import gradio as gr
import json
from sentence_transformers import SentenceTransformer
from sklearn.metrics.pairwise import cosine_similarity
import numpy as np
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

# 1. Embed model for semantic similarity
embedder = SentenceTransformer("paraphrase-MiniLM-L3-v2")  # Küçük ve hızlı

# 2. LLM model for question generation (TinyLLaMA)
llm_model_id = "TinyLlama/TinyLlama-1.1B-Chat-v1.0"
tokenizer = AutoTokenizer.from_pretrained(llm_model_id)
model = AutoModelForCausalLM.from_pretrained(llm_model_id)

# 3. Load memory-question data
with open("memory_questions.json", "r") as f:
    memory_data = json.load(f)

memory_texts = [item['description'] for item in memory_data]
memory_embeddings = embedder.encode(memory_texts)

# 4. En alakalı memory'yi bul + LLM ile soru üret
def generate_question(user_memory):
    # (a) En benzer memory'yi bul
    user_embedding = embedder.encode([user_memory])
    similarities = cosine_similarity(user_embedding, memory_embeddings)[0]
    best_match_index = np.argmax(similarities)
    matched_memory = memory_data[best_match_index]['description']

    # (b) Prompt hazırlığı
    prompt = f"<|system|>You are a helpful assistant who asks clear, meaningful questions based on short memories.<|user|>Memory: {matched_memory}\nGenerate a question that starts with What, Why, Who, When, or How.<|assistant|>"

    # (c) LLM ile generate et
    input_ids = tokenizer(prompt, return_tensors="pt").input_ids
    output = model.generate(input_ids, max_new_tokens=50, do_sample=False)
    result = tokenizer.decode(output[0], skip_special_tokens=True)

    # (d) Sadece son üretilen kısmı al
    if "<|assistant|>" in result:
        result = result.split("<|assistant|>")[-1].strip()

    return result

# 5. Gradio UI
iface = gr.Interface(
    fn=generate_question,
    inputs=gr.Textbox(label="Your Memory"),
    outputs=gr.Textbox(label="Generated Question"),
    title="MemoRease – LLM-Enhanced Question Generator",
    description="Enter a memory. We'll find a similar one and generate a clear, meaningful question using TinyLLaMA."
)

iface.launch()