app.py

import streamlit as st
import pandas as pd
import numpy as np
import torch
import nltk
import os
import tempfile
import base64
from rank_bm25 import BM25Okapi
from sentence_transformers import SentenceTransformer, CrossEncoder
from nltk.tokenize import word_tokenize
import pdfplumber
import PyPDF2
from docx import Document
import csv
from datasets import load_dataset
import gc
from transformers import AutoModelForCausalLM, AutoTokenizer
import time
import faiss
import re

# Download NLTK resources
try:
    nltk.data.find('tokenizers/punkt')
except LookupError:
    nltk.download('punkt')

# Set page configuration
st.set_page_config(
    page_title="AI Resume Screener",
    page_icon="🎯",
    layout="wide",
    initial_sidebar_state="expanded"
)

# Sidebar configuration
with st.sidebar:
    st.title("⚙️ Configuration")
    
    # Ranking weights
    st.subheader("Ranking Weights")
    semantic_weight = st.slider("Semantic Similarity Weight", 0.0, 1.0, 0.7, 0.1)
    keyword_weight = 1.0 - semantic_weight
    st.write(f"Keyword Weight: {keyword_weight:.1f}")
    
    # Advanced options
    st.subheader("Advanced Options")
    top_k = st.selectbox("Number of results to display", options=[1, 2, 3, 4, 5], index=4)
    
    # LLM Settings
    st.subheader("LLM Settings")
    use_llm_explanations = st.checkbox("Generate AI Explanations", value=True)
    if use_llm_explanations:
        hf_token = st.text_input("Hugging Face Token (optional)", type="password", 
                                help="Enter your HF token for better rate limits")
    
    st.markdown("---")
    st.markdown("### 🤖 Advanced Pipeline")
    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")
    st.markdown("- **Final**: Combined Scoring (Top 5)")
    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 Explanations**: Qwen/Qwen3-14B")
    st.markdown("- **Intent Analysis**: Qwen/Qwen3-1.7B")
    st.markdown("### 📈 Scoring Formula")
    st.markdown("**Final Score = Cross-Encoder (0-1) + BM25 (0.1-0.2) + Intent (0-0.3)**")

# Initialize session state
if 'embedding_model' not in st.session_state:
    st.session_state.embedding_model = None
if 'cross_encoder' not in st.session_state:
    st.session_state.cross_encoder = None
if 'results' not in st.session_state:
    st.session_state.results = []
if 'resume_texts' not in st.session_state:
    st.session_state.resume_texts = []
if 'file_names' not in st.session_state:
    st.session_state.file_names = []
if 'explanations_generated' not in st.session_state:
    st.session_state.explanations_generated = False
if 'current_job_description' not in st.session_state:
    st.session_state.current_job_description = ""
if 'qwen3_tokenizer' not in st.session_state:
    st.session_state.qwen3_tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-14B")
if 'qwen3_model' not in st.session_state:
    st.session_state.qwen3_model = AutoModelForCausalLM.from_pretrained(
        "Qwen/Qwen3-14B",
        torch_dtype="auto",
        device_map="auto"
    )
# Separate smaller model for intent analysis
if 'qwen3_intent_tokenizer' not in st.session_state:
    st.session_state.qwen3_intent_tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-1.7B")
if 'qwen3_intent_model' not in st.session_state:
    st.session_state.qwen3_intent_model = AutoModelForCausalLM.from_pretrained(
        "Qwen/Qwen3-1.7B",
        torch_dtype="auto",
        device_map="auto"
    )

@st.cache_resource
def load_embedding_model():
    """Load and cache the BGE embedding model"""
    try:
        with st.spinner("🔄 Loading BAAI/bge-large-en-v1.5 model..."):
            model = SentenceTransformer('BAAI/bge-large-en-v1.5')
            st.success("✅ Embedding model loaded successfully!")
            return model
    except Exception as e:
        st.error(f"❌ Error loading embedding model: {str(e)}")
        return None

@st.cache_resource
def load_cross_encoder():
    """Load and cache the Cross-Encoder model"""
    try:
        with st.spinner("🔄 Loading Cross-Encoder ms-marco-MiniLM-L6-v2..."):
            from sentence_transformers import CrossEncoder
            model = CrossEncoder('cross-encoder/ms-marco-MiniLM-L6-v2')
            st.success("✅ Cross-Encoder model loaded successfully!")
            return model
    except Exception as e:
        st.error(f"❌ Error loading Cross-Encoder model: {str(e)}")
        return None

def generate_qwen3_response(prompt, tokenizer, model, max_new_tokens=200):
    messages = [{"role": "user", "content": prompt}]
    text = tokenizer.apply_chat_template(
        messages,
        tokenize=False,
        add_generation_prompt=True,
        enable_thinking=True
    )
    model_inputs = tokenizer([text], return_tensors="pt").to(model.device)
    generated_ids = model.generate(
        **model_inputs,
        max_new_tokens=max_new_tokens
    )
    output_ids = generated_ids[0][len(model_inputs.input_ids[0]):].tolist()
    response = tokenizer.decode(output_ids, skip_special_tokens=True).strip("\n")
    return response

class ResumeScreener:
    def __init__(self):
        # Load models
        self.embedding_model = load_embedding_model()
        self.cross_encoder = load_cross_encoder()
    
    def extract_text_from_file(self, file_path, file_type):
        """Extract text from various file types"""
        try:
            if file_type == "pdf":
                with open(file_path, 'rb') as file:
                    with pdfplumber.open(file) as pdf:
                        text = ""
                        for page in pdf.pages:
                            text += page.extract_text() or ""
                        
