Spaces:
Runtime error
Runtime error
Create app.py
Browse files
app.py
ADDED
|
@@ -0,0 +1,148 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import gradio as gr
|
| 2 |
+
import pandas as pd
|
| 3 |
+
import os
|
| 4 |
+
|
| 5 |
+
from langchain_community.document_loaders import PyPDFLoader
|
| 6 |
+
from langchain_text_splitters import RecursiveCharacterTextSplitter
|
| 7 |
+
from langchain_huggingface import HuggingFaceEmbeddings
|
| 8 |
+
from langchain_community.vectorstores import Chroma
|
| 9 |
+
from langchain.chains import RetrievalQA
|
| 10 |
+
|
| 11 |
+
from transformers import pipeline
|
| 12 |
+
from langchain_community.llms import HuggingFacePipeline
|
| 13 |
+
|
| 14 |
+
# =========================
|
| 15 |
+
# LOAD PDF GUIDELINES
|
| 16 |
+
# =========================
|
| 17 |
+
|
| 18 |
+
PDF_PATH = "Lemonade_Guidlines.pdf"
|
| 19 |
+
|
| 20 |
+
loader = PyPDFLoader(PDF_PATH)
|
| 21 |
+
docs = loader.load()
|
| 22 |
+
|
| 23 |
+
# =========================
|
| 24 |
+
# CHUNKING
|
| 25 |
+
# =========================
|
| 26 |
+
|
| 27 |
+
text_splitter = RecursiveCharacterTextSplitter(
|
| 28 |
+
chunk_size=500,
|
| 29 |
+
chunk_overlap=100
|
| 30 |
+
)
|
| 31 |
+
|
| 32 |
+
chunks = text_splitter.split_documents(docs)
|
| 33 |
+
|
| 34 |
+
# =========================
|
| 35 |
+
# EMBEDDINGS
|
| 36 |
+
# =========================
|
| 37 |
+
|
| 38 |
+
embedding_model = HuggingFaceEmbeddings(
|
| 39 |
+
model_name="sentence-transformers/all-MiniLM-L6-v2"
|
| 40 |
+
)
|
| 41 |
+
|
| 42 |
+
# =========================
|
| 43 |
+
# VECTOR DB (FASTER INIT)
|
| 44 |
+
# =========================
|
| 45 |
+
|
| 46 |
+
vectorstore = Chroma.from_documents(
|
| 47 |
+
documents=chunks,
|
| 48 |
+
embedding=embedding_model
|
| 49 |
+
)
|
| 50 |
+
|
| 51 |
+
retriever = vectorstore.as_retriever(search_kwargs={"k": 2})
|
| 52 |
+
|
| 53 |
+
# =========================
|
| 54 |
+
# LIGHTWEIGHT LLM (IMPORTANT FIX)
|
| 55 |
+
# =========================
|
| 56 |
+
|
| 57 |
+
pipe = pipeline(
|
| 58 |
+
"text-generation",
|
| 59 |
+
model="TinyLlama/TinyLlama-1.1B-Chat-v1.0", # 🔥 small + fast
|
| 60 |
+
max_new_tokens=120,
|
| 61 |
+
temperature=0.2
|
| 62 |
+
)
|
| 63 |
+
|
| 64 |
+
llm = HuggingFacePipeline(pipeline=pipe)
|
| 65 |
+
|
| 66 |
+
# =========================
|
| 67 |
+
# QA CHAIN
|
| 68 |
+
# =========================
|
| 69 |
+
|
| 70 |
+
qa_chain = RetrievalQA.from_chain_type(
|
| 71 |
+
llm=llm,
|
| 72 |
+
retriever=retriever
|
| 73 |
+
)
|
| 74 |
+
|
| 75 |
+
# =========================
|
| 76 |
+
# VIOLATION CHECK
|
| 77 |
+
# =========================
|
| 78 |
+
|
| 79 |
+
def check_violation(row):
|
| 80 |
+
|
| 81 |
+
prompt = f"""
|
| 82 |
+
You are a strict legal billing auditor.
|
| 83 |
+
|
| 84 |
+
Check if the below entry violates guidelines.
|
| 85 |
+
|
| 86 |
+
Narrative: {row['Narrative (Full Text)']}
|
| 87 |
+
Task Code: {row['Task Code']}
|
| 88 |
+
Timekeeper: {row['Timekeeper Name']}
|
| 89 |
+
Hours: {row['Units (Hrs / Qty)']}
|
| 90 |
+
|
| 91 |
+
Respond strictly:
|
| 92 |
+
|
| 93 |
+
Violation: Yes or No
|
| 94 |
+
Reason: <one line>
|
| 95 |
+
"""
|
| 96 |
+
|
| 97 |
+
try:
|
| 98 |
+
result = qa_chain.invoke({"query": prompt})
|
| 99 |
+
return result["result"]
|
| 100 |
+
except Exception as e:
|
| 101 |
+
return f"Error: {str(e)}"
|
| 102 |
+
|
| 103 |
+
# =========================
|
| 104 |
+
# FILE PROCESSING
|
| 105 |
+
# =========================
|
| 106 |
+
|
| 107 |
+
def process_file(file):
|
| 108 |
+
|
| 109 |
+
df = pd.read_excel(file, engine="openpyxl")
|
| 110 |
+
|
| 111 |
+
# Duplicate detection
|
| 112 |
+
df["Duplicate Count"] = df.duplicated(
|
| 113 |
+
subset=["Narrative (Full Text)", "Units (Hrs / Qty)", "Rate ($/Unit)"],
|
| 114 |
+
keep=False
|
| 115 |
+
).astype(int)
|
| 116 |
+
|
| 117 |
+
results = []
|
| 118 |
+
|
| 119 |
+
for i, row in df.iterrows():
|
| 120 |
+
|
| 121 |
+
# 🔥 LIMIT rows to avoid timeout (VERY IMPORTANT)
|
| 122 |
+
if i > 30:
|
| 123 |
+
results.append("Skipped (limit reached)")
|
| 124 |
+
continue
|
| 125 |
+
|
| 126 |
+
res = check_violation(row)
|
| 127 |
+
results.append(res)
|
| 128 |
+
|
| 129 |
+
df["Violation Analysis"] = results
|
| 130 |
+
|
| 131 |
+
output_path = "audited_output.xlsx"
|
| 132 |
+
df.to_excel(output_path, index=False)
|
| 133 |
+
|
| 134 |
+
return output_path
|
| 135 |
+
|
| 136 |
+
# =========================
|
| 137 |
+
# UI
|
| 138 |
+
# =========================
|
| 139 |
+
|
| 140 |
+
interface = gr.Interface(
|
| 141 |
+
fn=process_file,
|
| 142 |
+
inputs=gr.File(label="Upload Billing Excel"),
|
| 143 |
+
outputs=gr.File(label="Download Result"),
|
| 144 |
+
title="Legal Billing Auditor",
|
| 145 |
+
description="AI-powered billing audit using RAG"
|
| 146 |
+
)
|
| 147 |
+
|
| 148 |
+
interface.launch()
|