app.py

import gradio as gr
from operator import itemgetter
import os
# import pandas as pd 

from langchain_community.vectorstores import FAISS

from langchain_community.llms.huggingface_pipeline import HuggingFacePipeline
from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline

## not needed since we are loading previously saved vector store from file and not reading pdf on the run
# from langchain_community.document_loaders import PyPDFLoader
# from langchain.text_splitter import RecursiveCharacterTextSplitter

from langchain_community.embeddings import HuggingFaceEmbeddings
from langchain.chains import RetrievalQA

## models tried
## TinyLlama/TinyLlama-1.1B-Chat-v1.0
## meta-llama/Meta-Llama-3-8B
## google/gemma-1.1-7b-it

HF_TOKEN = os.environ.get("HF_TOKEN", None)

model_id = "google/gemma-1.1-2b-it"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)

embeddings = HuggingFaceEmbeddings()

pipe = pipeline("text-generation", model = model, tokenizer = tokenizer, max_new_tokens = 200)
hf = HuggingFacePipeline(pipeline=pipe)

## commenting this code because now we are loading vectors directly and not parsing the pdf
# pdfLoader = PyPDFLoader("./LangchainPaper/RAGInputPaper.pdf")
# documents = pdfLoader.load()

# text_splitter = RecursiveCharacterTextSplitter(chunk_size = 512, chunk_overlap = 30)
# docs = text_splitter.split_documents(documents)

## creating vector embeddings during run using FAISS
# vectorstore = FAISS.from_documents(
# docs, embedding=embeddings
# )
# retriever = vectorstore.as_retriever()

## loading previously saved vector embeddings from local space 

vectorstore = FAISS.load_local("./fi_LangchainPaper", embeddings, allow_dangerous_deserialization = True)
retriever = vectorstore.as_retriever()

qa = RetrievalQA.from_chain_type(
llm = hf, chain_type = "stuff", retriever = retriever, return_source_documents = False)

# queries=pd.read_csv('./interactions/queries.csv')

def greet(Question):
    answer = qa({"query": Question})

    pa = [a.split("Helpful Answer: ") for a in answer.get('result').split('\n') if "Helpful Answer" in a]
    
    # new=pd.DataFrame.from_dict({'query':Question,'response':pa[0][-1]},orient='index')
    # queries.append(new)
    # queries.to_csv('./interactions/queries.csv')
    
    return pa[0][-1]

if __name__ == "__main__":

    title = "RAG with LLMs"

    description = """
    <img src="https://superagi.com/wp-content/uploads/2023/10/Introduction-to-RAGA-Retrieval-Augmented-Generation-and-Actions-1200x600.png.webp" width=100%>
    <br>
    Demo using Vector store-backed retriever. This space demonstrate application of RAG on a small model and its effectiveness, I used small model because of the space constraint. The current space runs on mere <b>2GB of RAM</b>, hence there is some delay in generating output. Test this to your hearts content and let me know your thoughts, I will keep updating this space with tiny improvements on architecture and design
    <ul>
        <li>model: TinyLlama/TinyLlama-1.1B-Chat-v1.0</li>
        <li></li>
        
        <li>update1: This space now does not create a faiss index on build, it uses a locally saved faiss index</li>
        <li>update2: This space now uses google/gemma-1.1-2b-it model to generate output, reduces the response time to 1/3rd</li>
    </ul>
    """
    article = """<p style='text-align: center'> 
    <ul>You can ask questions like - 
        <li>What is langchain framework?</li>
        <li>What is Action Agent?</li>
        <li>What are forms of memory implementation in langchain</li>
        <li>What is question answering from documents</li>
    </ul>
    Go through this paper here to find more about langchain and then test how this solution performs. <a href='https://www.researchgate.net/publication/372669736_Creating_Large_Language_Model_Applications_Utilizing_LangChain_A_Primer_on_Developing_LLM_Apps_Fast' target='_blank'>This paper is the data source for this solution</a>
    Have you already used RAG? feel free to suggest improvements
    Feel excited about the implementation? You know where to find me!
    I would love to connect and have a chat.
    </p>"""

    iface = gr.Interface(fn = greet, inputs = "text", outputs = gr.Textbox(lines = 5, label = "Answer"), title = title,
    description = description,
    article = article,)
    iface.launch(share = True)