                        if not text.strip():
                            # Fallback to PyPDF2
                            file.seek(0)
                            reader = PyPDF2.PdfReader(file)
                            text = ""
                            for page in reader.pages:
                                text += page.extract_text() or ""
                        return text
                        
            elif file_type == "docx":
                doc = Document(file_path)
                return " ".join([paragraph.text for paragraph in doc.paragraphs])
                
            elif file_type == "txt":
                with open(file_path, 'r', encoding='utf-8') as file:
                    return file.read()
                    
            elif file_type == "csv":
                with open(file_path, 'r', encoding='utf-8') as file:
                    csv_reader = csv.reader(file)
                    return " ".join([" ".join(row) for row in csv_reader])
                    
        except Exception as e:
            st.error(f"Error extracting text from {file_path}: {str(e)}")
            return ""
    
    def get_embedding(self, text):
        """Generate embedding for text using BGE model"""
        if self.embedding_model is None:
            st.error("No embedding model loaded!")
            return np.zeros(1024)  # BGE-large dimension
            
        try:
            # BGE models recommend adding instruction for retrieval
            # For queries (job description)
            if len(text) < 500:  # Assuming shorter texts are queries
                text = "Represent this sentence for searching relevant passages: " + text
            
            # Truncate text to avoid memory issues
            text = text[:8192] if text else ""
            
            # Generate embedding
            embedding = self.embedding_model.encode(text, 
                                                  convert_to_numpy=True,
                                                  normalize_embeddings=True)
            return embedding
                    
        except Exception as e:
            st.error(f"Error generating embedding: {str(e)}")
            return np.zeros(1024)  # BGE-large dimension
    
    def calculate_bm25_scores(self, resume_texts, job_description):
        """Calculate BM25 scores for keyword matching"""
        try:
            job_tokens = word_tokenize(job_description.lower())
            corpus = [word_tokenize(text.lower()) for text in resume_texts if text and text.strip()]
            
            if not corpus:
                return [0.0] * len(resume_texts)
                
            bm25 = BM25Okapi(corpus)
            scores = bm25.get_scores(job_tokens)
            return scores.tolist()
            
        except Exception as e:
            st.error(f"Error calculating BM25 scores: {str(e)}")
            return [0.0] * len(resume_texts)
    
    def advanced_pipeline_ranking(self, resume_texts, job_description, final_top_k=5):
        """Advanced pipeline: FAISS recall -> Cross-encoder -> BM25 -> LLM intent -> Final ranking"""
        if not resume_texts:
            return []
            
        # Stage 1: FAISS Recall (Top 50)
        st.write("🔍 **Stage 1**: FAISS Recall - Finding top 50 candidates...")
        top_50_indices = self.faiss_recall(resume_texts, job_description, top_k=50)
        
        # Stage 2: Cross-Encoder Re-ranking (Top 20)
        st.write("🎯 **Stage 2**: Cross-Encoder Re-ranking - Selecting top 20...")
        top_20_results = self.cross_encoder_rerank(resume_texts, job_description, top_50_indices, top_k=20)
        
        # Stage 3: BM25 Keyword Matching
        st.write("🔤 **Stage 3**: BM25 Keyword Matching...")
        top_20_with_bm25 = self.add_bm25_scores(resume_texts, job_description, top_20_results)
        
        # Stage 4: LLM Intent Analysis (using Qwen3-1.7B)
        st.write("🤖 **Stage 4**: LLM Intent Analysis...")
        top_20_with_intent = self.add_intent_scores(resume_texts, job_description, top_20_with_bm25)
        
        # Stage 5: Final Combined Ranking
        st.write(f"🏆 **Stage 5**: Final Combined Ranking - Selecting top {final_top_k}...")
        final_results = self.calculate_final_scores(top_20_with_intent)
        
        return final_results[:final_top_k]  # Return top K as selected by user
    
    def faiss_recall(self, resume_texts, job_description, top_k=50):
        """Stage 1: Use FAISS for initial recall to find top 50 resumes"""
        try:
            # Get job embedding
            job_embedding = self.get_embedding(job_description)
            
            # Get resume embeddings
            resume_embeddings = []
            progress_bar = st.progress(0)
            
            for i, text in enumerate(resume_texts):
                if text:
                    embedding = self.embedding_model.encode(text[:8192], 
                                                          convert_to_numpy=True,
                                                          normalize_embeddings=True)
                    resume_embeddings.append(embedding)
                else:
                    resume_embeddings.append(np.zeros(1024))
                progress_bar.progress((i + 1) / len(resume_texts))
            
            progress_bar.empty()
            
