handler.py

import os
import json
import torch
from typing import Dict, List, Any, Optional, Union
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig

# Import PEFT for adapter handling
try:
    import peft
    from peft import PeftModel, PeftConfig
    PEFT_AVAILABLE = True
except ImportError:
    PEFT_AVAILABLE = False
    print("Warning: PEFT library not available. Adapter loading may fail.")


class EndpointHandler:
    def __init__(self, model_dir=None):
        """
        Initialize the endpoint handler.
        Args:
            model_dir: Directory where model and tokenizer are stored
        """
        self.model = None
        self.tokenizer = None
        self.device = "cuda" if torch.cuda.is_available() else "cpu"
        self.model_dir = model_dir or os.getenv("MODEL_PATH", "/model")
        
        # GPU performance optimization flags
        self.flash_attention_supported = False  # Will be set during model loading
        self.use_sampling = True                # Better quality but slightly slower than greedy
        
        # Load model immediately
        self.load_model()
        
    def generate_optimized(self, inputs, attention_mask=None, max_new_tokens=512):
        """
        Optimized generation function that maximizes GPU utilization
        while respecting model constraints.
        """
        # Check if we need to create an attention mask
        if attention_mask is None:
            attention_mask = inputs.ne(self.tokenizer.pad_token_id).long()
        
        # Find input length to properly calculate output length
        input_length = inputs.shape[1]
        
        # Basic generation parameters
        generation_kwargs = {
            "inputs": inputs,
            "attention_mask": attention_mask,
            "max_new_tokens": max_new_tokens,
            
            # Performance options
            "use_cache": True,                         # Use KV cache for faster generation
            
            # Quality vs. speed tradeoff
            "temperature": 0.7 if self.use_sampling else 1.0,
            "top_p": 0.9 if self.use_sampling else 1.0,
            "do_sample": self.use_sampling,            # Sampling is slightly slower but better quality
            "num_beams": 1,                            # Beam search is slower but better quality (1 = no beam search)
            
            # Token handling
            "pad_token_id": self.tokenizer.pad_token_id,
            "eos_token_id": self.tokenizer.eos_token_id,
            
            # Content quality
            "repetition_penalty": 1.1,                 # Reduce repetition
        }
        
        # Add Flash Attention parameters only if supported by the transformers version
        # We check the transformer version by testing in a safe way
        try:
            import importlib
            transformers_version = importlib.import_module('transformers').__version__
            major, minor = map(int, transformers_version.split('.')[:2])
            
            if major > 4 or (major == 4 and minor >= 32):
                # Flash Attention support was added in transformers 4.32.0
                if self.flash_attention_supported:
                    print("Using Flash Attention in generation")
                    generation_kwargs["flash_attn"] = True
                    generation_kwargs["flash_attn_cross_entropy"] = True
            else:
                print(f"Flash Attention not added - transformers version {transformers_version} doesn't support it")
        except Exception as e:
            print(f"Error checking transformers version, skipping Flash Attention: {e}")
        
        # Generate with optimized parameters for GPU performance
        outputs = self.model.generate(**generation_kwargs)
        
        return outputs, input_length
    
    def load_model(self):
        """Load the finetuned model and tokenizer."""
        try:
            print(f"Loading model from {self.model_dir} to {self.device}...")
            
            # Load tokenizer with explicit padding token configuration
            try:
                self.tokenizer = AutoTokenizer.from_pretrained(
                    self.model_dir,
                    padding_side="left",  # Set padding to left side for causal LM
                    trust_remote_code=False
                )
                
                # Ensure pad token is set properly (important for attention masks)
                if self.tokenizer.pad_token is None:
                    self.tokenizer.pad_token = self.tokenizer.eos_token
                    print("Set pad_token to eos_token")
                    
            except Exception as tokenizer_error:
                print(f"Error loading tokenizer from {self.model_dir}: {tokenizer_error}")
                print("Attempting to load base Phi-2 tokenizer...")
                
