app.py

import os
import logging
import numpy as np
import torch
from torch import nn
import pandas as pd
from torchvision import transforms, models
from PIL import Image
import faiss
from transformers import AutoTokenizer, AutoModel, T5ForConditionalGeneration, T5Tokenizer
import gradio as gr
import cv2
import traceback
from datetime import datetime
import re
import random
import functools
import gc
from collections import OrderedDict
import json
import sys
import time
from tqdm.auto import tqdm
import warnings
import matplotlib.pyplot as plt
from fastapi import FastAPI, File, UploadFile, Form
from fastapi.middleware.cors import CORSMiddleware
from pydantic import BaseModel
from typing import Optional, List, Dict, Any, Union
import base64
import io

# Suppress unnecessary warnings
warnings.filterwarnings("ignore", category=UserWarning)

# === Configuration ===
class Config:
    """Configuration for MediQuery system"""
    # Model configuration
    IMAGE_MODEL = "chexnet"  # Options: "chexnet", "densenet"
    TEXT_MODEL = "biobert"   # Options: "biobert", "clinicalbert"
    GEN_MODEL = "flan-t5-base-finetuned"  # Base generation model

    # Resource management
    CACHE_SIZE = 50  # Reduced from 200 for deployment
    CACHE_EXPIRY_TIME = 1800  # Cache expiry time in seconds (30 minutes)
    LAZY_LOADING = True       # Enable lazy loading of models
    USE_HALF_PRECISION = True # Use half precision for models if available

    # Feature flags
    DEBUG = True              # Enable detailed debugging
    PHI_DETECTION_ENABLED = True  # Enable PHI detection
    ANATOMY_MAPPING_ENABLED = True  # Enable anatomical mapping

    # Thresholds and parameters
    CONFIDENCE_THRESHOLD = 0.4  # Threshold for flagging low confidence
    TOP_K_RETRIEVAL = 10        # Reduced from 30 for deployment
    MAX_CONTEXT_DOCS = 3        # Reduced from 5 for deployment

    # Advanced retrieval settings
    DYNAMIC_RERANKING = True    # Dynamically adjust reranking weights
    DIVERSITY_PENALTY = 0.1     # Penalty for duplicate content

    # Performance optimization
    BATCH_SIZE = 1              # Reduced from 4 for deployment
    OPTIMIZE_MEMORY = True      # Optimize memory usage
    USE_CACHING = True          # Use caching for embeddings and queries

    # Path settings
    DEFAULT_KNOWLEDGE_BASE_DIR = "./knowledge_base"
    DEFAULT_MODEL_PATH = "./models/flan-t5-finetuned"
    LOG_DIR = "./logs"

    # Advanced settings
    EMBEDDING_AGGREGATION = "weighted_avg"  # Options: "avg", "weighted_avg", "cls", "pooled"
    EMBEDDING_NORMALIZE = True  # Normalize embeddings to unit length

    # Error recovery settings
    MAX_RETRIES = 2  # Reduced from 3 for deployment
    RECOVERY_WAIT_TIME = 1  # Seconds to wait between retries

# Set up logging with improved formatting
os.makedirs(Config.LOG_DIR, exist_ok=True)
logging.basicConfig(
    level=logging.DEBUG if Config.DEBUG else logging.INFO,
    format='%(asctime)s - %(name)s - %(levelname)s - %(message)s',
    handlers=[
        logging.FileHandler(os.path.join(Config.LOG_DIR, f"mediquery_{datetime.now().strftime('%Y%m%d_%H%M%S')}.log")),
        logging.StreamHandler()
    ]
)
logger = logging.getLogger("MediQuery")

def debug_print(msg):
    """Print and log debug messages"""
    if Config.DEBUG:
        logger.debug(msg)
        print(f"DEBUG: {msg}")

