ss

app.py

@@ -15,7 +15,7 @@ from docx import Document
 import csv
 from datasets import load_dataset
 import gc
-from transformers import AutoModelForCausalLM, AutoTokenizer
+from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
 import time
 import faiss
 import re
@@ -40,8 +40,8 @@ st.set_page_config(
 keys_to_initialize = {
     'embedding_model': None, 'embedding_model_error': None,
     'cross_encoder': None, 'cross_encoder_error': None,
-    'qwen3_1_7b_tokenizer': None, 'qwen3_1_7b_tokenizer_error': None,
-    'qwen3_1_7b_model': None, 'qwen3_1_7b_model_error': None,
+    'qwen3_4b_tokenizer': None, 'qwen3_4b_tokenizer_error': None,
+    'qwen3_4b_model': None, 'qwen3_4b_model_error': None,
     'results': [], 'resume_texts': [], 'file_names': [], 'current_job_description': ""
     # Add any other app-specific session state keys here if needed
 }
@@ -59,17 +59,18 @@ if st.session_state.embedding_model is None and st.session_state.embedding_model
         )
         print(f"[Global Init] Embedding Model (BAAI/bge-large-en-v1.5) LOADED with device_map='auto'.")
     except Exception as e:
-        if "device_map" in str(e).lower() and "unexpected keyword argument" in str(e).lower():
+        error_str = str(e) if e else ""
+        if "device_map" in error_str.lower() and "unexpected keyword argument" in error_str.lower():
             print("⚠️ [Global Init] device_map='auto' not supported for SentenceTransformer. Falling back to default device handling.")
             try:
                 st.session_state.embedding_model = SentenceTransformer('BAAI/bge-large-en-v1.5')
                 print(f"[Global Init] Embedding Model (BAAI/bge-large-en-v1.5) LOADED (fallback device handling).")
             except Exception as e_fallback:
-                error_msg = f"Failed to load Embedding Model (fallback): {str(e_fallback)}"
+                error_msg = f"Failed to load Embedding Model (fallback): {str(e_fallback) if e_fallback else 'Unknown error'}"
                 print(f"❌ [Global Init] {error_msg}")
                 st.session_state.embedding_model_error = error_msg
         else:
-            error_msg = f"Failed to load Embedding Model: {str(e)}"
+            error_msg = f"Failed to load Embedding Model: {error_str}"
             print(f"❌ [Global Init] {error_msg}")
             st.session_state.embedding_model_error = error_msg
 
@@ -83,57 +84,69 @@ if st.session_state.cross_encoder is None and st.session_state.cross_encoder_err
         )
         print(f"[Global Init] Cross-Encoder Model (ms-marco-MiniLM-L6-v2) LOADED with device_map='auto'.")
     except Exception as e:
-        if "device_map" in str(e).lower() and "unexpected keyword argument" in str(e).lower():
+        error_str = str(e) if e else ""
+        if "device_map" in error_str.lower() and "unexpected keyword argument" in error_str.lower():
             print("⚠️ [Global Init] device_map='auto' not supported for CrossEncoder. Falling back to default device handling.")
             try:
                 st.session_state.cross_encoder = CrossEncoder('cross-encoder/ms-marco-MiniLM-L6-v2')
                 print(f"[Global Init] Cross-Encoder Model (ms-marco-MiniLM-L6-v2) LOADED (fallback device handling).")
             except Exception as e_fallback:
-                error_msg = f"Failed to load Cross-Encoder Model (fallback): {str(e_fallback)}"
+                error_msg = f"Failed to load Cross-Encoder Model (fallback): {str(e_fallback) if e_fallback else 'Unknown error'}"
                 print(f"❌ [Global Init] {error_msg}")
                 st.session_state.cross_encoder_error = error_msg
         else:
-            error_msg = f"Failed to load Cross-Encoder Model: {str(e)}"
+            error_msg = f"Failed to load Cross-Encoder Model: {error_str}"
             print(f"❌ [Global Init] {error_msg}")
             st.session_state.cross_encoder_error = error_msg
 
