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: """ 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()