            # Create FAISS index
            resume_embeddings = np.array(resume_embeddings).astype('float32')
            dimension = resume_embeddings.shape[1]
            index = faiss.IndexFlatIP(dimension)  # Inner product for cosine similarity
            index.add(resume_embeddings)
            
            # Search for top K
            job_embedding = job_embedding.reshape(1, -1).astype('float32')
            scores, indices = index.search(job_embedding, min(top_k, len(resume_texts)))
            
            return indices[0].tolist()
            
        except Exception as e:
            st.error(f"Error in FAISS recall: {str(e)}")
            # Fallback: return all indices
            return list(range(min(top_k, len(resume_texts))))
    
    def cross_encoder_rerank(self, resume_texts, job_description, top_50_indices, top_k=20):
        """Stage 2: Use Cross-Encoder to re-rank top 50 and select top 20"""
        try:
            if not self.cross_encoder:
                st.error("Cross-encoder not loaded!")
                return [(idx, 0.0) for idx in top_50_indices[:top_k]]
            
            # Prepare pairs for cross-encoder
            pairs = []
            valid_indices = []
            
            for idx in top_50_indices:
                if idx < len(resume_texts) and resume_texts[idx]:
                    # Truncate texts for cross-encoder
                    job_snippet = job_description[:512]
                    resume_snippet = resume_texts[idx][:512]
                    pairs.append([job_snippet, resume_snippet])
                    valid_indices.append(idx)
            
            if not pairs:
                return [(idx, 0.0) for idx in top_50_indices[:top_k]]
            
            # Get cross-encoder scores
            progress_bar = st.progress(0)
            scores = []
            
            # Process in batches to avoid memory issues
            batch_size = 8
            for i in range(0, len(pairs), batch_size):
                batch = pairs[i:i+batch_size]
                batch_scores = self.cross_encoder.predict(batch)
                scores.extend(batch_scores)
                progress_bar.progress(min(1.0, (i + batch_size) / len(pairs)))
            
            progress_bar.empty()
            
            # Combine indices with scores and sort
            indexed_scores = list(zip(valid_indices, scores))
            indexed_scores.sort(key=lambda x: x[1], reverse=True)
            
            return indexed_scores[:top_k]
            
        except Exception as e:
            st.error(f"Error in cross-encoder re-ranking: {str(e)}")
            return [(idx, 0.0) for idx in top_50_indices[:top_k]]
    
    def add_bm25_scores(self, resume_texts, job_description, top_20_results):
        """Stage 3: Add BM25 scores to top 20 resumes"""
        try:
            # Get texts for top 20
            top_20_texts = [resume_texts[idx] for idx, _ in top_20_results]
            
            # Calculate BM25 scores
            bm25_scores = self.calculate_bm25_scores(top_20_texts, job_description)
            
            # Normalize BM25 scores to 0.1-0.2 range
            if bm25_scores and max(bm25_scores) > 0:
                max_bm25 = max(bm25_scores)
                min_bm25 = min(bm25_scores)
                if max_bm25 > min_bm25:
                    normalized_bm25 = [
                        0.1 + 0.1 * (score - min_bm25) / (max_bm25 - min_bm25) 
                        for score in bm25_scores
                    ]
                else:
                    normalized_bm25 = [0.15] * len(bm25_scores)
            else:
                normalized_bm25 = [0.15] * len(top_20_results)
            
            # Combine with existing results
            results_with_bm25 = []
            for i, (idx, cross_score) in enumerate(top_20_results):
                bm25_score = normalized_bm25[i] if i < len(normalized_bm25) else 0.15
                results_with_bm25.append((idx, cross_score, bm25_score))
            
            return results_with_bm25
            
        except Exception as e:
            st.error(f"Error adding BM25 scores: {str(e)}")
            return [(idx, cross_score, 0.15) for idx, cross_score in top_20_results]
    
    def add_intent_scores(self, resume_texts, job_description, top_20_with_bm25):
        """Stage 4: Add LLM intent analysis scores"""
        try:
            results_with_intent = []
            progress_bar = st.progress(0)
            
            for i, (idx, cross_score, bm25_score) in enumerate(top_20_with_bm25):
                intent_score = self.analyze_intent(resume_texts[idx], job_description)
                results_with_intent.append((idx, cross_score, bm25_score, intent_score))
                progress_bar.progress((i + 1) / len(top_20_with_bm25))
            
            progress_bar.empty()
            return results_with_intent
            
        except Exception as e:
            st.error(f"Error adding intent scores: {str(e)}")
            return [(idx, cross_score, bm25_score, 0.1) for idx, cross_score, bm25_score in top_20_with_bm25]
    
    def analyze_intent(self, resume_text, job_description):
        """Analyze candidate's intent using LLM"""
        try:
            # Truncate texts
            resume_snippet = resume_text[:1500] if len(resume_text) > 1500 else resume_text
            job_snippet = job_description[:800] if len(job_description) > 800 else job_description
            
            prompt = f"""You are given a job description and a candidate's resume.
Clearly answer: "Is the candidate likely seeking this job? Respond with 'Yes', 'Maybe', or 'No' and give a brief justification."

