api_mental_health.py

# api_mental_health.py
from fastapi import FastAPI, HTTPException, UploadFile, File, Form
from pydantic import BaseModel
import pandas as pd
import numpy as np
import joblib
import re
import nltk
from nltk.tokenize import word_tokenize
from nltk.stem import WordNetLemmatizer
from nltk.corpus import stopwords
from vaderSentiment.vaderSentiment import SentimentIntensityAnalyzer
import chromadb
from chromadb.config import Settings
from langchain_openai import OpenAIEmbeddings, ChatOpenAI
from langchain_chroma import Chroma
from openai import OpenAI
import os
from dotenv import load_dotenv
from langsmith import Client, traceable
from langchain_core.runnables import RunnablePassthrough
from langchain_core.prompts import ChatPromptTemplate
import logging
from typing import List, Dict, Optional, Any, Union, Annotated
from datetime import datetime
from uuid import uuid4, UUID
import json
import requests
from fastapi.responses import StreamingResponse
from io import BytesIO
import base64

# Set up logging
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)

# Load environment variables
load_dotenv()

# Set NLTK data path to a directory where we want to look for data
nltk_data_path = os.path.join(os.path.dirname(__file__), "nltk_data")
os.makedirs(nltk_data_path, exist_ok=True)
nltk.data.path.append(nltk_data_path)

# Skip downloading NLTK data to avoid permission errors
logger.info(f"Using NLTK data from {nltk_data_path} if available")

# Initialize FastAPI app
app = FastAPI(title="Mental Health Counselor API")

# Initialize global storage (to be replaced with proper database)
DATA_DIR = os.path.join(os.path.dirname(__file__), "data")
os.makedirs(DATA_DIR, exist_ok=True)
os.makedirs(os.path.join(DATA_DIR, "users"), exist_ok=True)
os.makedirs(os.path.join(DATA_DIR, "sessions"), exist_ok=True)
os.makedirs(os.path.join(DATA_DIR, "conversations"), exist_ok=True)
os.makedirs(os.path.join(DATA_DIR, "feedback"), exist_ok=True)

# Initialize components
STOPWORDS = set(stopwords.words("english"))
lemmatizer = WordNetLemmatizer()
analyzer = SentimentIntensityAnalyzer()
output_dir = "mental_health_model_artifacts"

# Global variables for models and vector store
response_clf = None
crisis_clf = None
vectorizer = None
le = None
selector = None
lda = None
vector_store = None
llm = None
openai_client = None
langsmith_client = None

# Load models and initialize ChromaDB at startup
@app.on_event("startup")
async def startup_event():
    global response_clf, crisis_clf, vectorizer, le, selector, lda, vector_store, llm, openai_client, langsmith_client
    
    # Check environment variables
    if not os.environ.get("OPENAI_API_KEY"):
        logger.warning("OPENAI_API_KEY not set in .env file. Some functionality will be limited.")
    if not os.environ.get("LANGCHAIN_API_KEY"):
        logger.warning("LANGCHAIN_API_KEY not set in .env file. Some functionality will be limited.")
    os.environ["LANGCHAIN_TRACING_V2"] = "true"
    os.environ["LANGCHAIN_PROJECT"] = "MentalHealthCounselorPOC"

    # Initialize LangSmith client if API key is available
    try:
        logger.info("Initializing LangSmith client")
        langsmith_client = Client()
    except Exception as e:
        logger.warning(f"Failed to initialize LangSmith client: {e}")
        langsmith_client = None

    # Try to load saved components, continue with limited functionality if not available
    logger.info("Loading model artifacts")
    models_available = True
    try:
        response_clf = joblib.load(f"{output_dir}/response_type_classifier.pkl")
        crisis_clf = joblib.load(f"{output_dir}/crisis_classifier.pkl")
        vectorizer = joblib.load(f"{output_dir}/tfidf_vectorizer.pkl")
        le = joblib.load(f"{output_dir}/label_encoder.pkl")
        selector = joblib.load(f"{output_dir}/feature_selector.pkl")
        
        try:
            lda = joblib.load(f"{output_dir}/lda_model.pkl")
        except Exception as lda_error:
            logger.warning(f"Failed to load LDA model: {lda_error}. Creating placeholder model.")
            from sklearn.decomposition import LatentDirichletAllocation
            lda = LatentDirichletAllocation(n_components=10, random_state=42)
            # Note: Placeholder is untrained; retrain for accurate results
            
    except FileNotFoundError as e:
        logger.warning(f"Missing model artifact: {e}. Running with limited functionality.")
        models_available = False
        # Set placeholder values for models to avoid errors
        from sklearn.ensemble import RandomForestClassifier
        from sklearn.feature_extraction.text import TfidfVectorizer
        from sklearn.feature_selection import SelectKBest
        from sklearn.preprocessing import LabelEncoder
        from sklearn.decomposition import LatentDirichletAllocation
        
        response_clf = RandomForestClassifier()
        crisis_clf = RandomForestClassifier()
        vectorizer = TfidfVectorizer()
        le = LabelEncoder()
        selector = SelectKBest()
        lda = LatentDirichletAllocation(n_components=10)

    # Initialize ChromaDB if possible
    chroma_db_path = f"{output_dir}/chroma_db"
    if not os.path.exists(chroma_db_path):
        logger.warning(f"ChromaDB not found at {chroma_db_path}. Vector search will be unavailable.")
        vector_store = None
    else:
        try:
            logger.info("Initializing ChromaDB")
            if os.environ.get("OPENAI_API_KEY"):
                chroma_client = chromadb.PersistentClient(
                    path=chroma_db_path,
                    settings=Settings(anonymized_telemetry=False)
                )

                embeddings = OpenAIEmbeddings(
                    model="text-embedding-ada-002",
                    api_key=os.environ["OPENAI_API_KEY"],
                    disallowed_special=(), 
                    chunk_size=1000  
                )
                vector_store = Chroma(
                    client=chroma_client,
                    collection_name="mental_health_conversations",
                    embedding_function=embeddings
                )
            else:
                logger.warning("Skipping ChromaDB initialization as OPENAI_API_KEY is not set")
                vector_store = None
        except Exception as e:
            logger.warning(f"Error initializing ChromaDB: {e}")
            vector_store = None

    # Initialize OpenAI client and LLM if API key is available
    logger.info("Initializing OpenAI client and LLM")
    if os.environ.get("OPENAI_API_KEY"):
        try:
            openai_client = OpenAI(api_key=os.environ["OPENAI_API_KEY"])
            llm = ChatOpenAI(
                model="gpt-4o-mini",
                temperature=0.7,
                api_key=os.environ["OPENAI_API_KEY"]
            )
        except Exception as e:
            logger.warning(f"Error initializing OpenAI client: {e}")
            openai_client = None
            llm = None
    else:
        logger.warning("OpenAI client not initialized as OPENAI_API_KEY is not set")
        openai_client = None
        llm = None
        