                # Fall back to base Phi-2 tokenizer if model dir tokenizer fails
                self.tokenizer = AutoTokenizer.from_pretrained(
                    "microsoft/phi-2",
                    padding_side="left",
                    trust_remote_code=False
                )
                if self.tokenizer.pad_token is None:
                    self.tokenizer.pad_token = self.tokenizer.eos_token
            
            # Try to load model with quantization with consistent dtype settings
            try:
                from bitsandbytes.nn import Linear4bit
                from transformers import BitsAndBytesConfig
                
                print("Using 4-bit quantization with float16 compute type")
                
                # Use consistent float16 for both compute and parameters
                quantization_config = BitsAndBytesConfig(
                    load_in_4bit=True,
                    bnb_4bit_compute_dtype=torch.float16,  # Match with model dtype
                    bnb_4bit_use_double_quant=True,
                    bnb_4bit_quant_type="nf4"
                )
                
                # Try to load with base model specification for better adapter compatibility
                if os.path.exists(os.path.join(self.model_dir, "adapter_model.safetensors")):
                    print("Found adapter model, loading Phi-2 base with adapter")
                    
                    # Check if PEFT is available - using the global variable
                    global PEFT_AVAILABLE
                    if not PEFT_AVAILABLE:
                        print("PEFT not available, installing...")
                        try:
                            import pip
                            pip.main(['install', 'peft'])
                            import peft
                            from peft import PeftModel, PeftConfig
                            PEFT_AVAILABLE = True
                        except Exception as e:
                            print(f"Failed to install PEFT: {e}")
                    
                    # First load base model with quantization
                    base_model = AutoModelForCausalLM.from_pretrained(
                        "microsoft/phi-2",
                        quantization_config=quantization_config,
                        torch_dtype=torch.float16,
                        device_map="auto"
                    )
                    
                    try:
                        # Then load adapter on top
                        self.model = PeftModel.from_pretrained(
                            base_model,
                            self.model_dir,
                            torch_dtype=torch.float16,
                            device_map="auto"
                        )
                        print("Successfully loaded adapter model")
                    except Exception as adapter_error:
                        print(f"Error loading adapter: {adapter_error}")
                        # Fall back to just using the base model
                        print("Falling back to base model without adapter")
                        self.model = base_model
                else:
                    # Load as a standard model if no adapter is found
                    print("Loading model directly from directory")
                    self.model = AutoModelForCausalLM.from_pretrained(
                        self.model_dir,
                        torch_dtype=torch.float16,
                        device_map="auto",
                        quantization_config=quantization_config
                    )
                
            except ImportError as e:
                print(f"Warning: Could not use bitsandbytes quantization, falling back to standard loading: {e}")
                
                # Fallback to standard FP16 loading without quantization
                try:
                    self.model = AutoModelForCausalLM.from_pretrained(
                        self.model_dir,
                        torch_dtype=torch.float16,
                        device_map="auto",
                    )
                except Exception as model_error:
                    print(f"Error loading from model directory: {model_error}")
                    print("Attempting to load base Phi-2 model...")
                    
                    # Final fallback - try loading just the base model
                    self.model = AutoModelForCausalLM.from_pretrained(
                        "microsoft/phi-2",
                        torch_dtype=torch.float16,
                        device_map="auto",
                    )
            
            # Check for Flash Attention support with better error handling
            try:
                # First check if the transformers version supports it
                import importlib
                transformers_version = importlib.import_module('transformers').__version__
                major, minor = map(int, transformers_version.split('.')[:2])
                
                if major > 4 or (major == 4 and minor >= 32):
                    # Flash Attention support was added in transformers 4.32.0
                    try:
                        import flash_attn
                        self.flash_attention_supported = True
                        print(f"Flash Attention {flash_attn.__version__} detected and will be used if available!")
                    except ImportError:
                        print("Flash Attention library not installed. Using standard attention mechanism.")
                        self.flash_attention_supported = False
                else:
                    print(f"Transformers version {transformers_version} doesn't support Flash Attention parameters. Using standard attention.")
                    self.flash_attention_supported = False
            except Exception as e:
                print(f"Error checking Flash Attention support: {e}")
                print("Falling back to standard attention mechanism.")
                self.flash_attention_supported = False
            