Job Description:
{job_snippet}

Candidate Resume:
{resume_snippet}

Response format:
Intent: [Yes/Maybe/No]
Reason: [Brief justification]"""
            
            response = generate_qwen3_response(
                prompt,
                st.session_state.qwen3_intent_tokenizer,
                st.session_state.qwen3_intent_model,
                max_new_tokens=100
            )
            
            # Parse response
            response_lower = response.lower()
            if 'intent: yes' in response_lower or 'intent:yes' in response_lower:
                return 0.3
            elif 'intent: maybe' in response_lower or 'intent:maybe' in response_lower:
                return 0.1
            else:
                return 0.0
                
        except Exception as e:
            st.warning(f"Error analyzing intent: {str(e)}")
            return 0.1  # Default to "Maybe"
    
    def calculate_final_scores(self, results_with_all_scores):
        """Stage 5: Calculate final combined scores"""
        try:
            final_results = []
            
            for idx, cross_score, bm25_score, intent_score in results_with_all_scores:
                # Normalize cross-encoder score to 0-1 range
                normalized_cross = max(0, min(1, cross_score))
                
                # Final Score = Cross-Encoder (0-1) + BM25 (0.1-0.2) + Intent (0-0.3)
                final_score = normalized_cross + bm25_score + intent_score
                
                final_results.append({
                    'index': idx,
                    'cross_encoder_score': normalized_cross,
                    'bm25_score': bm25_score,
                    'intent_score': intent_score,
                    'final_score': final_score
                })
            
            # Sort by final score
            final_results.sort(key=lambda x: x['final_score'], reverse=True)
            
            return final_results
            
        except Exception as e:
            st.error(f"Error calculating final scores: {str(e)}")
            return []
    
    def extract_skills(self, text, job_description):
        """Extract skills from resume based on job description"""
        if not text:
            return []
            
        # Common tech skills
        common_skills = [
            "python", "java", "javascript", "react", "angular", "vue", "node.js", 
            "express", "django", "flask", "spring", "sql", "nosql", "html", "css",
            "aws", "azure", "gcp", "docker", "kubernetes", "jenkins", "git", "github",
            "agile", "scrum", "jira", "ci/cd", "devops", "microservices", "rest", "api",
            "machine learning", "deep learning", "data science", "artificial intelligence", 
            "tensorflow", "pytorch", "keras", "scikit-learn", "pandas", "numpy", 
            "matplotlib", "seaborn", "jupyter", "r", "sas", "spss", "tableau", "powerbi",
            "excel", "mysql", "postgresql", "mongodb", "redis", "elasticsearch", 
            "kafka", "rabbitmq", "spark", "hadoop", "hive", "airflow", "linux", "unix"
        ]
        
        # Extract potential skills from job description
        job_words = set(word.lower() for word in word_tokenize(job_description) if len(word) > 2)
        
        # Find matching skills
        found_skills = []
        text_lower = text.lower()
        
        # Check common skills that appear in both resume and job description
        for skill in common_skills:
            if skill in text_lower and any(skill in job_word for job_word in job_words):
                found_skills.append(skill)
        
        # Check for skills mentioned in job description
        for word in job_words:
            if len(word) > 3 and word in text_lower and word not in found_skills:
                # Basic filter to avoid common words
                if word not in ['with', 'have', 'that', 'this', 'from', 'what', 'when', 'where']:
                    found_skills.append(word)
        
        return list(set(found_skills))[:15]  # Return top 15 unique skills
    
    def generate_simple_explanation(self, score, semantic_score, bm25_score, skills):
        """Generate simple explanation for the match (fallback)"""
        if score > 0.8:
            quality = "excellent"
        elif score > 0.6:
            quality = "strong"
        elif score > 0.4:
            quality = "moderate"
        else:
            quality = "limited"
        
        explanation = f"This candidate shows {quality} alignment with the position (score: {score:.2f}). "
        
        if semantic_score > bm25_score:
            explanation += f"The resume demonstrates strong conceptual relevance ({semantic_score:.2f}) suggesting good experience fit. "
        else:
            explanation += f"The resume has high keyword match ({bm25_score:.2f}) indicating direct skill alignment. "
        
        if skills:
            explanation += f"Key matching competencies include: {', '.join(skills[:5])}."
        
        return explanation
    
    def generate_llm_explanation(self, resume_text, job_description, score, skills, max_retries=3):
        """Generate detailed explanation using Qwen3-14B"""
        if not st.session_state.qwen3_model:
            return self.generate_simple_explanation(score, score, score, skills)
        
        # Truncate texts to manage token limits
        resume_snippet = resume_text[:2000] if len(resume_text) > 2000 else resume_text
        job_snippet = job_description[:1000] if len(job_description) > 1000 else job_description
        
        prompt = f"""You are an expert HR analyst. Analyze this individual candidate's resume against the job requirements and write EXACTLY 150 words explaining why this specific candidate is suitable for the position.