# === Helper Functions for Conditions ===
def get_mimic_cxr_conditions():
    """Return the comprehensive list of conditions in MIMIC-CXR dataset"""
    return [
        "atelectasis",
        "cardiomegaly",
        "consolidation",
        "edema",
        "enlarged cardiomediastinum",
        "fracture",
        "lung lesion",
        "lung opacity",
        "no finding",
        "pleural effusion",
        "pleural other",
        "pneumonia",
        "pneumothorax",
        "support devices"
    ]

def get_condition_synonyms():
    """Return synonyms for conditions to improve matching"""
    return {
        "atelectasis": ["atelectatic change", "collapsed lung", "lung collapse"],
        "cardiomegaly": ["enlarged heart", "cardiac enlargement", "heart enlargement"],
        "consolidation": ["airspace opacity", "air-space opacity", "alveolar opacity"],
        "edema": ["pulmonary edema", "fluid overload", "vascular congestion"],
        "fracture": ["broken bone", "bone fracture", "rib fracture"],
        "lung opacity": ["pulmonary opacity", "opacification", "lung opacification"],
        "pleural effusion": ["pleural fluid", "fluid in pleural space", "effusion"],
        "pneumonia": ["pulmonary infection", "lung infection", "bronchopneumonia"],
        "pneumothorax": ["air in pleural space", "collapsed lung", "ptx"],
        "support devices": ["tube", "line", "catheter", "pacemaker", "device"]
    }

def get_anatomical_regions():
    """Return mapping of anatomical regions with descriptions and conditions"""
    return {
        "upper_right_lung": {
            "description": "Upper right lung field",
            "conditions": ["pneumonia", "lung lesion", "pneumothorax", "atelectasis"]
        },
        "upper_left_lung": {
            "description": "Upper left lung field",
            "conditions": ["pneumonia", "lung lesion", "pneumothorax", "atelectasis"]
        },
        "middle_right_lung": {
            "description": "Middle right lung field",
            "conditions": ["pneumonia", "lung opacity", "atelectasis"]
        },
        "lower_right_lung": {
            "description": "Lower right lung field",
            "conditions": ["pneumonia", "pleural effusion", "atelectasis"]
        },
        "lower_left_lung": {
            "description": "Lower left lung field",
            "conditions": ["pneumonia", "pleural effusion", "atelectasis"]
        },
        "heart": {
            "description": "Cardiac silhouette",
            "conditions": ["cardiomegaly", "enlarged cardiomediastinum"]
        },
        "hilar": {
            "description": "Hilar regions",
            "conditions": ["enlarged cardiomediastinum", "adenopathy"]
        },
        "costophrenic_angles": {
            "description": "Costophrenic angles",
            "conditions": ["pleural effusion", "pneumothorax"]
        },
        "spine": {
            "description": "Spine",
            "conditions": ["fracture", "degenerative changes"]
        },
        "diaphragm": {
            "description": "Diaphragm",
            "conditions": ["elevated diaphragm", "flattened diaphragm"]
        }
    }

# === PHI Detection and Anonymization ===
def detect_phi(text):
    """Detect potential PHI (Protected Health Information) in text"""
    # Patterns for PHI detection
    patterns = {
        'name': r'\b[A-Z][a-z]+ [A-Z][a-z]+\b',
        'mrn': r'\b[A-Z]{0,3}[0-9]{4,10}\b',
        'ssn': r'\b[0-9]{3}[-]?[0-9]{2}[-]?[0-9]{4}\b',
        'date': r'\b(0?[1-9]|1[0-2])[\/\-](0?[1-9]|[12]\d|3[01])[\/\-](19|20)\d{2}\b',
        'phone': r'\b(\+\d{1,2}\s?)?\(?\d{3}\)?[\s.-]?\d{3}[\s.-]?\d{4}\b',
        'email': r'\b[A-Za-z0-9._%+-]+@[A-Za-z0-9.-]+\.[A-Z|a-z]{2,}\b',
        'address': r'\b\d+\s+[A-Z][a-z]+\s+[A-Z][a-z]+\.?\b'
    }

    # Check each pattern
    phi_detected = {}
    for phi_type, pattern in patterns.items():
        matches = re.findall(pattern, text)
        if matches:
            phi_detected[phi_type] = matches

    return phi_detected

def anonymize_text(text):
    """Replace potential PHI with [REDACTED]"""
    if not text:
        return ""

    if not Config.PHI_DETECTION_ENABLED:
        return text

    try:
        # Detect PHI
        phi_detected = detect_phi(text)