    # Add route to check model availability
    @app.get("/model-status")
    async def model_status():
        return {
            "models_available": models_available,
            "vector_store_available": vector_store is not None,
            "llm_available": llm is not None,
            "openai_api_key_set": os.environ.get("OPENAI_API_KEY") is not None,
            "langchain_api_key_set": os.environ.get("LANGCHAIN_API_KEY") is not None
        }

# Pydantic model for request
class PatientContext(BaseModel):
    context: str

# New Pydantic models for expanded API functionality
class UserProfile(BaseModel):
    user_id: Optional[str] = None
    username: str
    name: str
    role: str = "counselor"
    specializations: List[str] = []
    years_experience: Optional[int] = None
    custom_crisis_keywords: List[str] = []
    preferences: Dict[str, Any] = {}
    created_at: Optional[datetime] = None
    updated_at: Optional[datetime] = None

class SessionData(BaseModel):
    session_id: Optional[str] = None
    counselor_id: str
    patient_identifier: str  # Anonymized ID
    session_notes: str = ""
    session_preferences: Dict[str, Any] = {}
    crisis_keywords: List[str] = []
    created_at: Optional[datetime] = None
    updated_at: Optional[datetime] = None
    
class ConversationEntry(BaseModel):
    session_id: str
    message: str
    sender: str  # 'patient' or 'counselor'
    timestamp: Optional[datetime] = None
    suggested_response: Optional[str] = None
    response_type: Optional[str] = None
    crisis_flag: bool = False
    risk_level: Optional[str] = None
    
class FeedbackData(BaseModel):
    suggestion_id: str
    counselor_id: str
    rating: int  # 1-5 scale
    was_effective: bool
    comments: Optional[str] = None
    
class AnalysisRequest(BaseModel):
    text: str
    patient_background: Optional[Dict[str, Any]] = None
    patient_age: Optional[int] = None
    cultural_context: Optional[str] = None
    
class MultiModalInput(BaseModel):
    session_id: str
    counselor_id: str
    input_type: str  # 'text', 'audio', 'video'
    content: str  # Text content or file path/url
    metadata: Dict[str, Any] = {}
    
class InterventionRequest(BaseModel):
    patient_issue: str
    patient_background: Optional[Dict[str, Any]] = None
    intervention_type: Optional[str] = None  # e.g., 'CBT', 'DBT', 'mindfulness'

# Text preprocessing function
@traceable(run_type="tool", name="Clean Text")
def clean_text(text):
    if pd.isna(text):
        return ""
    text = str(text).lower()
    text = re.sub(r"[^a-zA-Z']", " ", text)
    
    # Simple tokenization by splitting on whitespace instead of using word_tokenize
    # This avoids the dependency on punkt_tab
    tokens = text.split()
    
    # Filter out stopwords and short tokens
    tokens = [lemmatizer.lemmatize(tok) for tok in tokens if tok not in STOPWORDS and len(tok) > 2]
    return " ".join(tokens)

# Feature engineering function
@traceable(run_type="tool", name="Engineer Features")
def engineer_features(context, response=""):
    try:
        context_clean = clean_text(context)
        context_len = len(context_clean.split())
        context_vader = analyzer.polarity_scores(context)['compound']
        context_questions = context.count('?')
        crisis_keywords = ['suicide', 'hopeless', 'worthless', 'kill', 'harm', 'desperate', 'overwhelmed', 'alone']
        context_crisis_score = sum(1 for word in crisis_keywords if word in context.lower())
        
        # Check if vectorizer is properly initialized
        if vectorizer is None or not hasattr(vectorizer, 'transform'):
            logger.warning("Vectorizer not properly initialized, using placeholder")
            # Create a simple placeholder for features
            features = pd.DataFrame({
                "context_len": [context_len],
                "context_vader": [context_vader],
                "context_questions": [context_questions],
                "crisis_flag": [1 if context_crisis_score > 0 else 0]
            })
            feature_cols = ["context_len", "context_vader", "context_questions", "crisis_flag"]
            return features, feature_cols
            
        # Use vectorizer if available
        context_tfidf = vectorizer.transform([context_clean]).toarray()
        tfidf_cols = [f"tfidf_context_{i}" for i in range(context_tfidf.shape[1])]
        response_tfidf = np.zeros_like(context_tfidf)
        
        # Check if LDA model is properly initialized
        if lda is None or not hasattr(lda, 'transform'):
            logger.warning("LDA model not properly initialized, using zeros")
            lda_topics = np.zeros((1, 10))
        else:
            lda_topics = lda.transform(context_tfidf)
        
        feature_cols = ["context_len", "context_vader", "context_questions", "crisis_flag"] + \
                      [f"topic_{i}" for i in range(10)] + tfidf_cols + \
                      [f"tfidf_response_{i}" for i in range(response_tfidf.shape[1])]
        
        features = pd.DataFrame({
            "context_len": [context_len],
            "context_vader": [context_vader],
            "context_questions": [context_questions],
            **{f"topic_{i}": [lda_topics[0][i]] for i in range(10)},
            **{f"tfidf_context_{i}": [context_tfidf[0][i]] for i in range(context_tfidf.shape[1])},
            **{f"tfidf_response_{i}": [response_tfidf[0][i]] for i in range(response_tfidf.shape[1])},
        })
        
        # Check if crisis classifier is properly initialized
        if crisis_clf is None or not hasattr(crisis_clf, 'predict'):
            logger.warning("Crisis classifier not properly initialized, using keyword detection")
            crisis_flag = 1 if context_crisis_score > 0 else 0
        else:
            crisis_features = features[["context_len", "context_vader", "context_questions"] + [f"topic_{i}" for i in range(10)]]
            crisis_flag = crisis_clf.predict(crisis_features)[0]
            if context_crisis_score > 0:
                crisis_flag = 1
                
        features["crisis_flag"] = crisis_flag
        
        return features, feature_cols
        
    except Exception as e:
        # Fallback to very basic features if anything goes wrong
        logger.error(f"Error in engineer_features: {e}")
        context_len = len(context.split())
        context_questions = context.count('?')
        crisis_keywords = ['suicide', 'hopeless', 'worthless', 'kill', 'harm', 'desperate', 'overwhelmed', 'alone']
        context_crisis_score = sum(1 for word in crisis_keywords if word in context.lower())
        
        features = pd.DataFrame({
            "context_len": [context_len],
            "context_vader": [0.0],  # Default neutral sentiment
            "context_questions": [context_questions],
            "crisis_flag": [1 if context_crisis_score > 0 else 0]
        })
        feature_cols = ["context_len", "context_vader", "context_questions", "crisis_flag"]
        return features, feature_cols