Structure your 150-word analysis as follows:
1. Experience alignment (40-50 words)
2. Key strengths and skills match (40-50 words) 
3. Unique value proposition (40-50 words)
4. Overall recommendation (10-20 words)

Job Requirements:
{job_snippet}

Candidate's Resume:
{resume_snippet}

Identified Matching Skills: {', '.join(skills[:10])}
Compatibility Score: {score:.1%}

Write a professional, detailed 150-word analysis for THIS INDIVIDUAL CANDIDATE:"""
        
        for attempt in range(max_retries):
            try:
                response = generate_qwen3_response(
                    prompt,
                    st.session_state.qwen3_tokenizer,
                    st.session_state.qwen3_model,
                    max_new_tokens=200
                )
                
                # Extract the response and ensure it's about 150 words
                explanation = response.strip()
                word_count = len(explanation.split())
                
                # If response is close to 150 words (130-170), accept it
                if 130 <= word_count <= 170:
                    return explanation
                
                # If response is too short or too long, try again with adjusted prompt
                if word_count < 130:
                    # Response too short, try again
                    continue
                elif word_count > 170:
                    # Response too long, truncate to approximately 150 words
                    words = explanation.split()
                    truncated = ' '.join(words[:150])
                    # Add proper ending if truncated
                    if not truncated.endswith('.'):
                        truncated += '.'
                    return truncated
                
                return explanation
                
            except Exception as e:
                if attempt < max_retries - 1:
                    time.sleep(2)  # Wait before retry
                    continue
                else:
                    # Fallback to simple explanation
                    return self.generate_simple_explanation(score, score, score, skills)
        
        # If all retries failed, use simple explanation
        return self.generate_simple_explanation(score, score, score, skills)

def create_download_link(df, filename="resume_screening_results.csv"):
    """Create download link for results"""
    csv = df.to_csv(index=False)
    b64 = base64.b64encode(csv.encode()).decode()
    return f'<a href="data:file/csv;base64,{b64}" download="{filename}" class="download-btn">📥 Download Results CSV</a>'

# Main App Interface
st.title("🎯 AI-Powered Resume Screener")
st.markdown("*Find the perfect candidates using BAAI/bge-large-en-v1.5 embeddings and Qwen3-14B explanations*")
st.markdown("---")

# Initialize screener
screener = ResumeScreener()

# Job Description Input
st.header("📝 Step 1: Enter Job Description")
job_description = st.text_area(
    "Enter the complete job description or requirements:",
    height=150,
    placeholder="Paste the job description here, including required skills, experience, and qualifications..."
)

# Resume Input Options
st.header("📄 Step 2: Upload Resumes")

# Show loaded resumes indicator
if st.session_state.resume_texts:
    col1, col2 = st.columns([3, 1])
    with col1:
        st.info(f"📚 {len(st.session_state.resume_texts)} resumes loaded and ready for analysis")
    with col2:
        if st.button("🗑️ Clear Resumes", type="secondary", help="Clear all loaded resumes to start fresh"):
            st.session_state.resume_texts = []
            st.session_state.file_names = []
            st.session_state.results = []
            st.session_state.explanations_generated = False
            st.session_state.current_job_description = ""
            st.rerun()

input_method = st.radio(
    "Choose input method:",
    ["📁 Upload Files", "🗂️ Load from CSV Dataset", "🔗 Load from Hugging Face Dataset"]
)

if input_method == "📁 Upload Files":
    uploaded_files = st.file_uploader(
        "Upload resume files",
        type=["pdf", "docx", "txt"],
        accept_multiple_files=True,
        help="Supported formats: PDF, DOCX, TXT"
    )
    
    if uploaded_files:
        with st.spinner(f"🔄 Processing {len(uploaded_files)} files..."):
            resume_texts = []
            file_names = []
            
            for file in uploaded_files:
                file_type = file.name.split('.')[-1].lower()
                
                with tempfile.NamedTemporaryFile(delete=False, suffix=f'.{file_type}') as tmp_file:
                    tmp_file.write(file.getvalue())
                    tmp_path = tmp_file.name
                
                text = screener.extract_text_from_file(tmp_path, file_type)
                if text.strip():
                    resume_texts.append(text)
                    file_names.append(file.name)
                
                os.unlink(tmp_path)
            
            st.session_state.resume_texts = resume_texts
            st.session_state.file_names = file_names
            
        if resume_texts:
            st.success(f"✅ Successfully processed {len(resume_texts)} resumes")

elif input_method == "🗂️ Load from CSV Dataset":
    csv_file = st.file_uploader("Upload CSV file with resume data", type=["csv"])
    
    if csv_file:
        try:
            df = pd.read_csv(csv_file)
            st.write("**CSV Preview:**")
            st.dataframe(df.head())
            
            text_column = st.selectbox(
                "Select column containing resume text:",
                df.columns.tolist()
            )
            
            name_column = st.selectbox(
                "Select column for candidate names/IDs (optional):",
                ["Use Index"] + df.columns.tolist()
            )
            
            if st.button("🚀 Process CSV Data"):
                with st.spinner("🔄 Processing CSV data..."):
                    resume_texts = []
                    file_names = []
                    
                    for idx, row in df.iterrows():
                        text = str(row[text_column])
                        if text and text.strip() and text.lower() != 'nan':
                            resume_texts.append(text)
                            
