import gradio as gr
from datasets import load_dataset
from sentence_transformers import SentenceTransformer, CrossEncoder
import faiss
import torch
import numpy as np
from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline
import os
import spaces

# Login automatically if HF_TOKEN is present
hf_token = os.getenv("HF_TOKEN")
if hf_token:
    from huggingface_hub import login
    login(token=hf_token)


# Load corpus
print("Loading dataset...")
dataset = load_dataset("rag-datasets/rag-mini-wikipedia", "text-corpus")
# corpus = [item for item in dataset["passages"]]

# Always clean + use this corpus consistently
corpus = []
for item in dataset["passages"]:
    text = str(item).strip()
    if text:
        corpus.append(text)

# Embedding model
print("Encoding corpus...")
embedder = SentenceTransformer("all-MiniLM-L6-v2")
corpus_embeddings = embedder.encode(corpus, convert_to_tensor=True, device='cpu')
corpus_embeddings_np = corpus_embeddings.numpy()

# FAISS index
index = faiss.IndexFlatL2(corpus_embeddings_np.shape[1])
index.add(corpus_embeddings_np)

# Reranker model
# reranker = CrossEncoder("cross-encoder/ms-marco-MiniLM-L-6-v2")

# Generator (choose one: local HF model or OpenAI)
tokenizer = AutoTokenizer.from_pretrained("mistralai/Mistral-7B-Instruct-v0.3")
model = AutoModelForCausalLM.from_pretrained("mistralai/Mistral-7B-Instruct-v0.3", torch_dtype=torch.float16)
generator = pipeline("text-generation", model=model, tokenizer=tokenizer, max_new_tokens=150)

@spaces.GPU
def rag_pipeline(query):
    # Embed query
    query_embedding = embedder.encode([query], convert_to_tensor=True, device='cpu').numpy()

    # Retrieve top-k from FAISS
    D, I = index.search(query_embedding, k=5)
    retrieved_docs = [corpus[idx] for idx in I[0]]
    
    print("Retrieved indices:", I[0])
    print("Retrieved docs:")
    for doc in retrieved_docs:
        print("-", repr(doc))

    # # Rerank
    # rerank_pairs = [[str(query), str(doc)] for doc in retrieved_docs]
    # scores = reranker.predict(rerank_pairs)
    # reranked_docs = [doc for _, doc in sorted(zip(scores, retrieved_docs), reverse=True)]

    # Combine for context
    context = "\n\n".join(retrieved_docs[:2])
    prompt = f"""Answer the following question using the provided context.\n\nContext:\n{context}\n\nQuestion: {query}\nAnswer:"""

    # Generate
    response = generator(prompt)[0]["generated_text"]
    return response.split("Answer:")[-1].strip()

# Gradio UI
iface = gr.Interface(fn=rag_pipeline,
                     inputs=gr.Textbox(lines=2, placeholder="How fast is a penguin?"),
                     outputs="text",
                     title="Mini RAG Wikipedia Demo",
                     description="Retrieval-Augmented Generation on a small Wikipedia subset.")

iface.launch()