            # Enable TF32 precision for higher performance on newer NVIDIA GPUs
            if self.device == "cuda":
                # Only available on Ampere+ GPUs (A100, RTX 3090, etc.)
                try:
                    if torch.cuda.get_device_capability()[0] >= 8:
                        print("Enabling TF32 precision for faster matrix operations")
                        torch.backends.cuda.matmul.allow_tf32 = True
                        torch.backends.cudnn.allow_tf32 = True
                except Exception as e:
                    print(f"Error enabling TF32 precision: {e}")
            
            print(f"Model loaded successfully on {self.device}")
            return True
        except Exception as e:
            print(f"Error loading model: {e}")
            import traceback
            print(traceback.format_exc())
            return False

    def format_candidates_for_prompt(self, candidates: List[Dict[str, Any]]) -> str:
        """Format candidate information for the prompt."""
        candidate_summary = ""
        for i, candidate in enumerate(candidates):
            # Create a concise summary for each candidate
            summary = f"Candidate {i+1}: {candidate.get('name', 'Unnamed')}\n"
            
            if candidate.get('education'):
                edu = candidate['education'][0] if isinstance(candidate['education'], list) and len(candidate['education']) > 0 else candidate['education']
                summary += f"Education: {edu.get('institution', 'Unknown')}, {edu.get('degree', 'Unknown')}\n"
            
            if candidate.get('experience'):
                exp = candidate['experience'][0] if isinstance(candidate['experience'], list) and len(candidate['experience']) > 0 else candidate['experience']
                summary += f"Experience: {exp.get('title', 'Unknown')} at {exp.get('company', 'Unknown')}\n"
            
            if candidate.get('skills'):
                skills = candidate['skills'] if isinstance(candidate['skills'], list) else [candidate['skills']]
                summary += f"Skills: {', '.join(skills[:5])}\n"
            
            candidate_summary += summary + "\n"
        
        return candidate_summary

    def generate_team(self, candidates: List[Dict[str, Any]], requirements: str = None, team_size: int = 5) -> str:
        """Generate a team based on candidates and requirements."""
        if self.model is None or self.tokenizer is None:
            self.load_model()
            if self.model is None:
                return "Model could not be loaded. The AI analysis service is currently unavailable."
        
        try:
            # Format candidate information
            candidate_summary = self.format_candidates_for_prompt(candidates)
            
            # Build the enhanced prompt for more detailed analysis
            prompt = f"""Analyze these candidates and create THREE different optimal startup team compositions of {team_size} people each.

CANDIDATES:
{candidate_summary}

TEAM REQUIREMENTS:
{requirements or "Create a balanced team with complementary skills"}

For EACH team composition, please provide:
1. Team Name: Give this team composition a memorable name based on its strengths
2. Selected Members: List each selected team member with:
   - Their name
   - Recommended role in the team
   - 2-3 sentences on WHY they specifically are valuable to this team composition
   - How they complement other team members

3. Team Analysis (minimum 250 words):
   - Detailed strengths of this specific team combination
   - Potential weaknesses or challenges this team might face
   - Assessment of skill coverage and diversity of thinking
   - Team dynamics and how members would likely work together
   - How this team aligns with the stated requirements

4. Alternative Applications:
   - What type of startup would be MOST successful with this team
   - What type of startup would be LEAST successful with this team

After presenting all three team compositions, provide a final recommendation on which team would be best and why.

Format your response carefully with clear headings and make it comprehensive enough for founders to make informed decisions.
"""

            # Format as chat with improved system prompt
            messages = [
                {"role": "system", "content": "You are an elite startup advisor with deep expertise in team composition and founder dynamics. You specialize in analyzing candidate profiles and determining optimal team compositions that maximize chances of startup success."},
                {"role": "user", "content": prompt}
            ]
            
            # Convert to model input format
            inputs = self.tokenizer.apply_chat_template(
                messages, 
                return_tensors="pt"
            ).to(self.device)
            
            # Generate with proper context limits and attention masks
            with torch.no_grad():
                # Find input length to set appropriate output length
                input_length = inputs.shape[1]
                # Phi-2 has a context limit of 2048
                max_context_length = 2048
                
                # Calculate max new tokens to avoid exceeding model's context limits
                max_new_tokens = max(100, min(1024, max_context_length - input_length))
                
                print(f"Input length: {input_length}, Max new tokens: {max_new_tokens}")
                