# Prediction function
@traceable(run_type="chain", name="Predict Response Type")
def predict_response_type(context):
    if response_clf is None or vectorizer is None or le is None or selector is None or lda is None:
        logger.warning("Models not available, returning dummy prediction")
        return {
            "response_type": "Empathetic Listening",
            "crisis_flag": False,
            "confidence": 0.5,
            "features": {},
            "models_available": False
        }
        
    features, feature_cols = engineer_features(context)
    selected_features = selector.transform(features[feature_cols])
    pred_encoded = response_clf.predict(selected_features)[0]
    pred_label = le.inverse_transform([pred_encoded])[0]
    confidence = response_clf.predict_proba(selected_features)[0].max()
    
    if "?" in context and context.count("?") > 0:
        pred_label = "Question"
    if "trying" in context.lower() and "hard" in context.lower() and not any(kw in context.lower() for kw in ["how", "what", "help"]):
        pred_label = "Validation"
    if "trying" in context.lower() and "positive" in context.lower() and not any(kw in context.lower() for kw in ["how", "what", "help"]):
        pred_label = "Question"
    
    crisis_flag = bool(features["crisis_flag"].iloc[0])
    
    return {
        "response_type": pred_label,
        "crisis_flag": crisis_flag,
        "confidence": confidence,
        "features": features.to_dict(),
        "models_available": True
    }

# RAG suggestion function
@traceable(run_type="chain", name="RAG Suggestion")
def generate_suggestion_rag(context, response_type, crisis_flag):
    # Check if essential components are available
    if vector_store is None or llm is None:
        logger.warning("Vector store or LLM not available for RAG suggestions, using fallback")
        risk_level = "High" if crisis_flag else "Low"
        
        # Simple fallback suggestions based on response type
        if response_type == "Empathetic Listening":
            suggestion = "I can hear that you're going through a difficult time. It sounds really challenging, and I appreciate you sharing this with me."
        elif response_type == "Question":
            suggestion = "Could you tell me more about how this has been affecting your daily life?"
        elif response_type == "Advice":
            suggestion = "It might be helpful to consider speaking with a mental health professional who can provide personalized support for what you're experiencing."
        elif response_type == "Validation":
            suggestion = "It's completely understandable to feel this way given what you're going through. Your feelings are valid."
        else:
            suggestion = "Thank you for sharing that with me. Let's explore this further together."
            
        # Add crisis resources if needed
        if crisis_flag:
            suggestion += " If you're in crisis, please remember help is available 24/7 through the National Suicide Prevention Lifeline at 988."
            
        return {
            "suggested_response": suggestion,
            "risk_level": risk_level
        }
    
    # If vector store is available, proceed with RAG
    try:
        results = vector_store.similarity_search_with_score(context, k=3)
        retrieved_contexts = [
            f"Patient: {res[0].page_content}\nCounselor: {res[0].metadata['response']} (Type: {res[0].metadata['response_type']}, Crisis: {res[0].metadata['crisis_flag']}, Score: {res[1]:.2f})"
            for res in results
        ]
        
        prompt_template = ChatPromptTemplate.from_template(
            """
            You are an expert mental health counseling assistant. A counselor has provided the following patient situation:
            
            Patient Situation: {context}
            
            Predicted Response Type: {response_type}
            Crisis Flag: {crisis_flag}
            
            Based on the predicted response type and crisis flag, provide a suggested response for the counselor to use with the patient. The response should align with the response type ({response_type}) and be sensitive to the crisis level.
            
            For reference, here are similar cases from past conversations:
            {retrieved_contexts}
            
            Guidelines:
            - If Crisis Flag is True, prioritize safety, empathy, and suggest immediate resources (e.g., National Suicide Prevention Lifeline at 988).
            - For 'Empathetic Listening', focus on validating feelings without giving direct advice or questions.
            - For 'Advice', provide practical, actionable suggestions.
            - For 'Question', pose an open-ended question to encourage further discussion.
            - For 'Validation', affirm the patient's efforts or feelings.
            
            Output in the following format:
            ```json
            {{
                "suggested_response": "Your suggested response here",
                "risk_level": "Low/Moderate/High"
            }}
            ```
            """
        )
        
        rag_chain = (
            {
                "context": RunnablePassthrough(),
                "response_type": lambda x: response_type,
                "crisis_flag": lambda x: "Crisis" if crisis_flag else "No Crisis",
                "retrieved_contexts": lambda x: "\n".join(retrieved_contexts)
            }
            | prompt_template
            | llm
        )
        
        response = rag_chain.invoke(context)
        return eval(response.content.strip("```json\n").strip("\n```"))
    except Exception as e:
        logger.error(f"Error generating RAG suggestion: {e}")
        risk_level = "High" if crisis_flag else "Low"
        
        # Fallback suggestion if RAG fails
        if crisis_flag:
            suggestion = "I'm hearing that you're going through a very difficult time. Your safety is the most important thing right now. Would it be helpful to talk about resources that are available to support you, like the National Suicide Prevention Lifeline at 988?"
        else:
            suggestion = "Thank you for sharing that with me. I want to understand more about your experience and how I can best support you right now."
            
        return {
            "suggested_response": suggestion,
            "risk_level": risk_level
        }

# Direct suggestion function
@traceable(run_type="chain", name="Direct Suggestion")
def generate_suggestion_direct(context, response_type, crisis_flag):
    # Check if essential components are available
    if llm is None:
        logger.warning("LLM not available for direct suggestions, using fallback")
        risk_level = "High" if crisis_flag else "Low"
        
        # Simple fallback suggestions based on response type
        if response_type == "Empathetic Listening":
            suggestion = "It sounds like this has been really difficult for you. I'm here to listen and support you."
        elif response_type == "Question":
            suggestion = "How have you been coping with these feelings recently?"
        elif response_type == "Advice":
            suggestion = "One thing that might help is establishing a simple morning routine with small, achievable steps."
        elif response_type == "Validation":
            suggestion = "What you're experiencing is a normal response to a difficult situation. Your feelings are valid."
        else:
            suggestion = "I appreciate you sharing this with me. Let's work through this together."
            
        # Add crisis resources if needed
        if crisis_flag:
            suggestion += " Given what you've shared, I want to make sure you know about resources like the National Suicide Prevention Lifeline at 988, which is available 24/7."
            
        return {
            "suggested_response": suggestion,
            "risk_level": risk_level
        }
    
    # If LLM is available, proceed with direct suggestion
    try:
        prompt_template = ChatPromptTemplate.from_template(
            """
            You are an expert mental health counseling assistant. A counselor has provided the following patient situation:
            
            Patient Situation: {context}
            
            Predicted Response Type: {response_type}
            Crisis Flag: {crisis_flag}
            
            Provide a suggested response for the counselor to use with the patient, aligned with the response type ({response_type}) and sensitive to the crisis level.
            
            Guidelines:
            - If Crisis Flag is True, prioritize safety, empathy, and suggest immediate resources (e.g., National Suicide Prevention Lifeline at 988).
            - For 'Empathetic Listening', focus on validating feelings without giving direct advice or questions.
            - For 'Advice', provide practical, actionable suggestions.
            - For 'Question', pose an open-ended question to encourage further discussion.
            - For 'Validation', affirm the patient's efforts or feelings.
            - Strictly adhere to the response type. For 'Empathetic Listening', do not include questions or advice.
            
