Heraali/cschatbot

1import torch2from transformers import AutoModelForQuestionAnswering, AutoTokenizer, pipeline3from sentence_transformers import SentenceTransformer, util4import gradio as gr5import json6 7# Load pre-trained BERT QA model and tokenizer from Hugging Face model hub8model_name = "bert-large-uncased-whole-word-masking-finetuned-squad"9model = AutoModelForQuestionAnswering.from_pretrained(model_name)10tokenizer = AutoTokenizer.from_pretrained(model_name)11 12# Dynamically handle device (CPU only)13device = -1  # Force CPU usage by setting device to -114 15# Initialize the QA pipeline with the correct device16qa_pipeline = pipeline("question-answering", model=model, tokenizer=tokenizer, device=device, max_answer_len=500)17 18# Load the knowledge base from JSON file19with open('knowledge_base.json', 'r') as f:20    knowledge_base = json.load(f)21 22# Load Sentence-BERT model for semantic search23embedding_model = SentenceTransformer('all-MiniLM-L6-v2')24 25# Function to create embeddings for the knowledge base content26def create_knowledge_base_embeddings(knowledge_base):27    embeddings = []28    for entry in knowledge_base:29        if 'title' in entry:30            # Prepare content, handle both text, steps, and faq31            content = entry['title'] + ' ' + ' '.join(32                [c.get('text', '') for c in entry.get('content', [])] +33                [34                    ' '.join(step['details']) if isinstance(step['details'], list) else step['details']35                    for c in entry.get('content', []) if 'steps' in c36                    for step in c['steps']37                ] +38                [39                    faq['question'] + ' ' + faq['answer']40                    for c in entry.get('content', []) if 'faq' in c41                    for faq in c['faq']42                ]43            )44            embeddings.append(embedding_model.encode(content, convert_to_tensor=True))45    return embeddings46 47# Create knowledge base embeddings48knowledge_base_embeddings = create_knowledge_base_embeddings(knowledge_base)49 50# Function to retrieve the best context using semantic similarity51def get_dynamic_context_semantic(question, knowledge_base, knowledge_base_embeddings):52    # Create embedding for the question53    question_embedding = embedding_model.encode(question, convert_to_tensor=True)54 55    # Calculate cosine similarity between the question and knowledge base entries56    cosine_scores = util.pytorch_cos_sim(question_embedding, torch.stack(knowledge_base_embeddings))57 58    # Get the index of the highest score (most similar context)59    best_match_idx = torch.argmax(cosine_scores).item()60    best_match_score = cosine_scores[0, best_match_idx].item()61 62    if best_match_score > 0.5:  # Set a threshold for semantic similarity63        best_match_entry = knowledge_base[best_match_idx]64        print(f"Best match: {best_match_entry['title']} with score {best_match_score}")65 66        # Check if FAQ section exists and prioritize FAQ answers67        for content_item in best_match_entry['content']:68            if 'faq' in content_item:  # Look for FAQ in content69                for faq in content_item['faq']:70                    if faq['question'].lower() in question.lower():  # Match the FAQ question71                        return faq['answer']  # Return the matching FAQ answer72 73        # If no FAQ is found, check for steps74        for content_item in best_match_entry['content']:75            if 'steps' in content_item:76                step_details = [step['details'] for step in content_item['steps']]77                return "\n".join(step_details)  # Return steps if available78 79        # Fallback to regular text (but exclude objective/purpose sections)80        for content_item in best_match_entry['content']:81            if 'text' in content_item and "Objetivo" not in content_item['text']:  # Skip metadata82                return content_item['text']  # Fallback to text83 84    return "No relevant context found."85 86# Answer function for the Gradio app87def answer_question(question):88    # Retrieve context using semantic search function89    context = get_dynamic_context_semantic(question, knowledge_base, knowledge_base_embeddings)90    return context91 92# Gradio interface setup93interface = gr.Interface(94    fn=answer_question,95    inputs="text",96    outputs="text",97    title="OCN Customer Support Chatbot",98    description="Ask questions and get answers from the OCN knowledge base."99)100 101# Launch the Gradio interface102interface.launch(share=True)103