                # Create attention mask (explicitly handle padding)
                attention_mask = inputs.ne(self.tokenizer.pad_token_id).long()
                
                # Use the optimized generator instead of direct model.generate call
                outputs, input_length = self.generate_optimized(
                    inputs,
                    attention_mask=attention_mask,
                    max_new_tokens=max_new_tokens
                )
                
            # Decode more carefully
            try:
                # Get only the generated part (exclude input tokens)
                generated_output = outputs[0][input_length:]
                
                # Decode just the new tokens
                generated_text = self.tokenizer.decode(
                    generated_output, 
                    skip_special_tokens=True,
                    clean_up_tokenization_spaces=True
                )
                
                # Remove any model-specific artifacts
                generated_text = generated_text.replace("<|im_end|>", "").replace("<|im_start|>", "")
                assistant_response = generated_text.strip()
                
                # If that failed, try traditional approach
                if not assistant_response:
                    full_response = self.tokenizer.decode(outputs[0], skip_special_tokens=True)
                    assistant_response = full_response.split(prompt)[-1].strip()
            
            except Exception as decode_error:
                print(f"Error decoding response: {decode_error}")
                # Fallback to simpler decoding
                full_response = self.tokenizer.decode(outputs[0], skip_special_tokens=True)
                assistant_response = full_response.split(prompt)[-1].strip()
            
            return assistant_response
        
        except Exception as e:
            import traceback
            print(f"Error generating team: {e}")
            print(traceback.format_exc())
            # Return a user-friendly error message as a string
            return f"We encountered an error while analyzing the team. The AI service may be temporarily unavailable. You can try again later or use the basic team creation features. Error details: {str(e)}"

    def analyze_team(self, team: List[Dict[str, Any]], include_startup_comparison: bool = True) -> str:
        """
        Analyze an existing team and provide insights.
        Args:
            team: List of team members
            include_startup_comparison: Whether to compare to successful startups
        Returns:
            Team analysis as string
        """
        if self.model is None or self.tokenizer is None:
            self.load_model()
            if self.model is None:
                return "Model could not be loaded. The AI analysis service is currently unavailable."
        
        try:
            # Format team information
            team_summary = self.format_candidates_for_prompt(team)
            
            # Build the prompt
            prompt = f"""Analyze this existing startup team in depth:

TEAM MEMBERS:
{team_summary}

Please provide:

1. Team Composition Analysis (minimum 150 words):
   - Overall assessment of the team's strengths and complementary skills
   - Key skill coverage and potential skill gaps
   - Team dynamics and how members would likely work together
   - Potential areas of conflict or collaboration challenges

2. Success Factors (minimum 100 words):
   - What types of startups would be MOST successful with this team
   - Key advantages this team has compared to typical startup teams
   - How team members' backgrounds create competitive advantages

3. Risk Factors (minimum 100 words):
   - What types of startups would be LEAST successful with this team
   - Potential blind spots or weaknesses in the team composition
   - Suggested additions or changes to strengthen the team
"""

            if include_startup_comparison:
                prompt += """
4. Comparison to Successful Startups (minimum 100 words):
   - How this team compares to founding teams of successful startups
   - Historical examples of similar team compositions that succeeded
   - Key differentiating factors from typical successful startup teams
"""

            # Format as chat with improved system prompt
            messages = [
                {"role": "system", "content": "You are an elite startup advisor with deep expertise in team composition and founder dynamics. You specialize in analyzing team profiles and providing actionable insights to maximize chances of startup success."},
                {"role": "user", "content": prompt}
            ]
            
            # Convert to model input format
            inputs = self.tokenizer.apply_chat_template(
                messages, 
                return_tensors="pt"
            ).to(self.device)
            
            # Generate with proper context limits and attention masks
            with torch.no_grad():
                # Find input length to set appropriate output length
                input_length = inputs.shape[1]
                # Phi-2 has a context limit of 2048
                max_context_length = 2048
                
                # Calculate max new tokens to avoid exceeding model's context limits
                max_new_tokens = max(100, min(1024, max_context_length - input_length))
                
                print(f"Team analysis - Input length: {input_length}, Max new tokens: {max_new_tokens}")
                
