app.py

# 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
from langchain_community.retrievers import ArxivRetriever
import bs4
from langchain_core.rate_limiters import InMemoryRateLimiter
from urllib.parse import urljoin

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.
)

# get data
urlsfile = open("urls.txt")
urls = urlsfile.readlines()
urls = [url.replace("\n","") for url in urls]
urlsfile.close()

# Load, chunk and index the contents of the blog.
loader = WebBaseLoader(urls)
docs = loader.load()

# load arxiv papers 
arxivfile = open("arxiv.txt")
arxivs = arxivfile.readlines()
arxivs = [arxiv.replace("\n","") for arxiv in arxivs]
arxivfile.close()

retriever = ArxivRetriever(
    load_max_docs=2,
    get_ful_documents=True,
)

for arxiv in arxivs: 
    doc = retriever.invoke(arxiv) 
    doc[0].metadata['Published'] = str(doc[0].metadata['Published'])
    docs.append(doc[0])
    

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

# LLM model
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()

# RAG chain
rag_chain = RAG(llm, docs, embeddings)

def handle_prompt(message, history):
    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 ChangBot, a chat bot here to assist you with any question related to Chang's research. I'm in pre-alpha stage, so please be patient."
example_questions = [
                    "Tell me more about SimBIG",
                    "How can you constrain neutrino mass with galaxies?", 
                    "What is the DESI BGS?",    
                    "What is SEDflow?", 
                    "What are normalizing flows?"
                     ]

demo = gr.ChatInterface(handle_prompt, type="messages", title="ChangBot", examples=example_questions, theme=gr.themes.Soft(), description=greetingsmessage)#, chatbot=chatbot)

demo.launch()