                            if name_column == "Use Index":
                                file_names.append(f"Resume_{idx}")
                            else:
                                file_names.append(str(row[name_column]))
                    
                    st.session_state.resume_texts = resume_texts
                    st.session_state.file_names = file_names
                
                if resume_texts:
                    st.success(f"✅ Successfully loaded {len(resume_texts)} resumes from CSV")
                    
        except Exception as e:
            st.error(f"❌ Error processing CSV: {str(e)}")

elif input_method == "🔗 Load from Hugging Face Dataset":
    st.markdown("**Popular Resume Datasets:**")
    st.markdown("- `ahmedheakl/resume-atlas`")
    st.markdown("- `InferenceFly/Resume-Dataset`")
    
    col1, col2 = st.columns([2, 1])
    with col1:
        dataset_name = st.text_input(
            "Dataset name:",
            value="ahmedheakl/resume-atlas",
            help="Enter Hugging Face dataset name"
        )
    with col2:
        dataset_split = st.selectbox("Split:", ["train", "test", "validation"], index=0)
    
    if st.button("🔗 Load from Hugging Face"):
        try:
            with st.spinner(f"🔄 Loading {dataset_name}..."):
                dataset = load_dataset(dataset_name, split=dataset_split)
                
            st.success(f"✅ Loaded dataset with {len(dataset)} entries")
            st.write("**Dataset Preview:**")
            
            preview_df = pd.DataFrame(dataset[:5])
            st.dataframe(preview_df)
            
            text_column = st.selectbox(
                "Select column with resume text:",
                dataset.column_names,
                index=dataset.column_names.index('resume_text') if 'resume_text' in dataset.column_names else 0
            )
            
            category_column = None
            if 'category' in dataset.column_names:
                categories = list(set(dataset['category']))
                category_column = st.selectbox(
                    "Filter by category (optional):",
                    ["All"] + categories
                )
            
            max_samples = st.slider("Maximum samples to load:", 10, min(1000, len(dataset)), 100)
            
            if st.button("🚀 Process Dataset"):
                with st.spinner("🔄 Processing dataset..."):
                    resume_texts = []
                    file_names = []
                    
                    filtered_dataset = dataset
                    
                    if category_column and category_column != "All":
                        filtered_dataset = dataset.filter(lambda x: x['category'] == category_column)
                    
                    sample_indices = list(range(min(max_samples, len(filtered_dataset))))
                    
                    for idx in sample_indices:
                        item = filtered_dataset[idx]
                        text = str(item[text_column])
                        
                        if text and text.strip() and text.lower() != 'nan':
                            resume_texts.append(text)
                            
                            if 'id' in item:
                                file_names.append(f"Resume_{item['id']}")
                            else:
                                file_names.append(f"Resume_{idx}")
                    
                    st.session_state.resume_texts = resume_texts
                    st.session_state.file_names = file_names
                
                if resume_texts:
                    st.success(f"✅ Successfully loaded {len(resume_texts)} resumes")
                    
        except Exception as e:
            st.error(f"❌ Error loading dataset: {str(e)}")

# Processing and Results
st.header("🔍 Step 3: Analyze Resumes")

# First button: Find top K candidates (fast ranking)
col1, col2 = st.columns([1, 1])

with col1:
    if st.button("🚀 Advanced Pipeline Analysis", 
                 disabled=not (job_description and st.session_state.resume_texts),
                 type="primary",
                 help="Run the complete 5-stage advanced pipeline"):
        if len(st.session_state.resume_texts) == 0:
            st.error("❌ Please upload resumes first!")
        elif not job_description.strip():
            st.error("❌ Please enter a job description!")
        else:
            with st.spinner("🚀 Running Advanced Pipeline Analysis..."):
                try:
                    # Run the advanced pipeline
                    pipeline_results = screener.advanced_pipeline_ranking(
                        st.session_state.resume_texts, job_description, final_top_k=top_k
                    )
                    
                    # Prepare results for display
                    results = []
                    
                    for rank, result_data in enumerate(pipeline_results, 1):
                        idx = result_data['index']
                        name = st.session_state.file_names[idx]
                        text = st.session_state.resume_texts[idx]
                        
                        # Extract skills
                        skills = screener.extract_skills(text, job_description)
                        
                        results.append({
                            'rank': rank,
                            'name': name,
                            'final_score': result_data['final_score'],
                            'cross_encoder_score': result_data['cross_encoder_score'],
                            'bm25_score': result_data['bm25_score'],
                            'intent_score': result_data['intent_score'],
                            'skills': skills,
                            'text': text,
                            'text_preview': text[:500] + "..." if len(text) > 500 else text,
                            'explanation': None  # No detailed explanation yet
                        })
                    
                    # Add simple explanations for now
                    for result in results:
                        result['explanation'] = screener.generate_simple_explanation(
                            result['final_score'], 
                            result['cross_encoder_score'], 
                            result['bm25_score'], 
                            result['skills']
                        )
                    