                # Create attention mask (explicitly handle padding)
                attention_mask = inputs.ne(self.tokenizer.pad_token_id).long()
                
                # Use the optimized generator instead of direct model.generate call
                outputs, input_length = self.generate_optimized(
                    inputs,
                    attention_mask=attention_mask,
                    max_new_tokens=max_new_tokens
                )
                
            # Decode more carefully
            try:
                # Get only the generated part (exclude input tokens)
                generated_output = outputs[0][input_length:]
                
                # Decode just the new tokens
                generated_text = self.tokenizer.decode(
                    generated_output, 
                    skip_special_tokens=True,
                    clean_up_tokenization_spaces=True
                )
                
                # Remove any model-specific artifacts
                generated_text = generated_text.replace("<|im_end|>", "").replace("<|im_start|>", "")
                assistant_response = generated_text.strip()
                
                # If that failed, try traditional approach
                if not assistant_response:
                    full_response = self.tokenizer.decode(outputs[0], skip_special_tokens=True)
                    assistant_response = full_response.split(prompt)[-1].strip()
            
            except Exception as decode_error:
                print(f"Error decoding team analysis response: {decode_error}")
                # Fallback to simpler decoding
                full_response = self.tokenizer.decode(outputs[0], skip_special_tokens=True)
                assistant_response = full_response.split(prompt)[-1].strip()
            
            return assistant_response
        
        except Exception as e:
            import traceback
            print(f"Error analyzing team: {e}")
            print(traceback.format_exc())
            # Return a user-friendly error message as a string
            return f"We encountered an error while analyzing the team. The AI service may be temporarily unavailable. You can try again later or use the basic team features. Error details: {str(e)}"

    def analyze_candidate(self, candidate: Dict[str, Any]) -> str:
        """
        Analyze an individual candidate and provide insights.
        Args:
            candidate: Candidate information
        Returns:
            Candidate analysis as string
        """
        if self.model is None or self.tokenizer is None:
            self.load_model()
            if self.model is None:
                return "Model could not be loaded. The AI analysis service is currently unavailable."
        
        try:
            # Extract candidate information
            name = candidate.get('name', 'Unnamed Candidate')
            
            education_info = ""
            if candidate.get('education'):
                if isinstance(candidate['education'], list):
                    for edu in candidate['education']:
                        education_info += f"- {edu.get('degree', 'Degree')} from {edu.get('institution', 'Institution')}\n"
                else:
                    edu = candidate['education']
                    education_info += f"- {edu.get('degree', 'Degree')} from {edu.get('institution', 'Institution')}\n"
            
            experience_info = ""
            if candidate.get('experience'):
                if isinstance(candidate['experience'], list):
                    for exp in candidate['experience']:
                        experience_info += f"- {exp.get('title', 'Role')} at {exp.get('company', 'Company')}\n"
                else:
                    exp = candidate['experience']
                    experience_info += f"- {exp.get('title', 'Role')} at {exp.get('company', 'Company')}\n"
            
            skills_info = ""
            if candidate.get('skills'):
                skills = candidate['skills'] if isinstance(candidate['skills'], list) else [candidate['skills']]
                skills_info = ", ".join(skills)
            
            # Build comprehensive prompt
            prompt = f"""Analyze this candidate in depth for a startup founder or early employee role:

CANDIDATE PROFILE:
Name: {name}

Education:
{education_info}

Experience:
{experience_info}

Skills:
{skills_info}

Please provide a comprehensive analysis including:

1. Strengths Analysis (minimum 150 words):
   - Key professional strengths based on background and skills
   - Notable accomplishments and their significance
   - Areas of deep expertise and how they apply to startups

2. Founder/Early Employee Fit (minimum 150 words):
   - Assessment of suitability for founder or early employee roles
   - Specific founder archetype this candidate represents
   - Optimal startup stages for this candidate
   - Ideal role recommendations in a startup team

3. Complementary Team Members (minimum 100 words):
   - What types of co-founders or team members would complement this candidate
   - Potential team dynamics when working with different personality types
   - Skills gaps that should be filled by other team members

