app.py

import gradio as gr
import requests
from pdfminer.high_level import extract_text
from langchain_community.vectorstores import Chroma
from langchain_huggingface import HuggingFaceEmbeddings, ChatHuggingFace
from langchain_core.runnables import RunnablePassthrough
from io import BytesIO
from langchain_core.output_parsers import StrOutputParser
from langchain_core.documents import Document
from langchain_core.prompts import ChatPromptTemplate
from langchain.text_splitter import CharacterTextSplitter
from huggingface_hub import InferenceClient

client = InferenceClient("HuggingFaceH4/zephyr-7b-beta")

def extract_pdf_text(url: str) -> str:
    response = requests.get(url)
    pdf_file = BytesIO(response.content)
    text = extract_text(pdf_file)
    return text

pdf_url = "https://huggingface.co/spaces/disLodge/Call_model/raw/main/temp.pdf"
text = extract_pdf_text(pdf_url)
docs_list = [Document(page_content=text)]

text_splitter = CharacterTextSplitter.from_tiktoken_encoder(chunk_size=7500, chunk_overlap=100)
docs_splits = text_splitter.split_documents(docs_list)

embeddings = HuggingFaceEmbeddings(model_name="sentence-transformers/all-MiniLM-L6-v2")
vectorstore = Chroma.from_documents(
    documents=docs_splits,
    collection_name="rag-chroma",
    embedding=embeddings,
)
retriever = vectorstore.as_retriever()

llm = ChatHuggingFace(
    huggingfacehub_api_token=None,
    model_id="HuggingFaceH4/zephyr-7b-beta",
    interference_client=client,
)

# Before RAG chain
before_rag_template = "What is {topic}"
before_rag_prompt = ChatPromptTemplate.from_template(before_rag_template)
before_rag_chain = before_rag_prompt | llm | StrOutputParser()

# After RAG chain
after_rag_template = """You are a {role}. Summarize the following content for yourself and speak in terms of first person.
Only include content relevant to that role like a resume summary.

Context:
{context}

Question: Give a one paragraph summary of the key skills a {role} can have from this document.
"""
after_rag_prompt = ChatPromptTemplate.from_template(after_rag_template)

def format_query(input_dict):
    return f"Give a one paragraph summary of the key skills a {input_dict['role']} can have from this document."

after_rag_chain = (
    {
        "context": format_query | retriever,
        "role": lambda x: x["role"],
    }
    | after_rag_prompt
    | llm
    | StrOutputParser()
)

def process_query(role, system_message, max_tokens, temperature, top_p):
    client.max_tokens = max_tokens
    client.temperature = temperature
    client.top_p = top_p

    # Before RAG
    before_rag_result = before_rag_chain.invoke({"topic": "Hugging Face"})

    # After RAG
    after_rag_result = after_rag_chain.invoke({"role": role})

    return f"**Before RAG**\n{before_rag_result}\n\n**After RAG**\n{after_rag_result}"


with gr.Blocks() as demo:
    gr.Markdown("## Zephyr Chatbot Controls")

    role_dropdown = gr.Dropdown(choices=["SDE", "BA"], label="Select Role", value="SDE")

    system_message = gr.Textbox(value="You are a friendly chatbot.", label="System message")
    max_tokens = gr.Slider(1, 2048, value=512, label="Max tokens")
    temperature = gr.Slider(0.1, 4.0, value=0.7, label="Temperature", step=0.1)
    top_p = gr.Slider(0.1, 1.0, value=0.95, label="Top-p", step=0.05)

    output = gr.Textbox(label="Output", lines=20)

    submit_btn = gr.Button("Submit")
    clear_btn = gr.Button("Clear")

    submit_btn.click(
        fn=process_query,
        inputs=[role_dropdown, system_message, max_tokens, temperature, top_p],
        outputs=output
    )

    clear_btn.click(
        fn=lambda: ("", gr.Info("Chat cleared!")),
        outputs=[output]
    )

if __name__ == "__main__":
    demo.launch()