app.py

import os
import gradio as gr
import pdfplumber
import re

from langchain.docstore.document import Document
from langchain.vectorstores import FAISS
from langchain.embeddings.base import Embeddings
from sklearn.feature_extraction.text import TfidfVectorizer
from langchain.chains import RetrievalQA
from langchain.prompts import PromptTemplate
from langchain_openai import ChatOpenAI



os.environ["OPENAI_API_KEY"] = os.environ["OPENROUTER_API_KEY"]
os.environ["OPENAI_API_BASE"] = "https://openrouter.ai/api/v1"
os.environ["OPENAI_API_HEADERS"] = '{"HTTP-Referer":"https://huggingface.co", "X-Title":"PDF-RAG"}'

#Load and clean the policy PDF
def extract_clean_sections(file_path):
    with pdfplumber.open(file_path) as pdf:
        full_text = ""
        for page in pdf.pages:
            text = page.extract_text()
            if text:
                text = re.sub(r'Systems Campus.*?Lahore', '', text)
                text = re.sub(r'E-mail:.*?systemsltd\.com', '', text)
                full_text += text + "\n"

    pattern = r"(?<=\n)([A-Z][^\n]{3,50}):"
    parts = re.split(pattern, full_text)

    docs = []
    for i in range(1, len(parts), 2):
        title = parts[i].strip()
        content = parts[i + 1].strip()
        if len(content) > 20:
            docs.append(Document(page_content=f"{title}:\n{content}", metadata={"section": title}))
    return docs

#TF-IDF Embeddings
class TfidfEmbedding(Embeddings):
    def __init__(self):
        self.vectorizer = TfidfVectorizer()

    def fit(self, texts):
        self.vectorizer.fit(texts)

    def embed_documents(self, texts):
        return self.vectorizer.transform(texts).toarray()

    def embed_query(self, text):
        return self.vectorizer.transform([text]).toarray()[0]

#  Prompt Template 
TEMPLATE = """
You are a strict healthcare policy checker for Systems Ltd.

Always begin your answer clearly:
- Say "Yes, ..." if the claim is valid
- Say "No, ..." if the claim is not valid
- Say "Partially, ..." if it's conditionally allowed

Use the following policy information to support your answer.

{context}

Question: {question}
Answer:
"""
custom_prompt = PromptTemplate(template=TEMPLATE, input_variables=["context", "question"])


# Load the policy at startup
def initialize_policy():
    global qa_chain
    docs = extract_clean_sections("healthcare_policy.pdf")
    texts = [doc.page_content for doc in docs]
    embedder = TfidfEmbedding()
    embedder.fit(texts)
    vectordb = FAISS.from_texts(texts, embedder)
    retriever = vectordb.as_retriever()

    llm = ChatOpenAI(
        model="tngtech/deepseek-r1t2-chimera:free",
        base_url="https://openrouter.ai/api/v1",
        api_key=os.getenv("OPENAI_API_KEY"),
        default_headers={
            "HTTP-Referer": "https://huggingface.co",
            "X-Title": "PDF-RAG"
        },
        temperature=0.0
    )

    qa_chain = RetrievalQA.from_chain_type(
        llm=llm,
        chain_type="stuff",
        retriever=retriever,
        return_source_documents=False,
        chain_type_kwargs={"prompt": custom_prompt}
    )

# Run QA on user question
def ask_policy_question(question):
    if qa_chain is None:
        return "The policy is still loading. Please wait."
    try:
        return qa_chain.run(question)
    except Exception as e:
        return f"Error: {str(e)}"


#  Gradio Interface
qa_chain = None
status_text = "Loading..." 

with gr.Blocks() as demo:
    gr.Markdown("## SL HealthCare Claim Checker (RAG)")
    status_box = gr.Textbox(label="Status", value=status_text, interactive=False)

    with gr.Row():
        question = gr.Textbox(label="Enter your claim question")
        ask_btn = gr.Button("Ask")

    answer = gr.Textbox(label="Answer", lines=6)
    ask_btn.click(fn=ask_policy_question, inputs=question, outputs=answer)

    # Load the policy on startup
    def startup():
        global status_text
        initialize_policy()
        status_text = "Policy loaded. You may now ask questions."
        return status_text

    demo.load(fn=startup, outputs=status_box)

demo.launch()