4. Risk Assessment (minimum 100 words):
   - Potential blind spots or weaknesses based on background
   - Areas where the candidate might need support or development
   - Situations where this candidate might struggle in a startup environment

Format your analysis with clear sections and detailed insights to help assess this candidate for startup roles.
"""

            # Format as chat with system prompt
            messages = [
                {"role": "system", "content": "You are an elite talent assessor specializing in startup founders and early employees. You provide in-depth analysis of candidates' strengths, founder fit, and team compatibility."},
                {"role": "user", "content": prompt}
            ]
            
            # Convert to model input format
            inputs = self.tokenizer.apply_chat_template(
                messages, 
                return_tensors="pt"
            ).to(self.device)
            
            # Generate with proper context limits and attention masks
            with torch.no_grad():
                # Find input length to set appropriate output length
                input_length = inputs.shape[1]
                # Phi-2 has a context limit of 2048
                max_context_length = 2048
                
                # Calculate max new tokens to avoid exceeding model's context limits
                max_new_tokens = max(100, min(1024, max_context_length - input_length))
                
                print(f"Candidate analysis - Input length: {input_length}, Max new tokens: {max_new_tokens}")
                
                # Create attention mask (explicitly handle padding)
                attention_mask = inputs.ne(self.tokenizer.pad_token_id).long()
                
                # Use the optimized generator instead of direct model.generate call
                outputs, input_length = self.generate_optimized(
                    inputs,
                    attention_mask=attention_mask,
                    max_new_tokens=max_new_tokens
                )
                
            # Decode more carefully
            try:
                # Get only the generated part (exclude input tokens)
                generated_output = outputs[0][input_length:]
                
                # Decode just the new tokens
                generated_text = self.tokenizer.decode(
                    generated_output, 
                    skip_special_tokens=True,
                    clean_up_tokenization_spaces=True
                )
                
                # Remove any model-specific artifacts
                generated_text = generated_text.replace("<|im_end|>", "").replace("<|im_start|>", "")
                assistant_response = generated_text.strip()
                
                # If that failed, try traditional approach
                if not assistant_response:
                    full_response = self.tokenizer.decode(outputs[0], skip_special_tokens=True)
                    assistant_response = full_response.split(prompt)[-1].strip()
            
            except Exception as decode_error:
                print(f"Error decoding candidate analysis response: {decode_error}")
                # Fallback to simpler decoding
                full_response = self.tokenizer.decode(outputs[0], skip_special_tokens=True)
                assistant_response = full_response.split(prompt)[-1].strip()
            
            return assistant_response
        
        except Exception as e:
            import traceback
            print(f"Error analyzing candidate: {e}")
            print(traceback.format_exc())
            # Return a user-friendly error message as a string
            return f"We encountered an error while analyzing this candidate. The AI service may be temporarily unavailable. You can try again later. Error details: {str(e)}"

    def parse_natural_language_requirements(self, text: str) -> Dict[str, Any]:
        """
        Parse natural language team requirements into structured format.
        Args:
            text: Natural language description of team requirements
        Returns:
            Dictionary with structured requirements
        """
        if self.model is None or self.tokenizer is None:
            self.load_model()
            if self.model is None:
                return {"error": "Model could not be loaded"}
        
        try:
            # Build parsing prompt
            prompt = f"""Parse the following team requirements into a structured format:

REQUIREMENTS:
{text}

Extract the following information:
1. Required skills (comma-separated list)
2. Required roles (comma-separated list)
3. Industry focus (single industry)
4. Startup stage (idea, mvp, growth, or scaling)
5. Special requirements (comma-separated list)

Format your response as a JSON object with these keys: requiredSkills, requiredRoles, industryFocus, startupStage, specialRequirements
"""

            # Format as chat
            messages = [
                {"role": "system", "content": "You are a helpful assistant that extracts structured information from text."},
                {"role": "user", "content": prompt}
            ]
            
            # Convert to model input format
            inputs = self.tokenizer.apply_chat_template(
                messages, 
                return_tensors="pt"
            ).to(self.device)
            
            # Generate
            with torch.no_grad():
                outputs = self.model.generate(
                    inputs,
                    max_length=1024,
                    temperature=0.2,  # Lower temperature for more deterministic output
                    top_p=0.9,
                )
                