-# Load Qwen3-1.7B Tokenizer
-if st.session_state.qwen3_1_7b_tokenizer is None and st.session_state.qwen3_1_7b_tokenizer_error is None:
-    print("[Global Init] Loading Qwen3-1.7B Tokenizer...")
+# Load Qwen3-4B Tokenizer
+if st.session_state.qwen3_4b_tokenizer is None and st.session_state.qwen3_4b_tokenizer_error is None:
+    print("[Global Init] Loading Qwen3-4B Tokenizer...")
     try:
-        st.session_state.qwen3_1_7b_tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-1.7B")
-        print("[Global Init] Qwen3-1.7B Tokenizer Loaded.")
+        st.session_state.qwen3_4b_tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-4B")
+        print("[Global Init] Qwen3-4B Tokenizer Loaded.")
     except Exception as e:
-        error_msg = f"Failed to load Qwen3-1.7B Tokenizer: {str(e)}"
+        error_msg = f"Failed to load Qwen3-4B Tokenizer: {str(e) if e else 'Unknown error'}"
         print(f"❌ [Global Init] {error_msg}")
-        st.session_state.qwen3_1_7b_tokenizer_error = error_msg
+        st.session_state.qwen3_4b_tokenizer_error = error_msg
 
-# Load Qwen3-1.7B Model
-if st.session_state.qwen3_1_7b_model is None and st.session_state.qwen3_1_7b_model_error is None:
-    print("[Global Init] Loading Qwen3-1.7B Model (attempting with device_map='auto')...")
+# Load Qwen3-4B Model with 4-bit quantization
+if st.session_state.qwen3_4b_model is None and st.session_state.qwen3_4b_model_error is None:
+    print("[Global Init] Loading Qwen3-4B Model with 4-bit quantization...")
     try:
-        st.session_state.qwen3_1_7b_model = AutoModelForCausalLM.from_pretrained(
-            "Qwen/Qwen3-1.7B", 
-            torch_dtype="auto", 
+        # Configure 4-bit quantization for better performance and memory efficiency
+        quantization_config = BitsAndBytesConfig(
+            load_in_4bit=True,
+            bnb_4bit_quant_type="nf4",
+            bnb_4bit_compute_dtype=torch.float16,
+            bnb_4bit_use_double_quant=True
+        )
+        
+        st.session_state.qwen3_4b_model = AutoModelForCausalLM.from_pretrained(
+            "Qwen/Qwen3-4B", 
+            quantization_config=quantization_config,
             device_map="auto",
-            trust_remote_code=True # if required by this specific model
+            trust_remote_code=True,
+            torch_dtype=torch.float16,
+            use_cache=True
         )
-        print("[Global Init] Qwen3-1.7B Model Loaded with device_map='auto'.")
-    except Exception as e_dev_map:
-        print(f"⚠️ [Global Init] Failed to load Qwen3-1.7B with device_map='auto': {str(e_dev_map)}")
-        print("[Global Init] Retrying Qwen3-1.7B load without device_map (will use default single device)...")
+        print("[Global Init] Qwen3-4B Model Loaded with 4-bit quantization and device_map='auto'.")
+    except Exception as e_quant:
+        error_str = str(e_quant) if e_quant else ""
+        print(f"⚠️ [Global Init] Failed to load Qwen3-4B with 4-bit quantization: {error_str}")
+        print("[Global Init] Retrying Qwen3-4B load without quantization...")
         try:
-            st.session_state.qwen3_1_7b_model = AutoModelForCausalLM.from_pretrained(
-                "Qwen/Qwen3-1.7B", 
+            st.session_state.qwen3_4b_model = AutoModelForCausalLM.from_pretrained(
+                "Qwen/Qwen3-4B", 
                 torch_dtype="auto", 
-                # No device_map here, let Hugging Face decide or use CUDA if available
-                trust_remote_code=True # if required
+                device_map="auto",
+                trust_remote_code=True
             )
-            print("[Global Init] Qwen3-1.7B Model Loaded (fallback device handling).")
+            print("[Global Init] Qwen3-4B Model Loaded without quantization.")
         except Exception as e_fallback:
-            error_msg = f"Failed to load Qwen3-1.7B Model (fallback): {str(e_fallback)}"
+            error_msg = f"Failed to load Qwen3-4B Model (fallback): {str(e_fallback) if e_fallback else 'Unknown error'}"
             print(f"❌ [Global Init] {error_msg}")
-            st.session_state.qwen3_1_7b_model_error = error_msg
+            st.session_state.qwen3_4b_model_error = error_msg
 