            Output in the following format:
            ```json
            {{
                "suggested_response": "Your suggested response here",
                "risk_level": "Low/Moderate/High"
            }}
            ```
            """
        )
        
        direct_chain = (
            {
                "context": RunnablePassthrough(),
                "response_type": lambda x: response_type,
                "crisis_flag": lambda x: "Crisis" if crisis_flag else "No Crisis"
            }
            | prompt_template
            | llm
        )
        
        response = direct_chain.invoke(context)
        return eval(response.content.strip("```json\n").strip("\n```"))
    except Exception as e:
        logger.error(f"Error generating direct suggestion: {e}")
        risk_level = "High" if crisis_flag else "Low"
        
        # Fallback suggestion if direct generation fails
        if crisis_flag:
            suggestion = "I'm concerned about what you're sharing. Your wellbeing is important, and I want to make sure you have support. The National Suicide Prevention Lifeline (988) has trained counselors available 24/7."
        else:
            suggestion = "I hear you're having a difficult time. Would you like to talk more about how these feelings have been affecting you?"
            
        return {
            "suggested_response": suggestion,
            "risk_level": risk_level
        }

# User Profile Endpoints
@app.post("/users/create", response_model=UserProfile)
async def create_user(profile: UserProfile):
    """Create a new counselor profile with preferences and specializations."""
    try:
        saved_profile = save_user_profile(profile)
        logger.info(f"Created user profile: {saved_profile.user_id}")
        return saved_profile
    except Exception as e:
        logger.error(f"Error creating user profile: {e}")
        raise HTTPException(status_code=500, detail=f"Error creating user profile: {str(e)}")

@app.get("/users/{user_id}", response_model=UserProfile)
async def get_user(user_id: str):
    """Get a counselor profile by user ID."""
    profile = get_user_profile(user_id)
    if not profile:
        raise HTTPException(status_code=404, detail=f"User profile not found: {user_id}")
    return profile

@app.put("/users/{user_id}", response_model=UserProfile)
async def update_user(user_id: str, profile_update: UserProfile):
    """Update a counselor profile."""
    existing_profile = get_user_profile(user_id)
    if not existing_profile:
        raise HTTPException(status_code=404, detail=f"User profile not found: {user_id}")
    
    # Preserve the original user_id
    profile_update.user_id = user_id
    # Preserve the original created_at timestamp
    profile_update.created_at = existing_profile.created_at
    
    try:
        updated_profile = save_user_profile(profile_update)
        logger.info(f"Updated user profile: {user_id}")
        return updated_profile
    except Exception as e:
        logger.error(f"Error updating user profile: {e}")
        raise HTTPException(status_code=500, detail=f"Error updating user profile: {str(e)}")

# Session Management Endpoints
@app.post("/sessions/create", response_model=SessionData)
async def create_session(session_data: SessionData):
    """Create a new session with patient identifier (anonymized)."""
    try:
        # Verify counselor exists
        counselor = get_user_profile(session_data.counselor_id)
        if not counselor:
            raise HTTPException(status_code=404, detail=f"Counselor not found: {session_data.counselor_id}")
        
        # If counselor has custom crisis keywords, add them to the session
        if counselor.custom_crisis_keywords:
            session_data.crisis_keywords.extend(counselor.custom_crisis_keywords)
        
        saved_session = save_session(session_data)
        logger.info(f"Created session: {saved_session.session_id}")
        return saved_session
    except HTTPException:
        raise
    except Exception as e:
        logger.error(f"Error creating session: {e}")
        raise HTTPException(status_code=500, detail=f"Error creating session: {str(e)}")

@app.get("/sessions/{session_id}", response_model=SessionData)
async def get_session_by_id(session_id: str):
    """Get a session by ID."""
    session = get_session(session_id)
    if not session:
        raise HTTPException(status_code=404, detail=f"Session not found: {session_id}")
    return session

@app.get("/sessions/counselor/{counselor_id}", response_model=List[SessionData])
async def get_counselor_sessions(counselor_id: str):
    """Get all sessions for a counselor."""
    sessions = get_user_sessions(counselor_id)
    return sessions

@app.put("/sessions/{session_id}", response_model=SessionData)
async def update_session(session_id: str, session_update: SessionData):
    """Update a session."""
    existing_session = get_session(session_id)
    if not existing_session:
        raise HTTPException(status_code=404, detail=f"Session not found: {session_id}")
    
    # Preserve the original session_id and created_at
    session_update.session_id = session_id
    session_update.created_at = existing_session.created_at
    
    try:
        updated_session = save_session(session_update)
        logger.info(f"Updated session: {session_id}")
        return updated_session
    except Exception as e:
        logger.error(f"Error updating session: {e}")
        raise HTTPException(status_code=500, detail=f"Error updating session: {str(e)}")

# Conversation History Endpoints
@app.post("/conversations/add", response_model=str)
async def add_conversation_entry(entry: ConversationEntry):
    """Add a new entry to a conversation."""
    try:
        # Verify session exists
        session = get_session(entry.session_id)
        if not session:
            raise HTTPException(status_code=404, detail=f"Session not found: {entry.session_id}")
        
        entry_id = save_conversation_entry(entry)
        logger.info(f"Added conversation entry: {entry_id}")
        return entry_id
    except HTTPException:
        raise
    except Exception as e:
        logger.error(f"Error adding conversation entry: {e}")
        raise HTTPException(status_code=500, detail=f"Error adding conversation entry: {str(e)}")

@app.get("/conversations/{session_id}", response_model=List[ConversationEntry])
async def get_conversation(session_id: str):
    """Get conversation history for a session."""
    try:
        # Verify session exists
        session = get_session(session_id)
        if not session:
            raise HTTPException(status_code=404, detail=f"Session not found: {session_id}")
        
        entries = get_conversation_history(session_id)
        return entries
    except HTTPException:
        raise
    except Exception as e:
        logger.error(f"Error retrieving conversation history: {e}")
        raise HTTPException(status_code=500, detail=f"Error retrieving conversation history: {str(e)}")

# API Endpoints
@app.post("/suggest")
async def get_suggestion(context: PatientContext):
    logger.info(f"Received suggestion request for context: {context.context}")
    prediction = predict_response_type(context.context)
    suggestion_rag = generate_suggestion_rag(context.context, prediction["response_type"], prediction["crisis_flag"])
    suggestion_direct = generate_suggestion_direct(context.context, prediction["response_type"], prediction["crisis_flag"])
    
    return {
        "context": context.context,
        "response_type": prediction["response_type"],
        "crisis_flag": prediction["crisis_flag"],
        "confidence": prediction["confidence"],
        "rag_suggestion": suggestion_rag["suggested_response"],
        "rag_risk_level": suggestion_rag["risk_level"],
        "direct_suggestion": suggestion_direct["suggested_response"],
        "direct_risk_level": suggestion_direct["risk_level"]
    }