            # Decode
            response = self.tokenizer.decode(outputs[0], skip_special_tokens=True)
            
            # Extract the JSON part
            assistant_response = response.split(prompt)[-1].strip()
            
            # Find JSON in the response - look for text between curly braces
            import re
            json_match = re.search(r'\{.*\}', assistant_response, re.DOTALL)
            if json_match:
                json_str = json_match.group(0)
                try:
                    return json.loads(json_str)
                except json.JSONDecodeError:
                    print(f"Error parsing JSON: {json_str}")
                    
            # Fallback to a default structure if JSON parsing fails
            return {
                "requiredSkills": [],
                "requiredRoles": [],
                "industryFocus": "",
                "startupStage": "",
                "specialRequirements": []
            }
            
        except Exception as e:
            print(f"Error parsing requirements: {e}")
            return {
                "requiredSkills": [],
                "requiredRoles": [],
                "industryFocus": "",
                "startupStage": "",
                "specialRequirements": []
            }
    
    def search_candidates_by_description(self, description: str, candidates: List[Dict[str, Any]]) -> List[Dict[str, Any]]:
        """
        Search for candidates matching a natural language description.
        Args:
            description: Natural language description of desired candidate
            candidates: List of candidates to search
        Returns:
            Filtered list of candidates matching the description
        """
        if self.model is None or self.tokenizer is None:
            self.load_model()
            if self.model is None:
                return []
        
        try:
            # Build search prompt
            candidate_summary = self.format_candidates_for_prompt(candidates)
            
            prompt = f"""I'm looking for candidates matching this description: "{description}"

Here are the available candidates:
{candidate_summary}

Return a JSON array containing ONLY the candidate numbers (starting from 1) that best match this description. Format it like this: [1, 4, 7]
"""

            # Format as chat
            messages = [
                {"role": "system", "content": "You are a helpful assistant that finds the best matching candidates based on descriptions."},
                {"role": "user", "content": prompt}
            ]
            
            # Convert to model input format
            inputs = self.tokenizer.apply_chat_template(
                messages, 
                return_tensors="pt"
            ).to(self.device)
            
            # Generate
            with torch.no_grad():
                outputs = self.model.generate(
                    inputs,
                    max_length=1024,
                    temperature=0.3,
                    top_p=0.9,
                )
                
            # Decode
            response = self.tokenizer.decode(outputs[0], skip_special_tokens=True)
            
            # Extract the array part
            assistant_response = response.split(prompt)[-1].strip()
            
            # Find array in the response
            import re
            array_match = re.search(r'\[.*\]', assistant_response, re.DOTALL)
            if array_match:
                array_str = array_match.group(0)
                try:
                    indices = json.loads(array_str)
                    # Convert 1-based indices to 0-based
                    indices = [i-1 for i in indices if i > 0 and i <= len(candidates)]
                    return [candidates[i] for i in indices]
                except json.JSONDecodeError:
                    print(f"Error parsing indices: {array_str}")
            
            # Fallback to empty list if parsing fails
            return []
            
        except Exception as e:
            print(f"Error searching candidates: {e}")
            return []

    def __call__(self, data: Dict[str, Any]) -> Dict[str, Any]:
        """
        Perform inference on the input data.
        Args:
            data: Dictionary with inputs key containing the operation parameters
        Returns:
            Dictionary with the generated team analysis
        """
        # Extract inputs field if present (standard Hugging Face format)
        if "inputs" in data:
            inputs_data = data["inputs"]
        else:
            inputs_data = data  # Fallback to backward compatibility
            
        # Check for ping operation first - always respond to status checks
        operation = inputs_data.get("operation", "generate_team")
        
        # Special handling for ping - always respond even if model is not loaded
        if operation == "ping":
            model_status = "loaded" if self.model is not None else "not loaded"
            return {
                "status": "ok",
                "model_status": model_status,
                "model_info": {
                    "device": str(self.device),
                    "model_type": "phi-2-qlora-finetuned"
                }
            }
        
        # Try loading the model again if it failed previously
        if self.model is None:
            print("Model wasn't loaded initially, trying to load it now...")
            self.load_model()
        