                    # Store in session state
                    st.session_state.results = results
                    st.session_state.explanations_generated = False
                    st.session_state.current_job_description = job_description
                    
                    st.success(f"🚀 Advanced pipeline complete! Found top {len(st.session_state.results)} candidates.")
                    
                except Exception as e:
                    st.error(f"❌ Error during analysis: {str(e)}")

# Second button: Generate AI explanations (slower, optional)
with col2:
    # Show this button only if we have results and LLM is enabled
    show_explanation_button = (
        st.session_state.results and 
        use_llm_explanations and 
        st.session_state.qwen3_model and
        not st.session_state.explanations_generated
    )
    
    if show_explanation_button:
        if st.button("🤖 Generate AI Explanations", 
                     type="secondary",
                     help="Generate detailed 150-word explanations using Qwen3-14B (takes longer)"):
            with st.spinner("🤖 Generating detailed AI explanations..."):
                try:
                    explanation_progress = st.progress(0)
                    explanation_text = st.empty()
                    
                    for i, result in enumerate(st.session_state.results):
                        explanation_text.text(f"🤖 Generating AI explanation for candidate {i+1}/{len(st.session_state.results)}...")
                        
                        llm_explanation = screener.generate_llm_explanation(
                            result['text'],
                            st.session_state.current_job_description,
                            result['final_score'],
                            result['skills']
                        )
                        result['explanation'] = llm_explanation
                        
                        explanation_progress.progress((i + 1) / len(st.session_state.results))
                    
                    explanation_progress.empty()
                    explanation_text.empty()
                    
                    # Mark explanations as generated
                    st.session_state.explanations_generated = True
                    
                    st.success(f"🤖 AI explanations generated for all {len(st.session_state.results)} candidates!")
                    
                except Exception as e:
                    st.error(f"❌ Error generating explanations: {str(e)}")
    
    elif st.session_state.results and st.session_state.explanations_generated:
        st.info("✅ AI explanations already generated!")
    
    elif st.session_state.results and not use_llm_explanations:
        st.info("💡 Enable 'Generate AI Explanations' in sidebar to use this feature")
    
    elif st.session_state.results and not st.session_state.qwen3_model:
        st.warning("⚠️ LLM model not available. Check your Hugging Face token.")

# Display Results
if st.session_state.results:
    st.header("🏆 Top Candidates")
    
    # Create tabs for different views
    tab1, tab2, tab3 = st.tabs(["📊 Summary", "📋 Detailed Analysis", "📈 Visualizations"])
    
    with tab1:
        # Create summary dataframe with new scoring system
        summary_data = []
        for result in st.session_state.results:
            # Map intent score to text
            intent_text = "Yes" if result['intent_score'] == 0.3 else "Maybe" if result['intent_score'] == 0.1 else "No"
            
            summary_data.append({
                "Rank": result['rank'],
                "Candidate": result['name'],
                "Final Score": f"{result['final_score']:.2f}",
                "Cross-Encoder": f"{result['cross_encoder_score']:.2f}",
                "BM25": f"{result['bm25_score']:.2f}",
                "Intent": f"{intent_text} ({result['intent_score']:.1f})",
                "Top Skills": ", ".join(result['skills'][:5])
            })
        
        summary_df = pd.DataFrame(summary_data)
        
        # Style the dataframe
        def color_scores(val):
            if isinstance(val, str) and any(char.isdigit() for char in val):
                try:
                    # Extract numeric value
                    numeric_val = float(''.join(c for c in val if c.isdigit() or c == '.'))
                    if 'Final Score' in val or numeric_val >= 1.0:
                        if numeric_val >= 1.2:
                            return 'background-color: #d4edda'
                        elif numeric_val >= 1.0:
                            return 'background-color: #fff3cd'
                        else:
                            return 'background-color: #f8d7da'
                    else:
                        if numeric_val >= 0.7:
                            return 'background-color: #d4edda'
                        elif numeric_val >= 0.5:
                            return 'background-color: #fff3cd'
                        else:
                            return 'background-color: #f8d7da'
                except:
                    pass
            return ''
        
        styled_df = summary_df.style.applymap(color_scores, subset=['Final Score', 'Cross-Encoder', 'BM25'])
        st.dataframe(styled_df, use_container_width=True)
        