 # --- End of Global Model Loading Section ---
 
@@ -248,7 +261,7 @@ class ResumeScreener: # Ensure this class definition is BEFORE it's instantiated
         top_20_results = self.cross_encoder_rerank(resume_texts, job_description, top_50_indices, top_k=20)
         st.info("🔤 Stage 3: BM25 Keyword Matching...")
         top_20_with_bm25 = self.add_bm25_scores(resume_texts, job_description, top_20_results)
-        st.info("🤖 Stage 4: LLM Intent Analysis (Qwen3-1.7B)...")
+        st.info("🤖 Stage 4: LLM Intent Analysis (Qwen3-4B)...")
         top_20_with_intent = self.add_intent_scores(resume_texts, job_description, top_20_with_bm25)
         st.info("🏆 Stage 5: Final Combined Ranking...")
         final_results = self.calculate_final_scores(top_20_with_intent)
@@ -331,7 +344,7 @@ class ResumeScreener: # Ensure this class definition is BEFORE it's instantiated
         return results_with_bm25
 
     def add_intent_scores(self, resume_texts, job_description, top_20_with_bm25):
-        st.text(f"LLM Intent: Analyzing intent for {len(top_20_with_bm25)} candidates (Qwen3-1.7B)...")
+        st.text(f"LLM Intent: Analyzing intent for {len(top_20_with_bm25)} candidates (Qwen3-4B)...")
         results_with_intent = []
         progress_bar = st.progress(0)
         for i, (idx, cross_score, bm25_score) in enumerate(top_20_with_bm25):
@@ -342,20 +355,48 @@ class ResumeScreener: # Ensure this class definition is BEFORE it's instantiated
         return results_with_intent
 
     def analyze_intent(self, resume_text, job_description):
-        print(f"[analyze_intent] Analyzing intent for one resume (Qwen3-1.7B)...")
-        st.text("LLM Intent: Analyzing intent (Qwen3-1.7B)...")
+        print(f"[analyze_intent] Analyzing intent for one resume (Qwen3-4B)...")
+        st.text("LLM Intent: Analyzing intent (Qwen3-4B)...")
         try:
             resume_snippet = resume_text[:15000]
             job_snippet = job_description[:5000]
-            prompt = f\"\"\"You are given a job description and a candidate's resume... (rest of prompt)\"\"\" # Ensure f-string is correct
-            # ... (rest of analyze_intent, using st.session_state.qwen3_1_7b_tokenizer and _model)
+            
+            prompt = f"""You are given a job description and a candidate's resume. Your task is to analyze whether the candidate is likely seeking this specific type of job.
+
+Job Description:
+{job_snippet}
+
+Candidate Resume:
+{resume_snippet}
+
+Please analyze the candidate's background, skills, experience, and career trajectory to determine if they would be genuinely interested in and likely to apply for this position.
+
+Consider:
+1. Does their experience align with the job requirements?
+2. Is this a logical career progression for them?
+3. Do their skills match what's needed?
+4. Would this role be appealing given their background?
+
+Think through your analysis step by step, then provide your final assessment.
+
+Respond with exactly one of these formats:
+- Intent: Yes (if they would likely seek this job)
+- Intent: Maybe (if it's uncertain or partially aligned)
+- Intent: No (if they would likely not seek this job)"""
+
+            # Check if models are available
+            if not st.session_state.get('qwen3_4b_tokenizer') or not st.session_state.get('qwen3_4b_model'):
+                st.warning("Qwen3-4B model not available, using fallback intent score.")
+                return 0.1
+                
             response_text = generate_qwen3_response(
                 prompt,
-                st.session_state.qwen3_1_7b_tokenizer,
-                st.session_state.qwen3_1_7b_model,
+                st.session_state.qwen3_4b_tokenizer,
+                st.session_state.qwen3_4b_model,
                 max_new_tokens=20000
             )
-            # ... (parsing logic for response_text) ...
+            
+            # Parse thinking content and intent decision
             thinking_content = "No detailed thought process extracted."
             intent_decision_part = response_text
             think_start_tag = "<think>"
@@ -365,15 +406,20 @@ class ResumeScreener: # Ensure this class definition is BEFORE it's instantiated
             if start_index != -1 and end_index != -1 and start_index < end_index:
                 thinking_content = response_text[start_index + len(think_start_tag):end_index].strip()
                 intent_decision_part = response_text[end_index + len(think_end_tag):].strip()
+            
             response_lower = intent_decision_part.lower()
-            intent_score = 0.1
+            intent_score = 0.1  # Default "Maybe" score
+            
             if 'intent: yes' in response_lower or 'intent:yes' in response_lower:
                 intent_score = 0.3
             elif 'intent: no' in response_lower or 'intent:no' in response_lower:
                 intent_score = 0.0
+            elif 'intent: maybe' in response_lower or 'intent:maybe' in response_lower:
+                intent_score = 0.1
+                
             return intent_score
         except Exception as e:
-            st.warning(f"Error analyzing intent with Qwen3-1.7B: {str(e)}")
+            st.warning(f"Error analyzing intent with Qwen3-4B: {str(e)}")
             return 0.1
 
