity)
     return final_similarity
 
-# Function to extract keywords from the model answer (simple keyword extraction)
+# Function to extract keywords from the model answer
 def extract_keywords(text):
     # Split text into words and count them
     return set(text.lower().split())
@@ -132,12 +130,16 @@ def check_keywords(student_answer, model_answer):
     student_keywords = extract_keywords(student_answer)
     teacher_keywords = extract_keywords(model_answer)
     keyword_overlap = len(student_keywords.intersection(teacher_keywords))
-    return keyword_overlap / len(teacher_keywords)  # Ratio of matching keywords
+    return keyword_overlap / (len(teacher_keywords) if len(teacher_keywords) > 0 else 1)  # Ratio of matching keywords
 
 # Function to evaluate student's answer by comparing it to a model answer
 def evaluate_answer(image, languages, model_answer):
     student_answer = extract_text_from_image(image, languages)
+    
+    # Calculate semantic similarity primarily with BERT
     semantic_similarity = compare_answers(student_answer, model_answer)
+    
+    # Include keyword similarity
     keyword_similarity = check_keywords(student_answer, model_answer)
     
     # Combine semantic similarity with keyword presence check
@@ -149,9 +151,6 @@ def evaluate_answer(image, languages, model_answer):
     prompt = f"The student got grade: {grade} when the student's answer is: {student_answer} and the teacher's answer is: {model_answer}. Justify the grade given to the student."
     return grade, combined_similarity * 100, feedback, badge, detailed_feedback_msg, prompt
 
-# The rest of your existing code...
-
-
 # Main interface function for Gradio
 async def gradio_interface(image, languages: List[str], model_answer="The process of photosynthesis helps plants produce glucose using sunlight.", prompt="", history=[]):
     grade, similarity_score, feedback, badge, detailed_feedback_msg, prompt = evaluate_answer(image, languages, model_answer)