        # Replace PHI with [REDACTED]
        anonymized = text
        for phi_type, matches in phi_detected.items():
            for match in matches:
                anonymized = anonymized.replace(match, "[REDACTED]")

        return anonymized
    except Exception as e:
        debug_print(f"Error in anonymize_text: {str(e)}")
        return text

# === LRU Cache Implementation with Enhanced Features ===
class LRUCache:
    """LRU (Least Recently Used) Cache implementation with TTL and size tracking"""
    def __init__(self, capacity=Config.CACHE_SIZE, expiry_time=Config.CACHE_EXPIRY_TIME):
        self.cache = OrderedDict()
        self.capacity = capacity
        self.expiry_time = expiry_time  # in seconds
        self.timestamps = {}
        self.size_tracking = {
            "current_size_bytes": 0,
            "max_size_bytes": 0,
            "items_evicted": 0,
            "cache_hits": 0,
            "cache_misses": 0
        }

    def get(self, key):
        """Get item from cache with statistics tracking"""
        if key not in self.cache:
            self.size_tracking["cache_misses"] += 1
            return None

        # Check expiry
        if self.is_expired(key):
            self._remove_with_tracking(key)
            self.size_tracking["cache_misses"] += 1
            return None

        # Move to end (recently used)
        self.size_tracking["cache_hits"] += 1
        value = self.cache.pop(key)
        self.cache[key] = value
        return value

    def put(self, key, value):
        """Add item to cache with size tracking"""
        # Calculate approximate size of the value
        value_size = self._estimate_size(value)

        if key in self.cache:
            old_value = self.cache.pop(key)
            old_size = self._estimate_size(old_value)
            self.size_tracking["current_size_bytes"] -= old_size

        # Make space if needed
        while len(self.cache) >= self.capacity or (
            Config.OPTIMIZE_MEMORY and
            self.size_tracking["current_size_bytes"] + value_size > 1e9  # 1 GB limit
        ):
            self._evict_least_recently_used()

        # Add new item and timestamp
        self.cache[key] = value
        self.timestamps[key] = datetime.now().timestamp()
        self.size_tracking["current_size_bytes"] += value_size

        # Update max size
        if self.size_tracking["current_size_bytes"] > self.size_tracking["max_size_bytes"]:
            self.size_tracking["max_size_bytes"] = self.size_tracking["current_size_bytes"]

    def is_expired(self, key):
        """Check if item has expired"""
        if key not in self.timestamps:
            return True

        current_time = datetime.now().timestamp()
        return (current_time - self.timestamps[key]) > self.expiry_time

    def _evict_least_recently_used(self):
        """Remove least recently used item with tracking"""
        if not self.cache:
            return

        # Get oldest item
        key, value = self.cache.popitem(last=False)
        # Remove from timestamps and update tracking
        self._remove_with_tracking(key)

    def _remove_with_tracking(self, key):
        """Remove item with size tracking"""
        if key in self.cache:
            value = self.cache.pop(key)
            value_size = self._estimate_size(value)
            self.size_tracking["current_size_bytes"] -= value_size
            self.size_tracking["items_evicted"] += 1

        if key in self.timestamps:
            self.timestamps.pop(key)

    def remove(self, key):
        """Remove item from cache"""
        self._remove_with_tracking(key)

    def clear(self):
        """Clear the cache"""
        self.cache.clear()
        self.timestamps.clear()
        self.size_tracking["current_size_bytes"] = 0

    def get_stats(self):
        """Get cache statistics"""
        return {
            "size_bytes": self.size_tracking["current_size_bytes"],
            "max_size_bytes": self.size_tracking["max_size_bytes"],
            "items": len(self.cache),
            "capacity": self.capacity,
            "items_evicted": self.size_tracking["items_evicted"],
            "hit_rate": self.size_tracking["cache_hits"] /
                        (self.size_tracking["cache_hits"] + self.size_tracking["cache_misses"] + 1e-8)
        }

    def _estimate_size(self, obj):
        """Estimate memory size of an object in bytes"""
        if obj is None:
            return 0

        if isinstance(obj, np.ndarray):
            return obj.nbytes
        elif isinstance(obj, torch.Tensor):
            return obj.element_size() * obj.nelement()
        elif isinstance(obj, (str, bytes)):
            return len(obj)
        elif isinstance(obj, (list, tuple)):
            return sum(self._estimate_size(x) for x in obj)
        elif isinstance(obj, dict):
            return sum(self._estimate_size(k) + self._estimate_size(v) for k, v in obj.items())
        else:
            # Fallback - rough estimate
            return sys.getsizeof(obj)