        # If the model still couldn't be loaded, return an error that the frontend can handle
        if self.model is None:
            # The frontend expects strings for analysis fields, not nested objects
            if operation == "generate_team":
                return {
                    "team_analysis": "Model could not be loaded. The AI analysis service is currently unavailable. We're working to restore it as soon as possible. In the meantime, you can still use the basic team creation features.",
                    "error": "Model could not be loaded, using fallback mode",
                    "fallback": True,
                    "model_info": {
                        "device": str(self.device),
                        "model_type": "phi-2-qlora-finetuned",
                        "status": "failed"
                    }
                }
            elif operation == "analyze_team":
                return {
                    "team_analysis": "Model could not be loaded. The AI analysis service is currently unavailable. We're working to restore it as soon as possible.",
                    "error": "Model could not be loaded, using fallback mode",
                    "fallback": True,
                    "model_info": {
                        "device": str(self.device),
                        "model_type": "phi-2-qlora-finetuned",
                        "status": "failed"
                    }
                }
            elif operation == "analyze_candidate":
                return {
                    "candidate_analysis": "Model could not be loaded. The AI analysis service is currently unavailable. We're working to restore it as soon as possible.",
                    "error": "Model could not be loaded, using fallback mode",
                    "fallback": True,
                    "model_info": {
                        "device": str(self.device),
                        "model_type": "phi-2-qlora-finetuned",
                        "status": "failed"
                    }
                }
            else:
                # For other operations
                return {
                    "error": "Model could not be loaded, using fallback mode",
                    "fallback": True,
                    "model_info": {
                        "device": str(self.device),
                        "model_type": "phi-2-qlora-finetuned",
                        "status": "failed"
                    }
                }
        
        # Continue with normal operation handling
        if operation == "generate_team":
            candidates = inputs_data.get("candidates", [])
            requirements = inputs_data.get("requirements", "")
            team_size = int(inputs_data.get("team_size", 5))
            
            if not candidates:
                return {"error": "No candidates provided"}
            
            team_analysis = self.generate_team(candidates, requirements, team_size)
            
            return {
                "team_analysis": team_analysis,
                "model_info": {
                    "device": str(self.device),
                    "model_type": "phi-2-qlora-finetuned"
                }
            }
            
        elif operation == "analyze_team":
            team = inputs_data.get("team", [])
            include_startup_comparison = inputs_data.get("include_startup_comparison", True)
            
            if not team:
                return {"error": "No team provided"}
            
            team_analysis = self.analyze_team(team, include_startup_comparison)
            
            return {
                "team_analysis": team_analysis,
                "model_info": {
                    "x": str(self.device),
                    "model_type": "phi-2-qlora-finetuned"
                }
            }
            
        elif operation == "analyze_candidate":
            candidate = inputs_data.get("candidate", {})
            
            if not candidate:
                return {"error": "No candidate provided"}
            
            candidate_analysis = self.analyze_candidate(candidate)
            
            return {
                "candidate_analysis": candidate_analysis,
                "model_info": {
                    "device": str(self.device),
                    "model_type": "phi-2-qlora-finetuned"
                }
            }
            
        elif operation == "parse_requirements":
            requirements_text = inputs_data.get("requirements_text", "")
            parsed_requirements = self.parse_natural_language_requirements(requirements_text)
            
            return {
                "parsed_requirements": parsed_requirements,
                "model_info": {
                    "device": str(self.device),
                    "model_type": "phi-2-qlora-finetuned"
                }
            }
            
        elif operation == "search_candidates":
            candidates = inputs_data.get("candidates", [])
            description = inputs_data.get("description", "")
            
            if not candidates:
                return {"error": "No candidates provided"}
                
            if not description:
                return {"error": "No search description provided"}
                
            matching_candidates = self.search_candidates_by_description(description, candidates)
            
            return {
                "matching_candidates": matching_candidates,
                "count": len(matching_candidates),
                "model_info": {
                    "device": str(self.device),
                    "model_type": "phi-2-qlora-finetuned"
                }
            }
            
        # Ping operation is now handled at the beginning for better reliability
            
        else:
            return {"error": f"Unknown operation: {operation}"}