@app.post("/session/suggest")
async def get_session_suggestion(request: dict):
    """Get suggestion within a session context, with enhanced crisis detection based on session keywords."""
    try:
        session_id = request.get("session_id")
        if not session_id:
            raise HTTPException(status_code=400, detail="session_id is required")
        
        context = request.get("context")
        if not context:
            raise HTTPException(status_code=400, detail="context is required")
        
        # Get session for custom crisis keywords
        session = get_session(session_id)
        if not session:
            raise HTTPException(status_code=404, detail=f"Session not found: {session_id}")
        
        # Get conversation history for context
        conversation_history = get_conversation_history(session_id)
        
        # Regular prediction
        prediction = predict_response_type(context)
        crisis_flag = prediction["crisis_flag"]
        
        # Enhanced crisis detection with custom keywords
        if not crisis_flag and session.crisis_keywords:
            for keyword in session.crisis_keywords:
                if keyword.lower() in context.lower():
                    crisis_flag = True
                    logger.info(f"Crisis flag triggered by custom keyword: {keyword}")
                    break
        
        # Generate suggestions
        suggestion_rag = generate_suggestion_rag(context, prediction["response_type"], crisis_flag)
        suggestion_direct = generate_suggestion_direct(context, prediction["response_type"], crisis_flag)
        
        # Create response
        response = {
            "context": context,
            "response_type": prediction["response_type"],
            "crisis_flag": crisis_flag,
            "confidence": prediction["confidence"],
            "rag_suggestion": suggestion_rag["suggested_response"],
            "rag_risk_level": suggestion_rag["risk_level"],
            "direct_suggestion": suggestion_direct["suggested_response"],
            "direct_risk_level": suggestion_direct["risk_level"],
            "session_id": session_id
        }
        
        # Save the conversation entry
        entry = ConversationEntry(
            session_id=session_id,
            message=context,
            sender="patient",
            suggested_response=suggestion_rag["suggested_response"],
            response_type=prediction["response_type"],
            crisis_flag=crisis_flag,
            risk_level=suggestion_rag["risk_level"]
        )
        save_conversation_entry(entry)
        
        return response
    except HTTPException:
        raise
    except Exception as e:
        logger.error(f"Error getting session suggestion: {e}")
        raise HTTPException(status_code=500, detail=f"Error getting session suggestion: {str(e)}")

# Feedback Endpoints
@app.post("/feedback")
async def add_feedback(feedback: FeedbackData):
    """Add feedback about a suggestion's effectiveness."""
    try:
        feedback_id = save_feedback(feedback)
        logger.info(f"Added feedback: {feedback_id}")
        return {"feedback_id": feedback_id}
    except Exception as e:
        logger.error(f"Error adding feedback: {e}")
        raise HTTPException(status_code=500, detail=f"Error adding feedback: {str(e)}")

# Tone & Cultural Sensitivity Analysis
@traceable(run_type="chain", name="Cultural Sensitivity Analysis")
def analyze_cultural_sensitivity(text: str, cultural_context: Optional[str] = None):
    """Analyze text for cultural appropriateness and sensitivity."""
    prompt_template = ChatPromptTemplate.from_template(
        """
        You are a cultural sensitivity expert. Analyze the following text for cultural appropriateness:
        
        Text: {text}
        
        Cultural Context: {cultural_context}
        
        Provide an analysis of:
        1. Cultural appropriateness
        2. Potential bias or insensitivity
        3. Suggestions for improvement
        
        Output in the following format:
        ```json
        {{
            "cultural_appropriateness_score": 0-10,
            "issues_detected": ["issue1", "issue2"],
            "suggestions": ["suggestion1", "suggestion2"],
            "explanation": "Brief explanation of analysis"
        }}
        ```
        """
    )
    
    analysis_chain = (
        {
            "text": RunnablePassthrough(),
            "cultural_context": lambda x: cultural_context if cultural_context else "General"
        }
        | prompt_template
        | llm
    )
    
    try:
        response = analysis_chain.invoke(text)
        return eval(response.content.strip("```json\n").strip("\n```"))
    except Exception as e:
        logger.error(f"Error analyzing cultural sensitivity: {e}")
        raise HTTPException(status_code=500, detail=f"Error analyzing cultural sensitivity: {str(e)}")

@traceable(run_type="chain", name="Age Appropriate Analysis")
def analyze_age_appropriateness(text: str, age: Optional[int] = None):
    """Analyze text for age-appropriate language."""
    prompt_template = ChatPromptTemplate.from_template(
        """
        You are an expert in age-appropriate communication. Analyze the following text for age appropriateness:
        
        Text: {text}
        
        Target Age: {age}
        
        Provide an analysis of:
        1. Age appropriateness
        2. Complexity level
        3. Suggestions for improvement
        
        Output in the following format:
        ```json
        {{
            "age_appropriateness_score": 0-10,
            "complexity_level": "Simple/Moderate/Complex",
            "issues_detected": ["issue1", "issue2"],
            "suggestions": ["suggestion1", "suggestion2"],
            "explanation": "Brief explanation of analysis"
        }}
        ```
        """
    )
    
    analysis_chain = (
        {
            "text": RunnablePassthrough(),
            "age": lambda x: str(age) if age else "Adult"
        }
        | prompt_template
        | llm
    )
    
    try:
        response = analysis_chain.invoke(text)
        return eval(response.content.strip("```json\n").strip("\n```"))
    except Exception as e:
        logger.error(f"Error analyzing age appropriateness: {e}")
        raise HTTPException(status_code=500, detail=f"Error analyzing age appropriateness: {str(e)}")

@app.post("/analyze/sensitivity")
async def analyze_text_sensitivity(request: AnalysisRequest):
    """Analyze text for cultural sensitivity and age appropriateness."""
    try:
        cultural_analysis = analyze_cultural_sensitivity(request.text, request.cultural_context)
        age_analysis = analyze_age_appropriateness(request.text, request.patient_age)
        
        return {
            "text": request.text,
            "cultural_analysis": cultural_analysis,
            "age_analysis": age_analysis
        }
    except Exception as e:
        logger.error(f"Error analyzing text sensitivity: {e}")
        raise HTTPException(status_code=500, detail=f"Error analyzing text sensitivity: {str(e)}")

# Guided Intervention Workflows
@traceable(run_type="chain", name="Generate Intervention")
def generate_intervention_workflow(issue: str, intervention_type: Optional[str] = None, background: Optional[Dict] = None):
    """Generate a structured intervention workflow for a specific issue."""
    prompt_template = ChatPromptTemplate.from_template(
        """
        You are an expert mental health counselor. Generate a structured intervention workflow for the following patient issue:
        