# === Improved Lazy Model Loading ===
class LazyModel:
    """Lazy loading wrapper for models with proper method forwarding and error recovery"""
    def __init__(self, model_name, model_class, device, **kwargs):
        self.model_name = model_name
        self.model_class = model_class
        self.device = device
        self.kwargs = kwargs
        self._model = None
        self.last_error = None
        self.last_used = datetime.now()
        debug_print(f"LazyModel initialized for {model_name}")

    def _ensure_loaded(self, retries=Config.MAX_RETRIES):
        """Ensure model is loaded with retry mechanism"""
        if self._model is None:
            debug_print(f"Lazy loading model: {self.model_name}")
            for attempt in range(retries):
                try:
                    self._model = self.model_class.from_pretrained(self.model_name, **self.kwargs)

                    # Apply memory optimizations
                    if Config.OPTIMIZE_MEMORY:
                        # Convert to half precision if available and enabled
                        if Config.USE_HALF_PRECISION and self.device.type == 'cuda' and hasattr(self._model, 'half'):
                            self._model = self._model.half()
                            debug_print(f"Using half precision for {self.model_name}")

                    self._model = self._model.to(self.device)
                    self._model.eval()  # Set to evaluation mode
                    debug_print(f"Model {self.model_name} loaded successfully")
                    self.last_error = None
                    break
                except Exception as e:
                    self.last_error = str(e)
                    debug_print(f"Error loading model {self.model_name} (attempt {attempt+1}/{retries}): {str(e)}")
                    if attempt < retries - 1:
                        # Wait before retrying
                        time.sleep(Config.RECOVERY_WAIT_TIME)
                    else:
                        raise RuntimeError(f"Failed to load model {self.model_name} after {retries} attempts: {str(e)}")

        # Update last used timestamp
        self.last_used = datetime.now()
        return self._model

    def __call__(self, *args, **kwargs):
        """Call the model"""
        model = self._ensure_loaded()
        return model(*args, **kwargs)

    # Forward common model methods
    def generate(self, *args, **kwargs):
        """Forward generate method to model with error recovery"""
        model = self._ensure_loaded()
        try:
            return model.generate(*args, **kwargs)
        except Exception as e:
            # If generation fails, try reloading the model once
            debug_print(f"Generation failed, reloading model: {str(e)}")
            self.unload()
            model = self._ensure_loaded()
            return model.generate(*args, **kwargs)

    def to(self, device):
        """Move model to specified device"""
        self.device = device
        if self._model is not None:
            self._model = self._model.to(device)
        return self

    def eval(self):
        """Set model to evaluation mode"""
        if self._model is not None:
            self._model.eval()
        return self

    def unload(self):
        """Unload model from memory"""
        if self._model is not None:
            del self._model
            self._model = None
            gc.collect()
            if torch.cuda.is_available():
                torch.cuda.empty_cache()
            debug_print(f"Model {self.model_name} unloaded")

# === MediQuery Core System ===
class MediQuery:
    """Core MediQuery system for medical image and text analysis"""
    def __init__(self, knowledge_base_dir=Config.DEFAULT_KNOWLEDGE_BASE_DIR, model_path=Config.DEFAULT_MODEL_PATH):
        self.knowledge_base_dir = knowledge_base_dir
        self.model_path = model_path
        self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
        debug_print(f"Using device: {self.device}")
        
        # Create directories if they don't exist
        os.makedirs(knowledge_base_dir, exist_ok=True)
        os.makedirs(os.path.dirname(model_path), exist_ok=True)
        
        # Initialize caches
        self.embedding_cache = LRUCache(capacity=Config.CACHE_SIZE)
        self.query_cache = LRUCache(capacity=Config.CACHE_SIZE)
        
        # Initialize models
        self._init_models()
        
        # Load knowledge base
        self._init_knowledge_base()
        
        debug_print("MediQuery system initialized")

    def _init_models(self):
        """Initialize all required models with lazy loading"""
        debug_print("Initializing models...")
        
