import torch
from transformers import AutoModelForQuestionAnswering, AutoTokenizer, pipeline
from sentence_transformers import SentenceTransformer, util
import gradio as gr
import json

# Load pre-trained BERT QA model and tokenizer from Hugging Face model hub
model_name = "bert-large-uncased-whole-word-masking-finetuned-squad"
model = AutoModelForQuestionAnswering.from_pretrained(model_name)
tokenizer = AutoTokenizer.from_pretrained(model_name)

# Dynamically handle device (CPU only)
device = -1  # Force CPU usage by setting device to -1

# Initialize the QA pipeline with the correct device
qa_pipeline = pipeline("question-answering", model=model, tokenizer=tokenizer, device=device)

# Load the knowledge base from JSON file
with open('knowledge_base.json', 'r') as f:
    knowledge_base = json.load(f)

# Load the expanded QA dataset
with open('expanded_qa_dataset.json', 'r') as f:
    expanded_qa_dataset = json.load(f)

# Load Sentence-BERT model for semantic search
embedding_model = SentenceTransformer('all-MiniLM-L6-v2')

# Function to create embeddings for the knowledge base content
def create_knowledge_base_embeddings(knowledge_base):
    embeddings = []
    for entry in knowledge_base:
        if 'title' in entry:
            content = entry['title'] + ' ' + ' '.join(
                [c.get('text', '') for c in entry.get('content', [])] +
                [
                    ' '.join(step['details']) if isinstance(step['details'], list) else step['details']
                    for c in entry.get('content', []) if 'steps' in c
                    for step in c['steps']
                ] +
                [
                    faq['question'] + ' ' + faq['answer']
                    for c in entry.get('content', []) if 'faq' in c
                    for faq in c['faq']
                ]
            )
            embeddings.append(embedding_model.encode(content, convert_to_tensor=True))
    return embeddings

# Create knowledge base embeddings
knowledge_base_embeddings = create_knowledge_base_embeddings(knowledge_base)

# Function to retrieve the best context using semantic similarity
def get_dynamic_context_semantic(question, knowledge_base, knowledge_base_embeddings, threshold=0.5):
    # Create embedding for the question
    question_embedding = embedding_model.encode(question, convert_to_tensor=True)

    # Calculate cosine similarity between the question and knowledge base entries
    cosine_scores = util.pytorch_cos_sim(question_embedding, torch.stack(knowledge_base_embeddings))

    # Get the index of the highest score (most similar context)
    best_match_idx = torch.argmax(cosine_scores).item()
    best_match_score = cosine_scores[0, best_match_idx].item()

    if best_match_score > threshold:  # Set a threshold for semantic similarity
        best_match_entry = knowledge_base[best_match_idx]

        # Check if FAQ section exists and prioritize FAQ answers
        for content_item in best_match_entry['content']:
            if 'faq' in content_item:
                for faq in content_item['faq']:
                    if faq['question'].lower() in question.lower():
                        return faq['answer']

        # If no FAQ is found, check for steps
        for content_item in best_match_entry['content']:
            if 'steps' in content_item:
                step_details = [step['details'] for step in content_item['steps']]
                return "\n".join(step_details)

        # Fallback to regular text
        for content_item in best_match_entry['content']:
            if 'text' in content_item:
                return content_item['text']
    
    return "Lo siento, no encontré una respuesta adecuada para tu pregunta."

# Check expanded QA dataset first for a direct answer
def get_answer_from_expanded_qa(question, expanded_qa_dataset):
    for item in expanded_qa_dataset:
        if item['question'].lower() == question.lower():
            return item['answer']
    return None

# Answer function for the Gradio app
def answer_question(question):
    # Check if the question matches any entry in the expanded QA dataset
    direct_answer = get_answer_from_expanded_qa(question, expanded_qa_dataset)
    if direct_answer:
        return direct_answer

    # If no direct answer found, use the knowledge base with semantic search
    context = get_dynamic_context_semantic(question, knowledge_base, knowledge_base_embeddings, threshold=0.55)
    return context

# Gradio interface setup
interface = gr.Interface(
    fn=answer_question,
    inputs="text",
    outputs="text",
    title="OCN Customer Support Chatbot",
    description="Ask questions and get answers from the OCN knowledge base."
)

# Launch the Gradio interface
interface.launch()