import PyPDF2 from openpyxl import load_workbook from pptx import Presentation import gradio as gr import io from huggingface_hub import InferenceClient import re import zipfile import xml.etree.ElementTree as ET # Constants CHUNK_SIZE = 32000 MAX_NEW_TOKENS = 4096 # Initialize the Mistral chat model client = InferenceClient("mistralai/Mistral-Nemo-Instruct-2407") # --- Utility Functions --- def xml2text(xml): """Extracts text from XML data.""" text = u'' root = ET.fromstring(xml) for child in root.iter(): text += child.text + " " if child.text is not None else '' return text def clean_text(content): """Cleans text content based on the 'clean' parameter.""" content = content.replace('\n', ' ') content = content.replace('\r', ' ') content = content.replace('\t', ' ') content = re.sub(r'\s+', ' ', content) return content def split_content(content, chunk_size=CHUNK_SIZE): """Splits content into chunks of a specified size.""" chunks = [] for i in range(0, len(content), chunk_size): chunks.append(content[i:i + chunk_size]) return chunks # --- Document Reading Functions --- def extract_text_from_docx(docx_data, clean=True): """Extracts text from DOCX files.""" text = u'' zipf = zipfile.ZipFile(io.BytesIO(docx_data)) filelist = zipf.namelist() header_xmls = 'word/header[0-9]*.xml' for fname in filelist: if re.match(header_xmls, fname): text += xml2text(zipf.read(fname)) doc_xml = 'word/document.xml' text += xml2text(zipf.read(doc_xml)) footer_xmls = 'word/footer[0-9]*.xml' for fname in filelist: if re.match(footer_xmls, fname): text += xml2text(zipf.read(fname)) zipf.close() if clean: text = clean_text(text) return text, len(text) def extract_text_from_pptx(pptx_data, clean=True): """Extracts text from PPT files.""" text = u'' zipf = zipfile.ZipFile(io.BytesIO(pptx_data)) filelist = zipf.namelist() # Extract text from slide notes notes_xmls = 'ppt/notesSlides/notesSlide[0-9]*.xml' for fname in filelist: if re.match(notes_xmls, fname): text += xml2text(zipf.read(fname)) # Extract text from slide content (shapes and text boxes) slide_xmls = 'ppt/slides/slide[0-9]*.xml' for fname in filelist: if re.match(slide_xmls, fname): text += xml2text(zipf.read(fname)) zipf.close() if clean: text = clean_text(text) return text, len(text) def read_document(file, clean=True): """Reads content from various document formats.""" file_path = file.name file_extension = file_path.split('.')[-1].lower() with open(file_path, "rb") as f: file_content = f.read() if file_extension == 'pdf': try: pdf_reader = PyPDF2.PdfReader(io.BytesIO(file_content)) content = '' for page in range(len(pdf_reader.pages)): content += pdf_reader.pages[page].extract_text() if clean: content = clean_text(content) return content, len(content) except Exception as e: return f"Error reading PDF: {e}", 0 elif file_extension == 'xlsx': try: wb = load_workbook(io.BytesIO(file_content)) content = '' for sheet in wb.worksheets: for row in sheet.rows: for cell in row: if cell.value is not None: content += str(cell.value) + ' ' if clean: content = clean_text(content) return content, len(content) except Exception as e: return f"Error reading XLSX: {e}", 0 elif file_extension == 'pptx': try: return extract_text_from_pptx(file_content, clean) except Exception as e: return f"Error reading PPTX: {e}", 0 elif file_extension == 'doc' or file_extension == 'docx': try: return extract_text_from_docx(file_content, clean) except Exception as e: return f"Error reading DOC/DOCX: {e}", 0 else: try: content = file_content.decode('utf-8') if clean: content = clean_text(content) return content, len(content) except Exception as e: return f"Error reading file: {e}", 0 # --- Chat Functions --- def generate_mistral_response(message): """Generates a response from the Mistral API.""" stream = client.text_generation( message, max_new_tokens=MAX_NEW_TOKENS, stream=True, details=True, return_full_text=False ) output = "" for response in stream: if not response.token.text == "": output += response.token.text yield output def chat_document(file, question, clean=True): """Chats with a document using a single Mistral API call.""" content, length = read_document(file, clean) if length > CHUNK_SIZE: content = content[:CHUNK_SIZE] # Limit to max chunk size system_prompt = """ You are a helpful and informative assistant that can answer questions based on the content of documents. You will receive the content of a document and a question about it. Your task is to provide a concise and accurate answer to the question based solely on the provided document content. If the document does not contain enough information to answer the question, simply state that you cannot answer the question based on the provided information. """ message = f"""[INST] [SYSTEM] {system_prompt} Document Content: {content} Question: {question} Answer:""" yield from generate_mistral_response(message) def chat_document_v2(file, question, clean=True): """Chats with a document using chunk-based Mistral API calls and summarizes the answers.""" content, length = read_document(file, clean) chunks = split_content(content) system_prompt = """ You are a helpful and informative assistant that can answer questions based on the content of documents. You will receive the content of a document and a question about it. Your task is to provide a concise and accurate answer to the question based solely on the provided document content. If the document does not contain enough information to answer the question, simply state that you cannot answer the question based on the provided information. """ all_answers = [] for chunk in chunks: message = f"""[INST] [SYSTEM] {system_prompt} Document Content: {chunk[:CHUNK_SIZE]} Question: {question} Answer:""" response = "" for stream_response in generate_mistral_response(message): response = stream_response # Update with latest response all_answers.append(response) # Summarize all answers using Mistral summary_prompt = """ You are a helpful and informative assistant that can summarize multiple answers related to the same question. You will receive a list of answers to a question, and your task is to generate a concise and comprehensive summary that incorporates the key information from all the answers. Avoid repeating information unnecessarily and focus on providing the most relevant and accurate summary based on the provided answers. Answers: """ all_answers_str = "\n".join(all_answers) summary_message = f"""[INST] [SYSTEM] {summary_prompt} {all_answers_str[:30000]} Summary:""" yield from generate_mistral_response(summary_message) # --- Gradio Interface --- with gr.Blocks() as demo: with gr.Tabs(): with gr.TabItem("Document Reader"): iface1 = gr.Interface( fn=read_document, inputs=[ gr.File(label="Upload a Document"), gr.Checkbox(label="Clean Text", value=True), ], outputs=[ gr.Textbox(label="Document Content"), gr.Number(label="Document Length (characters)"), ], title="Document Reader", description="Upload a document (PDF, XLSX, PPTX, TXT, CSV, DOC, DOCX and Code or text file) to read its content." ) with gr.TabItem("Document Chat"): iface2 = gr.Interface( fn=chat_document, inputs=[ gr.File(label="Upload a Document"), gr.Textbox(label="Question"), gr.Checkbox(label="Clean and Compress Text", value=True), ], outputs=gr.Markdown(label="Answer"), title="Document Chat", description="Upload a document and ask questions about its content." ) with gr.TabItem("Document Chat V2"): iface3 = gr.Interface( fn=chat_document_v2, inputs=[ gr.File(label="Upload a Document"), gr.Textbox(label="Question"), gr.Checkbox(label="Clean Text", value=True), ], outputs=gr.Markdown(label="Answer"), title="Document Chat V2", description="Upload a document and ask questions about its content (using chunk-based approach)." ) demo.launch()