# ================================================================ # Wine Quality Research Assistant # Simple RAG Chatbot for Hugging Face Spaces # Gradio 6.x Compatible # ================================================================ import os import re import numpy as np import torch import gradio as gr from pypdf import PdfReader from sentence_transformers import SentenceTransformer from transformers import AutoTokenizer, AutoModelForSeq2SeqLM # ================================================================ # 1. CONFIGURATION # ================================================================ APP_TITLE = "🍷 Wine Quality Research Assistant" EMBEDDING_MODEL_ID = "sentence-transformers/all-MiniLM-L6-v2" LLM_MODEL_ID = "google/flan-t5-small" TOP_K = 3 CHUNK_SIZE = 700 CHUNK_OVERLAP = 100 DEVICE = "cuda" if torch.cuda.is_available() else "cpu" print("=" * 70) print("Starting Wine Quality Research Assistant") print("=" * 70) print(f"Device: {DEVICE}") print(f"Gradio: {gr.__version__}") # ================================================================ # 2. FIND PDF # ================================================================ def find_pdf(): pdf_files = [ file for file in os.listdir(".") if file.lower().endswith(".pdf") ] if not pdf_files: raise FileNotFoundError( "No PDF found. Upload the capstone PDF " "to the root folder of the Space." ) print("\nPDF files detected:") for pdf in pdf_files: print(" -", pdf) return pdf_files[0] PDF_FILE = find_pdf() print(f"\nUsing PDF: {PDF_FILE}") # ================================================================ # 3. CLEAN PDF TEXT # ================================================================ def clean_text(text): if not text: return "" text = text.replace("\x00", " ") text = re.sub( r"\s+", " ", text ) return text.strip() # ================================================================ # 4. READ PDF # ================================================================ def read_pdf(pdf_path): reader = PdfReader(pdf_path) pages = [] for page_number, page in enumerate( reader.pages, start=1 ): try: text = clean_text( page.extract_text() ) if text: pages.append( { "page": page_number, "text": text } ) except Exception as error: print( f"Could not read page " f"{page_number}: {error}" ) return pages pages = read_pdf(PDF_FILE) if not pages: raise ValueError( "PDF found, but no readable text " "could be extracted." ) print( f"Loaded {len(pages)} PDF pages." ) # ================================================================ # 5. CREATE CHUNKS # ================================================================ def create_chunks(pages): chunks = [] for page_data in pages: text = page_data["text"] page_number = page_data["page"] start = 0 while start < len(text): end = start + CHUNK_SIZE chunk_text = text[ start:end ].strip() if len(chunk_text) > 50: chunks.append( { "page": page_number, "text": chunk_text } ) start += ( CHUNK_SIZE - CHUNK_OVERLAP ) return chunks chunks = create_chunks(pages) print( f"Created {len(chunks)} " f"knowledge chunks." ) # ================================================================ # 6. LOAD EMBEDDING MODEL # ================================================================ print("Loading embedding model...") embedding_model = SentenceTransformer( EMBEDDING_MODEL_ID, device=DEVICE ) document_texts = [ chunk["text"] for chunk in chunks ] document_embeddings = embedding_model.encode( document_texts, convert_to_numpy=True, normalize_embeddings=True, show_progress_bar=False ) print("Embedding model ready.") # ================================================================ # 7. LOAD LANGUAGE MODEL # ================================================================ print( f"Loading language model: " f"{LLM_MODEL_ID}" ) tokenizer = AutoTokenizer.from_pretrained( LLM_MODEL_ID ) llm_model = AutoModelForSeq2SeqLM.from_pretrained( LLM_MODEL_ID ) llm_model.to(DEVICE) llm_model.eval() print("Language model ready.") # ================================================================ # 8. RETRIEVE RELEVANT PDF CONTENT # ================================================================ def retrieve_context(question): question_embedding = embedding_model.encode( [question], convert_to_numpy=True, normalize_embeddings=True )[0] similarity_scores = np.dot( document_embeddings, question_embedding ) top_indices = np.argsort( similarity_scores )[-TOP_K:][::-1] results = [] for index in top_indices: results.append( { "page": chunks[index]["page"], "text": chunks[index]["text"], "score": float( similarity_scores[ index ] ) } ) return results # ================================================================ # 9. BUILD CONTEXT # ================================================================ def build_context(results): sections = [] for item in results: sections.append( f""" PAGE {item['page']}: {item['text']} """ ) return "\n".join(sections) # ================================================================ # 10. FORMAT CHAT HISTORY # ================================================================ def format_history(history): if not history: return "" conversation = [] # Only retain recent conversation for item in history[-4:]: if isinstance(item, dict): role = item.get( "role", "" ) content = item.get( "content", "" ) conversation.append( f"{role}: {content}" ) return "\n".join(conversation) # ================================================================ # 11. GENERATE ANSWER # ================================================================ def generate_answer( question, context, history ): prompt = f""" You are a helpful academic research assistant. You answer questions about the research project: "Machine Learning-Based Prediction of Wine Quality Using Physicochemical Properties." Answer the user's question using ONLY the information provided in the DOCUMENT CONTEXT. RULES: 1. Answer the question directly. 2. Use clear and simple language. 3. Be academically accurate. 4. Do not invent information. 5. Do not invent research results. 6. Clearly distinguish proposed methodology from completed findings. 7. Use bullet points when helpful. 8. Keep the answer focused on the user's question. 9. If the document does not contain the answer, say: "This information is not specified in the current capstone synopsis." DOCUMENT CONTEXT: {context} RECENT CONVERSATION: {history} QUESTION: {question} ANSWER: """ inputs = tokenizer( prompt, return_tensors="pt", truncation=True, max_length=1800 ) inputs = { key: value.to(DEVICE) for key, value in inputs.items() } with torch.no_grad(): output = llm_model.generate( **inputs, max_new_tokens=220, num_beams=2, do_sample=False, repetition_penalty=1.15, no_repeat_ngram_size=3 ) answer = tokenizer.decode( output[0], skip_special_tokens=True ) return answer.strip() # ================================================================ # 12. CHAT FUNCTION # ================================================================ def chat(message, history): if not message.strip(): return ( "Please enter a question " "about the document." ) print("\n" + "=" * 60) print(f"QUESTION: {message}") # Retrieve relevant content results = retrieve_context( message ) print("Retrieved:") for item in results: print( f"Page {item['page']} | " f"Score: {item['score']:.3f}" ) context = build_context( results ) history_text = format_history( history ) answer = generate_answer( message, context, history_text ) # Unique page numbers source_pages = sorted( set( item["page"] for item in results ) ) page_text = ", ".join( str(page) for page in source_pages ) return f""" {answer} --- 📚 **Source page(s): {page_text}** """ # ================================================================ # 13. CSS # ================================================================ css = """ .gradio-container { max-width: 1000px !important; margin: auto !important; } """ # ================================================================ # 14. USER INTERFACE # ================================================================ with gr.Blocks( title=APP_TITLE ) as demo: # ------------------------------------------------------------ # SIMPLE HEADER # ------------------------------------------------------------ gr.Markdown( """ # 🍷 Wine Quality Research Assistant ### Ask questions about the QM640 Wine Quality Capstone This chatbot searches the uploaded capstone document and answers your questions using the relevant information from it. **Type your question below to begin.** """ ) # ------------------------------------------------------------ # SMALL STATUS LINE # ------------------------------------------------------------ gr.Markdown( f""" 📄 **Document loaded:** `{PDF_FILE}` 📑 **Pages:** {len(pages)}    |    🧩 **Knowledge chunks:** {len(chunks)} """ ) # ------------------------------------------------------------ # CHATBOT # ------------------------------------------------------------ chatbot_component = gr.Chatbot( height=480, show_label=False, placeholder=( "🍷 Ask me anything about " "the Wine Quality Capstone." ) ) # ------------------------------------------------------------ # CHAT INTERFACE # ------------------------------------------------------------ chat_interface = gr.ChatInterface( fn=chat, chatbot=chatbot_component, textbox=gr.Textbox( placeholder=( "Ask a question about " "the document..." ), show_label=False ), submit_btn="Ask", stop_btn=False, examples=[ "What is the main objective of the study?", "What dataset is used in this project?", "What are the research questions?", "Which machine learning models will be compared?", "How will model performance be evaluated?", "What is the business value of this research?" ] ) # ------------------------------------------------------------ # FOOTER # ------------------------------------------------------------ gr.Markdown( """ --- 🔎 **RAG-powered:** Your question → Relevant PDF sections → AI-generated answer *QM640 Data Analytics Capstone* """ ) # ================================================================ # 15. LAUNCH # ================================================================ print("\n" + "=" * 70) print("Wine Quality Research Assistant is ready.") print("=" * 70) if __name__ == "__main__": demo.queue() demo.launch( css=css )