RAG_Model / app.py
Panda1902's picture
Update app.py
02430ea verified
Raw History Blame
3.29 kB
import gradio as gr
import pandas as pd
import os
from langchain_community.document_loaders import PyPDFLoader
from langchain_text_splitters import RecursiveCharacterTextSplitter
from langchain_huggingface import HuggingFaceEmbeddings
from langchain_community.vectorstores import Chroma
from langchain.chains import RetrievalQA
from transformers import pipeline
from langchain_community.llms import HuggingFacePipeline
print("Hi")
# =========================
# LOAD PDF GUIDELINES
# =========================
PDF_PATH = "Lemonade_Guidlines.pdf"
loader = PyPDFLoader(PDF_PATH)
docs = loader.load()
# =========================
# CHUNKING
# =========================
text_splitter = RecursiveCharacterTextSplitter(
chunk_size=500,
chunk_overlap=100
)
chunks = text_splitter.split_documents(docs)
# =========================
# EMBEDDINGS
# =========================
embedding_model = HuggingFaceEmbeddings(
model_name="sentence-transformers/all-MiniLM-L6-v2"
)
# =========================
# VECTOR DB (FASTER INIT)
# =========================
vectorstore = Chroma.from_documents(
documents=chunks,
embedding=embedding_model
)
retriever = vectorstore.as_retriever(search_kwargs={"k": 2})
# =========================
# LIGHTWEIGHT LLM (IMPORTANT FIX)
# =========================
pipe = pipeline(
"text-generation",
model="TinyLlama/TinyLlama-1.1B-Chat-v1.0", # 🔥 small + fast
max_new_tokens=120,
temperature=0.2
)
llm = HuggingFacePipeline(pipeline=pipe)
# =========================
# QA CHAIN
# =========================
qa_chain = RetrievalQA.from_chain_type(
llm=llm,
retriever=retriever
)
# =========================
# VIOLATION CHECK
# =========================
def check_violation(row):
prompt = f"""
You are a strict legal billing auditor.
Check if the below entry violates guidelines.
Narrative: {row['Narrative (Full Text)']}
Task Code: {row['Task Code']}
Timekeeper: {row['Timekeeper Name']}
Hours: {row['Units (Hrs / Qty)']}
Respond strictly:
Violation: Yes or No
Reason: <one line>
"""
try:
result = qa_chain.invoke({"query": prompt})
return result["result"]
except Exception as e:
return f"Error: {str(e)}"
# =========================
# FILE PROCESSING
# =========================
def process_file(file):
df = pd.read_excel(file, engine="openpyxl")
# Duplicate detection
df["Duplicate Count"] = df.duplicated(
subset=["Narrative (Full Text)", "Units (Hrs / Qty)", "Rate ($/Unit)"],
keep=False
).astype(int)
results = []
for i, row in df.iterrows():
# 🔥 LIMIT rows to avoid timeout (VERY IMPORTANT)
if i > 30:
results.append("Skipped (limit reached)")
continue
res = check_violation(row)
results.append(res)
df["Violation Analysis"] = results
output_path = "audited_output.xlsx"
df.to_excel(output_path, index=False)
return output_path
# =========================
# UI
# =========================
interface = gr.Interface(
fn=process_file,
inputs=gr.File(label="Upload Billing Excel"),
outputs=gr.File(label="Download Result"),
title="Legal Billing Auditor",
description="AI-powered billing audit using RAG"
)
interface.launch()