        Patient Issue: {issue}
        
        Intervention Type: {intervention_type}
        Patient Background: {background}
        
        Provide a step-by-step intervention plan based on evidence-based practices. Include:
        1. Initial assessment questions
        2. Specific techniques to apply
        3. Homework or practice exercises
        4. Follow-up guidance
        
        Output in the following format:
        ```json
        {{
            "intervention_type": "CBT/DBT/ACT/Mindfulness/etc.",
            "assessment_questions": ["question1", "question2", "question3"],
            "techniques": [
                {{
                    "name": "technique name",
                    "description": "brief description",
                    "instructions": "step-by-step instructions"
                }}
            ],
            "exercises": [
                {{
                    "name": "exercise name",
                    "description": "brief description",
                    "instructions": "step-by-step instructions"
                }}
            ],
            "follow_up": ["follow-up step 1", "follow-up step 2"],
            "resources": ["resource1", "resource2"]
        }}
        ```
        """
    )
    
    intervention_chain = (
        {
            "issue": RunnablePassthrough(),
            "intervention_type": lambda x: intervention_type if intervention_type else "Best fit",
            "background": lambda x: str(background) if background else "Not provided"
        }
        | prompt_template
        | llm
    )
    
    try:
        response = intervention_chain.invoke(issue)
        return eval(response.content.strip("```json\n").strip("\n```"))
    except Exception as e:
        logger.error(f"Error generating intervention workflow: {e}")
        raise HTTPException(status_code=500, detail=f"Error generating intervention workflow: {str(e)}")

@app.post("/interventions/generate")
async def get_intervention_workflow(request: InterventionRequest):
    """Get a structured intervention workflow for a specific patient issue."""
    try:
        intervention = generate_intervention_workflow(
            request.patient_issue, 
            request.intervention_type, 
            request.patient_background
        )
        
        return {
            "patient_issue": request.patient_issue,
            "intervention": intervention
        }
    except Exception as e:
        logger.error(f"Error generating intervention workflow: {e}")
        raise HTTPException(status_code=500, detail=f"Error generating intervention workflow: {str(e)}")

@app.get("/health")
async def health_check():
    if all([response_clf, crisis_clf, vectorizer, le, selector, lda, vector_store, llm]):
        return {"status": "healthy", "message": "All models and vector store loaded successfully"}
    logger.error("Health check failed: One or more components not loaded")
    raise HTTPException(status_code=500, detail="One or more components failed to load")

@app.get("/metadata")
async def get_metadata():
    try:
        collection = vector_store._client.get_collection("mental_health_conversations")
        count = collection.count()
        return {"collection_name": "mental_health_conversations", "document_count": count}
    except Exception as e:
        logger.error(f"Error retrieving metadata: {e}")
        raise HTTPException(status_code=500, detail=f"Error retrieving metadata: {str(e)}")

# Database utility functions
def save_user_profile(profile: UserProfile):
    if not profile.user_id:
        profile.user_id = str(uuid4())
    
    if not profile.created_at:
        profile.created_at = datetime.now()
    
    profile.updated_at = datetime.now()
    
    file_path = os.path.join(DATA_DIR, "users", f"{profile.user_id}.json")
    with open(file_path, "w") as f:
        # Convert datetime to string for JSON serialization
        profile_dict = profile.dict()
        for key in ["created_at", "updated_at"]:
            if profile_dict[key]:
                profile_dict[key] = profile_dict[key].isoformat()
        f.write(json.dumps(profile_dict, indent=2))
    
    return profile

def get_user_profile(user_id: str) -> Optional[UserProfile]:
    file_path = os.path.join(DATA_DIR, "users", f"{user_id}.json")
    if not os.path.exists(file_path):
        return None
    
    with open(file_path, "r") as f:
        data = json.loads(f.read())
        # Convert string dates back to datetime
        for key in ["created_at", "updated_at"]:
            if data[key]:
                data[key] = datetime.fromisoformat(data[key])
        return UserProfile(**data)

def save_session(session: SessionData):
    if not session.session_id:
        session.session_id = str(uuid4())
    
    if not session.created_at:
        session.created_at = datetime.now()
    
    session.updated_at = datetime.now()
    
    file_path = os.path.join(DATA_DIR, "sessions", f"{session.session_id}.json")
    with open(file_path, "w") as f:
        # Convert datetime to string for JSON serialization
        session_dict = session.dict()
        for key in ["created_at", "updated_at"]:
            if session_dict[key]:
                session_dict[key] = session_dict[key].isoformat()
        f.write(json.dumps(session_dict, indent=2))
    
    return session

def get_session(session_id: str) -> Optional[SessionData]:
    file_path = os.path.join(DATA_DIR, "sessions", f"{session_id}.json")
    if not os.path.exists(file_path):
        return None
    
    with open(file_path, "r") as f:
        data = json.loads(f.read())
        # Convert string dates back to datetime
        for key in ["created_at", "updated_at"]:
            if data[key]:
                data[key] = datetime.fromisoformat(data[key])
        return SessionData(**data)

def get_user_sessions(counselor_id: str) -> List[SessionData]:
    sessions = []
    sessions_dir = os.path.join(DATA_DIR, "sessions")
    for filename in os.listdir(sessions_dir):
        if not filename.endswith(".json"):
            continue
        
        file_path = os.path.join(sessions_dir, filename)
        with open(file_path, "r") as f:
            data = json.loads(f.read())
            if data["counselor_id"] == counselor_id:
                for key in ["created_at", "updated_at"]:
                    if data[key]:
                        data[key] = datetime.fromisoformat(data[key])
                sessions.append(SessionData(**data))
    
    return sessions

def save_conversation_entry(entry: ConversationEntry):
    conversation_dir = os.path.join(DATA_DIR, "conversations", entry.session_id)
    os.makedirs(conversation_dir, exist_ok=True)
    
    if not entry.timestamp:
        entry.timestamp = datetime.now()
    
    entry_id = str(uuid4())
    file_path = os.path.join(conversation_dir, f"{entry_id}.json")
    
    with open(file_path, "w") as f:
        # Convert datetime to string for JSON serialization
        entry_dict = entry.dict()
        entry_dict["entry_id"] = entry_id
        if entry_dict["timestamp"]:
            entry_dict["timestamp"] = entry_dict["timestamp"].isoformat()
        f.write(json.dumps(entry_dict, indent=2))
    
    return entry_id

def get_conversation_history(session_id: str) -> List[ConversationEntry]:
    conversation_dir = os.path.join(DATA_DIR, "conversations", session_id)
    if not os.path.exists(conversation_dir):
        return []
    
    entries = []
    for filename in os.listdir(conversation_dir):
        if not filename.endswith(".json"):
            continue
        
        file_path = os.path.join(conversation_dir, filename)
        with open(file_path, "r") as f:
            data = json.loads(f.read())
            if data["timestamp"]:
                data["timestamp"] = datetime.fromisoformat(data["timestamp"])
            entries.append(ConversationEntry(**data))
    