        # Download link
        detailed_data = []
        for result in st.session_state.results:
            intent_text = "Yes" if result['intent_score'] == 0.3 else "Maybe" if result['intent_score'] == 0.1 else "No"
            
            detailed_data.append({
                "Rank": result['rank'],
                "Candidate": result['name'],
                "Final_Score": result['final_score'],
                "Cross_Encoder_Score": result['cross_encoder_score'],
                "BM25_Score": result['bm25_score'],
                "Intent_Score": result['intent_score'],
                "Intent_Analysis": intent_text,
                "Skills": "; ".join(result['skills']),
                "AI_Explanation": result['explanation'],
                "Resume_Preview": result['text_preview']
            })
        
        download_df = pd.DataFrame(detailed_data)
        st.markdown(create_download_link(download_df), unsafe_allow_html=True)
    
    with tab2:
        # Detailed results with new scoring breakdown
        for result in st.session_state.results:
            intent_text = "Yes" if result['intent_score'] == 0.3 else "Maybe" if result['intent_score'] == 0.1 else "No"
            
            with st.expander(f"#{result['rank']}: {result['name']} (Final Score: {result['final_score']:.2f})"):
                col1, col2 = st.columns([1, 2])
                
                with col1:
                    st.metric("🏆 Final Score", f"{result['final_score']:.2f}")
                    
                    st.write("**📊 Score Breakdown:**")
                    st.metric("🎯 Cross-Encoder", f"{result['cross_encoder_score']:.2f}", help="Semantic relevance (0-1)")
                    st.metric("🔤 BM25 Keywords", f"{result['bm25_score']:.2f}", help="Keyword matching (0.1-0.2)")
                    st.metric("🤖 Intent Analysis", f"{intent_text} ({result['intent_score']:.1f})", help="Job seeking likelihood (0-0.3)")
                    
                    st.write("**🎯 Matching Skills:**")
                    skills_per_column = 5
                    skill_cols = st.columns(2)
                    for idx, skill in enumerate(result['skills'][:10]):
                        with skill_cols[idx % 2]:
                            st.write(f"• {skill}")
                
                with col2:
                    st.write("**💡 AI-Generated Match Analysis:**")
                    st.info(result['explanation'])
                    
                    st.write("**📄 Resume Preview:**")
                    st.text_area("", result['text_preview'], height=200, disabled=True, key=f"preview_{result['rank']}")
    
    with tab3:
        # Score visualization
        if len(st.session_state.results) > 1:
            # Bar chart
            st.subheader("Score Comparison")
            
            chart_data = pd.DataFrame({
                'Candidate': [r['name'][:20] + '...' if len(r['name']) > 20 else r['name'] 
                             for r in st.session_state.results],
                'Final Score': [r['final_score'] for r in st.session_state.results],
                'Cross-Encoder': [r['cross_encoder_score'] for r in st.session_state.results],
                'BM25': [r['bm25_score'] for r in st.session_state.results],
                'Intent': [r['intent_score'] for r in st.session_state.results]
            })
            
            st.bar_chart(chart_data.set_index('Candidate'))
            
            # Score distribution
            col1, col2 = st.columns(2)
            
            with col1:
                st.subheader("Score Distribution")
                score_ranges = {
                    'Excellent (≥1.2)': sum(1 for r in st.session_state.results if r['final_score'] >= 1.2),
                    'Good (1.0-1.2)': sum(1 for r in st.session_state.results if 1.0 <= r['final_score'] < 1.2),
                    'Fair (0.8-1.0)': sum(1 for r in st.session_state.results if 0.8 <= r['final_score'] < 1.0),
                    'Poor (<0.8)': sum(1 for r in st.session_state.results if r['final_score'] < 0.8),
                }
                
                dist_df = pd.DataFrame({
                    'Range': score_ranges.keys(),
                    'Count': score_ranges.values()
                })
                st.bar_chart(dist_df.set_index('Range'))
            
            with col2:
                st.subheader("Average Scores")
                avg_final = np.mean([r['final_score'] for r in st.session_state.results])
                avg_cross = np.mean([r['cross_encoder_score'] for r in st.session_state.results])
                avg_bm25 = np.mean([r['bm25_score'] for r in st.session_state.results])
                avg_intent = np.mean([r['intent_score'] for r in st.session_state.results])
                
                st.metric("Average Final Score", f"{avg_final:.2f}")
                st.metric("Average Cross-Encoder", f"{avg_cross:.2f}")
                st.metric("Average BM25", f"{avg_bm25:.2f}")
                st.metric("Average Intent", f"{avg_intent:.2f}")

# Memory cleanup
st.markdown("---")
st.subheader("🧹 Reset Application")
col1, col2, col3 = st.columns([1, 1, 3])
with col1:
    if st.button("🗑️ Clear Resumes Only", type="secondary", help="Clear only the loaded resumes"):
        st.session_state.resume_texts = []
        st.session_state.file_names = []
        st.session_state.results = []
        st.session_state.explanations_generated = False
        st.session_state.current_job_description = ""
        st.success("✅ Resumes cleared!")
        st.rerun()

with col2:
    if st.button("🧹 Clear Everything", type="primary", help="Clear all data and free memory"):
        st.session_state.resume_texts = []
        st.session_state.file_names = []
        st.session_state.results = []
        st.session_state.explanations_generated = False
        st.session_state.current_job_description = ""
        
        if torch.cuda.is_available():
            torch.cuda.empty_cache()
        gc.collect()
        st.success("✅ Everything cleared!")
        st.rerun()

# Footer
st.markdown("---")
st.markdown(
    """
    <div style='text-align: center; color: #666;'>
        🚀 Powered by BAAI/bge-large-en-v1.5 & Qwen3-14B | Built with Streamlit
    </div>
    """, 
    unsafe_allow_html=True
)