     def calculate_final_scores(self, results_with_all_scores):
@@ -432,12 +478,12 @@ with st.sidebar:
     st.markdown("- **Stage 1**: FAISS Recall (Top 50)")
     st.markdown("- **Stage 2**: Cross-Encoder Re-ranking (Top 20)")
     st.markdown("- **Stage 3**: BM25 Keyword Matching")
-    st.markdown("- **Stage 4**: LLM Intent Analysis (Qwen3-1.7B)")
+    st.markdown("- **Stage 4**: LLM Intent Analysis (Qwen3-4B)")
     st.markdown("- **Final**: Combined Scoring") # Updated this line
     st.markdown("### 📊 Models Used")
     st.markdown("- **Embedding**: BAAI/bge-large-en-v1.5")
     st.markdown("- **Cross-Encoder**: ms-marco-MiniLM-L6-v2")
-    st.markdown("- **LLM**: Qwen/Qwen3-1.7B")
+    st.markdown("- **LLM**: Qwen/Qwen3-4B (4-bit quantized)")
     st.markdown("### 📈 Scoring Formula")
     st.markdown("**Final Score = Cross-Encoder (0-1) + BM25 (0.1-0.2) + Intent (0-0.3)**")
 
@@ -445,7 +491,7 @@ with st.sidebar:
 st.title("🎯 AI-Powered Resume Screener")
 # ... (Model Loading Status display as before)
 # ...
-st.markdown("*Find the perfect candidates using BAAI/bge-large-en-v1.5 embeddings and Qwen3-1.7B for intent analysis*")
+st.markdown("*Find the perfect candidates using BAAI/bge-large-en-v1.5 embeddings and Qwen3-4B for intent analysis*")
 
 st.subheader("🤖 Model Loading Status")
 col1, col2 = st.columns(2)
@@ -463,18 +509,18 @@ with col1:
     else:
         st.warning("⏳ Cross-Encoder Model loading or not found (check console).")
 with col2:
-    if st.session_state.get('qwen3_1_7b_tokenizer_error'):
-        st.error(f"Qwen3-1.7B Tokenizer: {st.session_state.qwen3_1_7b_tokenizer_error}")
-    elif st.session_state.get('qwen3_1_7b_tokenizer'):
-        st.success("✅ Qwen3-1.7B Tokenizer loaded.")
+    if st.session_state.get('qwen3_4b_tokenizer_error'):
+        st.error(f"Qwen3-4B Tokenizer: {st.session_state.qwen3_4b_tokenizer_error}")
+    elif st.session_state.get('qwen3_4b_tokenizer'):
+        st.success("✅ Qwen3-4B Tokenizer loaded.")
     else:
-        st.warning("⏳ Qwen3-1.7B Tokenizer loading or not found (check console).")
-    if st.session_state.get('qwen3_1_7b_model_error'):
-        st.error(f"Qwen3-1.7B Model: {st.session_state.qwen3_1_7b_model_error}")
-    elif st.session_state.get('qwen3_1_7b_model'):
-        st.success("✅ Qwen3-1.7B Model loaded.")
+        st.warning("⏳ Qwen3-4B Tokenizer loading or not found (check console).")
+    if st.session_state.get('qwen3_4b_model_error'):
+        st.error(f"Qwen3-4B Model: {st.session_state.qwen3_4b_model_error}")
+    elif st.session_state.get('qwen3_4b_model'):
+        st.success("✅ Qwen3-4B Model loaded (4-bit quantized).")
     else:
-        st.warning("⏳ Qwen3-1.7B Model loading or not found (check console).")
+        st.warning("⏳ Qwen3-4B Model loading or not found (check console).")
 st.markdown("---")
 
 # Initialize screener (This line was causing NameError, ensure class is defined above)
@@ -671,8 +717,8 @@ with col1:
                  disabled=not (job_description and st.session_state.resume_texts and 
                               st.session_state.get('embedding_model') and
                               st.session_state.get('cross_encoder') and
-                              st.session_state.get('qwen3_1_7b_model') and
-                              st.session_state.get('qwen3_1_7b_tokenizer')),
+                              st.session_state.get('qwen3_4b_model') and
+                              st.session_state.get('qwen3_4b_tokenizer')),
                  type="primary",
                  help="Run the complete 5-stage advanced pipeline"):
         print("--- Advanced Pipeline Analysis Button Clicked ---")
@@ -901,7 +947,7 @@ st.markdown("---")
 st.markdown(
     """
     <div style='text-align: center; color: #666;'>
-        🚀 Powered by BAAI/bge-large-en-v1.5 & Qwen3-1.7B | Built with Streamlit
+        🚀 Powered by BAAI/bge-large-en-v1.5 & Qwen3-4B (4-bit) | Built with Streamlit
     </div>
     """, 
     unsafe_allow_html=True