    # Sort by timestamp
    entries.sort(key=lambda x: x.timestamp)
    return entries

def save_feedback(feedback: FeedbackData):
    feedback_id = str(uuid4())
    file_path = os.path.join(DATA_DIR, "feedback", f"{feedback_id}.json")
    
    with open(file_path, "w") as f:
        feedback_dict = feedback.dict()
        feedback_dict["feedback_id"] = feedback_id
        feedback_dict["timestamp"] = datetime.now().isoformat()
        f.write(json.dumps(feedback_dict, indent=2))
    
    return feedback_id

# Multi-modal Input Support
@app.post("/input/process")
async def process_multimodal_input(input_data: MultiModalInput):
    """Process multi-modal input (text, audio, video)."""
    try:
        if input_data.input_type not in ["text", "audio", "video"]:
            raise HTTPException(status_code=400, detail=f"Unsupported input type: {input_data.input_type}")
        
        # For now, handle text directly and simulate processing for audio/video
        if input_data.input_type == "text":
            # Process text normally
            prediction = predict_response_type(input_data.content)
            
            return {
                "input_type": "text",
                "processed_content": input_data.content,
                "analysis": {
                    "response_type": prediction["response_type"],
                    "crisis_flag": prediction["crisis_flag"],
                    "confidence": prediction["confidence"]
                }
            }
        elif input_data.input_type == "audio":
            # Simulate audio transcription and emotion detection
            # In a production system, this would use a speech-to-text API and emotion analysis
            prompt_template = ChatPromptTemplate.from_template(
                """
                Simulate audio processing for this description: {content}
                
                Generate a simulated transcription and emotion detection as if this were real audio.
                
                Output in the following format:
                ```json
                {{
                    "transcription": "Simulated transcription of the audio",
                    "emotion_detected": "primary emotion",
                    "secondary_emotions": ["emotion1", "emotion2"],
                    "confidence": 0.85
                }}
                ```
                """
            )
            
            process_chain = prompt_template | llm
            response = process_chain.invoke({"content": input_data.content})
            audio_result = eval(response.content.strip("```json\n").strip("\n```"))
            
            # Now process the transcription
            prediction = predict_response_type(audio_result["transcription"])
            
            return {
                "input_type": "audio",
                "processed_content": audio_result["transcription"],
                "emotion_analysis": {
                    "primary_emotion": audio_result["emotion_detected"],
                    "secondary_emotions": audio_result["secondary_emotions"],
                    "confidence": audio_result["confidence"]
                },
                "analysis": {
                    "response_type": prediction["response_type"],
                    "crisis_flag": prediction["crisis_flag"],
                    "confidence": prediction["confidence"]
                }
            }
        elif input_data.input_type == "video":
            # Simulate video analysis
            # In a production system, this would use video analytics API
            prompt_template = ChatPromptTemplate.from_template(
                """
                Simulate video processing for this description: {content}
                
                Generate a simulated analysis as if this were real video with facial expressions and body language.
                
                Output in the following format:
                ```json
                {{
                    "transcription": "Simulated transcription of speech in the video",
                    "facial_expressions": ["expression1", "expression2"],
                    "body_language": ["posture observation", "gesture observation"],
                    "primary_emotion": "primary emotion",
                    "confidence": 0.80
                }}
                ```
                """
            )
            
            process_chain = prompt_template | llm
            response = process_chain.invoke({"content": input_data.content})
            video_result = eval(response.content.strip("```json\n").strip("\n```"))
            
            # Process the transcription
            prediction = predict_response_type(video_result["transcription"])
            
            return {
                "input_type": "video",
                "processed_content": video_result["transcription"],
                "nonverbal_analysis": {
                    "facial_expressions": video_result["facial_expressions"],
                    "body_language": video_result["body_language"],
                    "primary_emotion": video_result["primary_emotion"],
                    "confidence": video_result["confidence"]
                },
                "analysis": {
                    "response_type": prediction["response_type"],
                    "crisis_flag": prediction["crisis_flag"],
                    "confidence": prediction["confidence"]
                }
            }
    except Exception as e:
        logger.error(f"Error processing multimodal input: {e}")
        raise HTTPException(status_code=500, detail=f"Error processing multimodal input: {str(e)}")

# Therapeutic Technique Suggestions
@traceable(run_type="chain", name="Therapeutic Techniques")
def suggest_therapeutic_techniques(context: str, technique_type: Optional[str] = None):
    """Suggest specific therapeutic techniques based on the patient context."""
    prompt_template = ChatPromptTemplate.from_template(
        """
        You are an expert mental health professional with extensive knowledge of therapeutic techniques. Based on the following patient context, suggest therapeutic techniques that would be appropriate:
        
        Patient Context: {context}
        
        Technique Type (if specified): {technique_type}
        
        Suggest specific therapeutic techniques, exercises, or interventions that would be helpful for this patient. Include:
        1. Name of technique
        2. Brief description
        3. How to apply it in this specific case
        4. Expected benefits
        
        Provide a range of options from different therapeutic approaches (CBT, DBT, ACT, mindfulness, motivational interviewing, etc.) unless a specific technique type was requested.
        
        Output in the following format:
        ```json
        {{
            "primary_approach": "The most appropriate therapeutic approach",
            "techniques": [
                {{
                    "name": "Technique name",
                    "approach": "CBT/DBT/ACT/etc.",
                    "description": "Brief description",
                    "application": "How to apply to this specific case",
                    "benefits": "Expected benefits"
                }}
            ],
            "rationale": "Brief explanation of why these techniques were selected"
        }}
        ```
        """
    )
    
    technique_chain = (
        {
            "context": RunnablePassthrough(),
            "technique_type": lambda x: technique_type if technique_type else "Any appropriate"
        }
        | prompt_template
        | llm
    )
    
    try:
        response = technique_chain.invoke(context)
        return eval(response.content.strip("```json\n").strip("\n```"))
    except Exception as e:
        logger.error(f"Error suggesting therapeutic techniques: {e}")
        raise HTTPException(status_code=500, detail=f"Error suggesting therapeutic techniques: {str(e)}")

@app.post("/techniques/suggest")
async def get_therapeutic_techniques(request: dict):
    """Get suggested therapeutic techniques for a patient context."""
    try:
        context = request.get("context")
        if not context:
            raise HTTPException(status_code=400, detail="context is required")
        
        technique_type = request.get("technique_type")
        
        techniques = suggest_therapeutic_techniques(context, technique_type)
        
        return {
            "context": context,
            "techniques": techniques
        }
    except Exception as e:
        logger.error(f"Error getting therapeutic techniques: {e}")
        raise HTTPException(status_code=500, detail=f"Error getting therapeutic techniques: {str(e)}")

# Ethical AI Guardrails - Confidence Indicator
@app.post("/suggest/with_confidence")
async def get_suggestion_with_confidence(context: PatientContext):
    """Get suggestion with detailed confidence indicators and uncertainty flags."""
    try:
        # Get standard prediction
        prediction = predict_response_type(context.context)
        
        # Set confidence thresholds
        high_confidence = 0.8
        medium_confidence = 0.6
        
        # Determine confidence level
        confidence_value = prediction["confidence"]
        if confidence_value >= high_confidence:
            confidence_level = "High"
        elif confidence_value >= medium_confidence:
            confidence_level = "Medium"
        else:
            confidence_level = "Low"
        
        # Analyze for potential biases
        bias_prompt = ChatPromptTemplate.from_template(
            """
            You are an AI ethics expert. Analyze the following patient context and proposed response type for potential biases:
            
            Patient Context: {context}
            Predicted Response Type: {response_type}
            
            Identify any potential biases in interpretation or response. Consider gender, cultural, socioeconomic, and other potential biases.
            
