# https://python.langchain.com/docs/tutorials/rag/
import gradio as gr
from langchain import hub
from langchain_chroma import Chroma
from langchain_core.output_parsers import StrOutputParser
from langchain_core.runnables import RunnablePassthrough
from langchain_mistralai import MistralAIEmbeddings
from langchain_community.embeddings import HuggingFaceInstructEmbeddings
from langchain_text_splitters import RecursiveCharacterTextSplitter
from langchain_mistralai import ChatMistralAI
from langchain_community.document_loaders import PyPDFLoader
import requests
from pathlib import Path
from langchain_community.document_loaders import WebBaseLoader, ArxivLoader
import bs4
from langchain_core.rate_limiters import InMemoryRateLimiter
from urllib.parse import urljoin


def initialize(arxivcode):  
    #loader = ArxivLoader(query=str(arxivcode),)
    #docs = loader.load()
    #retriever = ArxivRetriever(
    #    load_max_docs=2,
    #    get_full_documents=True,
    #)
    #docs = retriever.invoke(str(arxivcode))
    #for i in range(len(docs)): 
    #    docs[i].metadata['Published'] = str(docs[i].metadata['Published'])

    # Load, chunk and index the contents of the blog.
    url = ['https://arxiv.org/abs/%s' % arxivcode]
    loader = WebBaseLoader(url)
    docs = loader.load()


    # LLM model
    rate_limiter = InMemoryRateLimiter(
        requests_per_second=0.1,  # <-- MistralAI free. We can only make a request once every second
        check_every_n_seconds=0.01,  # Wake up every 100 ms to check whether allowed to make a request,
        max_bucket_size=10,  # Controls the maximum burst size.
    )    
    llm = ChatMistralAI(model="mistral-large-latest", rate_limiter=rate_limiter)
    
    # Embeddings
    embed_model = "sentence-transformers/multi-qa-distilbert-cos-v1"
    # embed_model = "nvidia/NV-Embed-v2"
    embeddings = HuggingFaceInstructEmbeddings(model_name=embed_model)
    # embeddings = MistralAIEmbeddings()

    def format_docs(docs):
        return "\n\n".join(doc.page_content for doc in docs)
    
    def RAG(llm, docs, embeddings):
    
        # Split text
        text_splitter = RecursiveCharacterTextSplitter(chunk_size=1000, chunk_overlap=200)
        splits = text_splitter.split_documents(docs)
    
        # Create vector store
        vectorstore = Chroma.from_documents(documents=splits, embedding=embeddings)
    
        # Retrieve and generate using the relevant snippets of the documents
        retriever = vectorstore.as_retriever()
    
        # Prompt basis example for RAG systems
        prompt = hub.pull("rlm/rag-prompt")
    
        # Create the chain
        rag_chain = (
            {"context": retriever | format_docs, "question": RunnablePassthrough()}
            | prompt
            | llm
            | StrOutputParser()
        )
    
        return rag_chain

    return RAG(llm, docs, embeddings)

def handle_prompt(message, history, arxivcode, rag_chain): 
    try:
        # Stream output
        out=""
        for chunk in rag_chain.stream(message):
            out += chunk
            yield out
    except:
        raise gr.Error("Requests rate limit exceeded")


greetingsmessage = "Hi, I'm your personal arXiv reader. Ask me questions about the arXiv paper above"

with gr.Blocks() as demo:     
  arxiv_code = gr.Textbox("", label="arxiv.number")
  rag_chain = initialize(arxiv_code)
    
  gr.ChatInterface(handle_prompt, type="messages", theme=gr.themes.Soft(), 
                          description=greetingsmessage, 
                   additional_inputs=[arxiv_code, rag_chain]
                  )
                          
if __name__=='__main__': 
    demo.launch()