        # Image model
        if Config.IMAGE_MODEL == "chexnet":
            self.image_model = models.densenet121(pretrained=False)
            # For deployment, we'll download the weights during initialization
            try:
                # Simplified for deployment - would need to download weights
                self.image_model = nn.Sequential(*list(self.image_model.children())[:-1])
                debug_print("CheXNet model initialized")
            except Exception as e:
                debug_print(f"Error initializing CheXNet: {str(e)}")
                # Fallback to standard DenseNet
                self.image_model = nn.Sequential(*list(models.densenet121(pretrained=True).children())[:-1])
        else:
            self.image_model = nn.Sequential(*list(models.densenet121(pretrained=True).children())[:-1])
        
        self.image_model = self.image_model.to(self.device).eval()
        
        # Text model - lazy loaded
        text_model_name = "dmis-lab/biobert-v1.1" if Config.TEXT_MODEL == "biobert" else "emilyalsentzer/Bio_ClinicalBERT"
        self.text_tokenizer = AutoTokenizer.from_pretrained(text_model_name)
        self.text_model = LazyModel(
            text_model_name,
            AutoModel,
            self.device
        )
        
        # Generation model - lazy loaded
        if os.path.exists(self.model_path):
            gen_model_path = self.model_path
        else:
            gen_model_path = "google/flan-t5-base"  # Fallback to base model
        
        self.gen_tokenizer = T5Tokenizer.from_pretrained(gen_model_path)
        self.gen_model = LazyModel(
            gen_model_path,
            T5ForConditionalGeneration,
            self.device
        )
        
        # Image transformation
        self.image_transform = transforms.Compose([
            transforms.Resize((224, 224)),
            transforms.ToTensor(),
            transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
        ])
        
        debug_print("Models initialized")

    def _init_knowledge_base(self):
        """Initialize knowledge base with FAISS indices"""
        debug_print("Initializing knowledge base...")
        
        # For deployment, we'll create a minimal knowledge base
        # In a real deployment, you would download the knowledge base files
        
        # Create dummy knowledge base for demonstration
        self.text_data = pd.DataFrame({
            'combined_text': [
                "The chest X-ray shows clear lung fields with no evidence of consolidation, effusion, or pneumothorax. The heart size is normal. No acute cardiopulmonary abnormality.",
                "Bilateral patchy airspace opacities consistent with multifocal pneumonia. No pleural effusion or pneumothorax. Heart size is normal.",
                "Cardiomegaly with pulmonary vascular congestion and bilateral pleural effusions, consistent with congestive heart failure. No pneumothorax or pneumonia.",
                "Right upper lobe opacity concerning for pneumonia. No pleural effusion or pneumothorax. Heart size is normal.",
                "Left lower lobe atelectasis. No pneumothorax or pleural effusion. Heart size is normal.",
                "Bilateral pleural effusions with bibasilar atelectasis. Cardiomegaly present. Findings consistent with heart failure.",
                "Right pneumothorax with partial lung collapse. No pleural effusion. Heart size is normal.",
                "Endotracheal tube, central venous catheter, and nasogastric tube in place. No pneumothorax or pleural effusion.",
                "Hyperinflated lungs with flattened diaphragms, consistent with COPD. No acute infiltrate or effusion.",
                "Multiple rib fractures on the right side. No pneumothorax or hemothorax. Lung fields are clear."
            ],
            'valid_index': list(range(10))
        })
        
        # Create dummy FAISS indices
        self.image_index = None  # Will be created on first use
        self.text_index = None   # Will be created on first use
        
        debug_print("Knowledge base initialized")

    def _create_dummy_indices(self):
        """Create dummy FAISS indices for demonstration"""
        # Text embeddings (768 dimensions for BERT-based models)
        text_dim = 768
        text_embeddings = np.random.rand(len(self.text_data), text_dim).astype('float32')
        
        # Image embeddings (1024 dimensions for DenseNet121)
        image_dim = 1024
        image_embeddings = np.random.rand(len(self.text_data), image_dim).astype('float32')
        