            Output in the following format:
            ```json
            {{
                "bias_detected": true/false,
                "bias_types": ["bias type 1", "bias type 2"],
                "explanation": "Brief explanation of potential biases"
            }}
            ```
            """
        )
        
        bias_chain = (
            {
                "context": lambda x: context.context,
                "response_type": lambda x: prediction["response_type"]
            }
            | bias_prompt
            | llm
        )
        
        bias_analysis = eval(bias_chain.invoke({}).content.strip("```json\n").strip("\n```"))
        
        # Generate suggestions
        suggestion_rag = generate_suggestion_rag(context.context, prediction["response_type"], prediction["crisis_flag"])
        suggestion_direct = generate_suggestion_direct(context.context, prediction["response_type"], prediction["crisis_flag"])
        
        return {
            "context": context.context,
            "response_type": prediction["response_type"],
            "crisis_flag": prediction["crisis_flag"],
            "confidence": {
                "value": prediction["confidence"],
                "level": confidence_level,
                "uncertainty_flag": confidence_level == "Low"
            },
            "bias_analysis": bias_analysis,
            "rag_suggestion": suggestion_rag["suggested_response"],
            "rag_risk_level": suggestion_rag["risk_level"],
            "direct_suggestion": suggestion_direct["suggested_response"],
            "direct_risk_level": suggestion_direct["risk_level"],
            "attribution": {
                "ai_generated": True,
                "model_version": "Mental Health Counselor API v2.0",
                "human_reviewed": False
            }
        }
    except Exception as e:
        logger.error(f"Error getting suggestion with confidence: {e}")
        raise HTTPException(status_code=500, detail=f"Error getting suggestion with confidence: {str(e)}")

# Text to Speech with Eleven Labs API
@app.post("/api/text-to-speech")
async def text_to_speech(request: dict):
    """Convert text to speech using Eleven Labs API."""
    try:
        text = request.get("text")
        voice_id = request.get("voice_id", "pNInz6obpgDQGcFmaJgB") # Default to "Adam" voice
        
        if not text:
            raise HTTPException(status_code=400, detail="Text is required")
        
        # Get API key from environment
        api_key = os.getenv("ELEVEN_API_KEY")
        if not api_key:
            raise HTTPException(status_code=500, detail="Eleven Labs API key not configured")
        
        # Prepare the request to Eleven Labs
        url = f"https://api.elevenlabs.io/v1/text-to-speech/{voice_id}"
        
        headers = {
            "Accept": "audio/mpeg",
            "Content-Type": "application/json",
            "xi-api-key": api_key
        }
        
        payload = {
            "text": text,
            "model_id": "eleven_multilingual_v2",
            "voice_settings": {
                "stability": 0.5,
                "similarity_boost": 0.75
            }
        }
        
        # Make the request to Eleven Labs
        response = requests.post(url, json=payload, headers=headers)
        
        if response.status_code != 200:
            logger.error(f"Eleven Labs API error: {response.text}")
            raise HTTPException(status_code=response.status_code, 
                               detail=f"Eleven Labs API error: {response.text}")
        
        # Return audio as streaming response
        return StreamingResponse(
            BytesIO(response.content),
            media_type="audio/mpeg"
        )
        
    except Exception as e:
        logger.error(f"Error in text-to-speech: {str(e)}")
        if not isinstance(e, HTTPException):
            raise HTTPException(status_code=500, detail=f"Text-to-speech error: {str(e)}")
        raise e

# Multimedia file processing (speech to text)
@app.post("/api/input/process")
async def process_audio_input(
    audio: UploadFile = File(...),
    session_id: str = Form(...)
):
    """Process audio input for speech-to-text using Eleven Labs."""
    try:
        # Get API key from environment
        api_key = os.getenv("ELEVEN_API_KEY")
        if not api_key:
            raise HTTPException(status_code=500, detail="Eleven Labs API key not configured")
        
        # Read the audio file content
        audio_content = await audio.read()
        
        # Call Eleven Labs Speech-to-Text API
        url = "https://api.elevenlabs.io/v1/speech-to-text"
        
        headers = {
            "xi-api-key": api_key
        }
        
        # Create form data with the audio file
        files = {
            'audio': ('audio.webm', audio_content, 'audio/webm')
        }
        
        data = {
            'model_id': 'whisper-1'  # Using Whisper model
        }
        
        # Make the request to Eleven Labs
        response = requests.post(url, headers=headers, files=files, data=data)
        
        if response.status_code != 200:
            logger.error(f"Eleven Labs API error: {response.text}")
            raise HTTPException(status_code=response.status_code, 
                              detail=f"Eleven Labs API error: {response.text}")
        
        result = response.json()
        
        # Extract the transcribed text
        text = result.get('text', '')
        
        # Return the transcribed text
        return {
            "text": text,
            "session_id": session_id
        }
        
    except Exception as e:
        logger.error(f"Error processing audio: {str(e)}")
        if not isinstance(e, HTTPException):
            raise HTTPException(status_code=500, detail=f"Audio processing error: {str(e)}")
        raise e

# Add a custom encoder for bytes objects to prevent UTF-8 decode errors
def custom_encoder(obj):
    if isinstance(obj, bytes):
        try:
            return obj.decode('utf-8')
        except UnicodeDecodeError:
            return base64.b64encode(obj).decode('ascii')
    raise TypeError(f"Object of type {type(obj)} is not JSON serializable")

# Override the jsonable_encoder function to handle bytes properly
from fastapi.encoders import jsonable_encoder as original_jsonable_encoder

def safe_jsonable_encoder(*args, **kwargs):
    try:
        return original_jsonable_encoder(*args, **kwargs)
    except UnicodeDecodeError:
        # If the standard encoder fails with a decode error, 
        # ensure all bytes are properly handled
        if args and isinstance(args[0], bytes):
            return custom_encoder(args[0])
        raise

# Monkey patch the jsonable_encoder in FastAPI
import fastapi.encoders
fastapi.encoders.jsonable_encoder = safe_jsonable_encoder