        # Create FAISS indices
        self.text_index = faiss.IndexFlatL2(text_dim)
        self.text_index.add(text_embeddings)
        
        self.image_index = faiss.IndexFlatL2(image_dim)
        self.image_index.add(image_embeddings)
        
        debug_print("Dummy FAISS indices created")

    def process_image(self, image_path):
        """Process an X-ray image and return analysis results"""
        try:
            debug_print(f"Processing image: {image_path}")
            
            # Check cache
            if Config.USE_CACHING:
                cached_result = self.query_cache.get(f"img_{image_path}")
                if cached_result:
                    debug_print("Using cached image result")
                    return cached_result
            
            # Load and preprocess image
            image = Image.open(image_path).convert('RGB')
            image_tensor = self.image_transform(image).unsqueeze(0).to(self.device)
            
            # Generate image embedding
            with torch.no_grad():
                image_embedding = self.image_model(image_tensor)
                image_embedding = nn.functional.avg_pool2d(image_embedding, kernel_size=7).squeeze().cpu().numpy()
            
            # Initialize FAISS indices if needed
            if self.image_index is None:
                self._create_dummy_indices()
            
            # Retrieve similar cases
            distances, indices = self.image_index.search(np.array([image_embedding]), k=Config.TOP_K_RETRIEVAL)
            
            # Get relevant text data
            retrieved_texts = [self.text_data.iloc[idx]['combined_text'] for idx in indices[0]]
            
            # Generate context for the model
            context = "\n\n".join(retrieved_texts[:Config.MAX_CONTEXT_DOCS])
            
            # Generate analysis
            prompt = f"Analyze this chest X-ray based on similar cases:\n\n{context}\n\nProvide a detailed radiological assessment including findings and impression:"
            
            analysis = self._generate_text(prompt)
            
            # Generate attention map (simplified for deployment)
            attention_map = self._generate_attention_map(image)
            
            # Prepare result
            result = {
                "analysis": analysis,
                "attention_map": attention_map,
                "confidence": 0.85,  # Placeholder
                "similar_cases": retrieved_texts[:3]  # Return top 3 similar cases
            }
            
            # Cache result
            if Config.USE_CACHING:
                self.query_cache.put(f"img_{image_path}", result)
            
            return result
            
        except Exception as e:
            error_msg = f"Error processing image: {str(e)}\n{traceback.format_exc()}"
            debug_print(error_msg)
            return {"error": error_msg}

    def process_query(self, query_text):
        """Process a text query and return relevant information"""
        try:
            debug_print(f"Processing query: {query_text}")
            
            # Check cache
            if Config.USE_CACHING:
                cached_result = self.query_cache.get(f"txt_{query_text}")
                if cached_result:
                    debug_print("Using cached query result")
                    return cached_result
            
            # Anonymize query
            query_text = anonymize_text(query_text)
            
            # Generate text embedding
            query_embedding = self._generate_text_embedding(query_text)
            
            # Initialize FAISS indices if needed
            if self.text_index is None:
                self._create_dummy_indices()
            
            # Retrieve similar texts
            distances, indices = self.text_index.search(np.array([query_embedding]), k=Config.TOP_K_RETRIEVAL)
            
            # Get relevant text data
            retrieved_texts = [self.text_data.iloc[idx]['combined_text'] for idx in indices[0]]
            
            # Generate context for the model
            context = "\n\n".join(retrieved_texts[:Config.MAX_CONTEXT_DOCS])
            
            # Generate response
            prompt = f"Answer this medical question based on the following information:\n\nQuestion: {query_text}\n\nRelevant information:\n{context}\n\nDetailed answer:"
            
            response = self._generate_text(prompt)
            
            # Prepare result
            result = {
                "response": response,
                "confidence": 0.9,  # Placeholder
                "sources": retrieved_texts[:3]  # Return top 3 sources
            }
            
            # Cache result
            if Config.USE_CACHING:
                self.query_cache.put(f"txt_{query_text}", result)
            
            return result
            
        except Exception as e:
            error_msg = f"Error processing query: {str(e)}\n{traceback.format_exc()}"
            debug_print(error_msg)
            return {"error": error_msg}

    def _generate_text_embedding(self, text):
        """Generate embedding for text using the text model"""
        try:
            # Check cache
            if Config.USE_CACHING:
                cached_embedding = self.embedding_cache.get(f"txt_emb_{text}")
                if cached_embedding is not None:
                    return cached_embedding
            
            # Tokenize
            inputs = self.text_tokenizer(
                text,
                padding=True,
                truncation=True,
                return_tensors="pt",
                max_length=512
            ).to(self.device)
            
            # Generate embedding
            with torch.no_grad():
                outputs = self.text_model(**inputs)
                
            # Use mean pooling
            embedding = outputs.last_hidden_state.mean(dim=1).cpu().numpy()[0]
            
            # Cache embedding
            if Config.USE_CACHING:
                self.embedding_cache.put(f"txt_emb_{text}", embedding)
            
            return embedding
            
        except Exception as e:
            debug_print(f"Error generating text embedding: {str(e)}")
            # Return random embedding as fallback
            return np.random.rand(768).astype('float32')

    def _generate_text(self, prompt):
        """Generate text using the language model"""
        try:
            # Tokenize
            inputs = self.gen_tokenizer(
                prompt,
                padding=True,
                truncation=True,
                return_tensors="pt",
                max_length=512
            ).to(self.device)
            
            # Generate
            with torch.no_grad():
                output_ids = self.gen_model.generate(
                    inputs.input_ids,
                    max_length=256,
                    num_beams=4,
                    early_stopping=True
                )
            
            # Decode
            output_text = self.gen_tokenizer.decode(output_ids[0], skip_special_tokens=True)
            
            return output_text
            
        except Exception as e:
            debug_print(f"Error generating text: {str(e)}")
            return "I apologize, but I'm unable to generate a response at this time. Please try again later."

    def _generate_attention_map(self, image):
        """Generate a simplified attention map for the image"""
        try:
            # Convert to numpy array
            img_np = np.array(image.resize((224, 224)))
            
            # Create a simple heatmap (this is a placeholder - real implementation would use model attention)
            heatmap = np.zeros((224, 224), dtype=np.float32)
            
            # Add some random "attention" areas
            for _ in range(3):
                x, y = np.random.randint(50, 174, 2)
                radius = np.random.randint(20, 50)
                for i in range(224):
                    for j in range(224):
                        dist = np.sqrt((i - x)**2 + (j - y)**2)
                        if dist < radius:
                            heatmap[i, j] += max(0, 1 - dist/radius)
            
            # Normalize
            heatmap = heatmap / heatmap.max()
            
            # Apply colormap
            heatmap_colored = cv2.applyColorMap((heatmap * 255).astype(np.uint8), cv2.COLORMAP_JET)
            
            # Overlay on original image
            img_rgb = cv2.cvtColor(img_np, cv2.COLOR_RGB2BGR)
            overlay = cv2.addWeighted(img_rgb, 0.7, heatmap_colored, 0.3, 0)
            
            # Convert to base64 for API response
            _, buffer = cv2.imencode('.png', overlay)
            img_str = base64.b64encode(buffer).decode('utf-8')
            
            return img_str
            
        except Exception as e:
            debug_print(f"Error generating attention map: {str(e)}")
            return None

    def cleanup(self):
        """Clean up resources"""
        debug_print("Cleaning up resources...")
        
        # Unload models
        if hasattr(self, 'text_model') and isinstance(self.text_model, LazyModel):
            self.text_model.unload()
        
        if hasattr(self, 'gen_model') and isinstance(self.gen_model, LazyModel):
            self.gen_model.unload()
        
        # Clear caches
        if hasattr(self, 'embedding_cache'):
            self.embedding_cache.clear()
        
        if hasattr(self, 'query_cache'):
            self.query_cache.clear()
        
        # Force garbage collection
        gc.collect()
        if torch.cuda.is_available():
            torch.cuda.empty_cache()
        
        debug_print("Cleanup complete")

# === FastAPI Application ===
app = FastAPI(title="MediQuery API", description="API for MediQuery AI medical assistant")

# Add CORS middleware
app.add_middleware(
    CORSMiddleware,
    allow_origins=["*"],  # For production, specify the actual frontend domain
    allow_credentials=True,
    allow_methods=["*"],
    allow_headers=["*"],
)

# Initialize MediQuery system
mediquery = MediQuery()

# Define API models
class QueryRequest(BaseModel):
    text: str

class QueryResponse(BaseModel):
    response: str
    confidence: float
    sources: List[str]
    error: Optional[str] = None

class ImageAnalysisResponse(BaseModel):
    analysis: str
    attention_map: Optional[str] = None
    confidence: float
    similar_cases: List[str]
    error: Optional[str] = None

@app.post("/api/query", response_model=QueryResponse)
async def process_text_query(query: QueryRequest):
    """Process a text query and return relevant information"""
    result = mediquery.process_query(query.text)
    return result

@app.post("/api/analyze-image", response_model=ImageAnalysisResponse)
async def analyze_image(file: UploadFile = File(...)):
    """Analyze an X-ray image and return results"""
    # Save uploaded file temporarily
    temp_file = f"/tmp/{file.filename}"
    with open(temp_file, "wb") as f:
        f.write(await file.read())
    
    # Process image
    result = mediquery.process_image(temp_file)
    
    # Clean up
    os.remove(temp_file)
    
    return result

@app.get("/api/health")
async def health_check():
    """Health check endpoint"""
    return {"status": "ok", "version": "1.0.0"}

# === Gradio Interface ===
def create_gradio_interface():
    """Create a Gradio interface for the MediQuery system"""
    # Define processing functions
    def process_image_gradio(image):
        # Save image temporarily
        temp_file = "/tmp/gradio_image.png"
        image.save(temp_file)
        
        # Process image
        result = mediquery.process_image(temp_file)
        
        # Clean up
        os.remove(temp_file)
        
        # Prepare output
        analysis = result.get("analysis", "Error processing image")
        attention_map_b64 = result.get("attention_map")
        
        # Convert base64 to image if available
        attention_map = None
        if attention_map_b64:
            try:
                attention_map = Image.open(io.BytesIO(base64.b64decode(attention_map_b64)))
            except:
                pass
        
        return analysis, attention_map
    
    def process_query_gradio(query):
        result = mediquery.process_query(query)
        return result.get("response", "Error processing query")
    
    # Create interface
    with gr.Blocks(title="MediQuery") as demo:
        gr.Markdown("# MediQuery - AI Medical Assistant")
        
        with gr.Tab("Image Analysis"):
            with gr.Row():
                with gr.Column():
                    image_input = gr.Image(type="pil", label="Upload Chest X-ray")
                    image_button = gr.Button("Analyze X-ray")
                
                with gr.Column():
                    text_output = gr.Textbox(label="Analysis Results", lines=10)
                    image_output = gr.Image(label="Attention Map")
            
            image_button.click(
                fn=process_image_gradio,
                inputs=[image_input],
                outputs=[text_output, image_output]
            )
        
        with gr.Tab("Text Query"):
            query_input = gr.Textbox(label="Medical Query", lines=3, placeholder="e.g., What does pneumonia look like on a chest X-ray?")
            query_button = gr.Button("Submit Query")
            query_output = gr.Textbox(label="Response", lines=10)
            
            query_button.click(
                fn=process_query_gradio,
                inputs=[query_input],
                outputs=[query_output]
            )
        
        gr.Markdown("## Example Queries")
        gr.Examples(
            examples=[
                ["What does pleural effusion look like?"],
                ["How to differentiate pneumonia from tuberculosis?"],
                ["What are the signs of cardiomegaly on X-ray?"]
            ],
            inputs=[query_input]
        )
    
    return demo

# Create Gradio interface
demo = create_gradio_interface()

# Mount Gradio app to FastAPI
app = gr.mount_gradio_app(app, demo, path="/")

# Startup and shutdown events
@app.on_event("startup")
async def startup_event():
    """Initialize resources on startup"""
    debug_print("API starting up...")

@app.on_event("shutdown")
async def shutdown_event():
    """Clean up resources on shutdown"""
    debug_print("API shutting down...")
    mediquery.cleanup()

# Run the FastAPI app with uvicorn when executed directly
if __name__ == "__main__":
    import uvicorn
    uvicorn.run(app, host="0.0.0.0", port=8000)