import base64 import io import json import os import string from typing import Any, Dict, List import chromadb import google.generativeai as palm import pandas as pd import requests import streamlit as st from chromadb.utils.embedding_functions import SentenceTransformerEmbeddingFunction from langchain.text_splitter import ( RecursiveCharacterTextSplitter, SentenceTransformersTokenTextSplitter, ) from PIL import Image, ImageDraw, ImageFont from pypdf import PdfReader from transformers import pipeline from utils.cnn_transformer import * from utils.helpers import * # API Key (You should set this in your environment variables) api_key = st.secrets["PALM_API_KEY"] palm.configure(api_key=api_key) # Main function of the Streamlit app def main(): st.set_page_config(layout="wide") st.title("Generative AI Demo on Camera Input/Image/PDF 💻") # Dropdown for user to choose the input method input_method = st.sidebar.selectbox( "Choose input method:", ["Camera", "Upload Image", "Upload PDF"] ) image, uploaded_file = None, None if input_method == "Camera": # Streamlit widget to capture an image from the user's webcam image = st.sidebar.camera_input("Take a picture 📸") elif input_method == "Upload Image": # Create a file uploader in the sidebar image = st.sidebar.file_uploader("Upload a JPG image", type=["jpg"]) elif input_method == "Upload PDF": # File uploader widget uploaded_file = st.sidebar.file_uploader("Choose a PDF file", type="pdf") # Add instruction st.sidebar.markdown( """ # 🌟 How to Use the App 🌟 """ ) with st.sidebar: with st.expander("Show/Hide"): st.markdown( """ 1) **🌈 User Input Magic**: - 📸 **Camera Snap**: Tap to capture a moment with your device's camera. Say cheese! - 🖼️ **Image Upload Extravaganza**: Got a cool pic? Upload it from your computer and let the magic begin! - 📄 **PDF Adventure**: Use gen AI as ctrl+F to search information on any PDF, like opening a treasure chest of information! - 🧐 **YOLO Algorithm**: Wanna detect the object in the image? Use our object detection algorithm to see if the objects can be detected. 2) **🤖 AI Interaction Wonderland**: - 🌟 **Gemini's AI**: Google's Gemini AI is your companion, ready to dive deep into your uploads. - 🌐 **Chroma Database**: As you upload, we're crafting a colorful Chroma database in our secret lab, making your interaction even more awesome! 3) **💬 Chit-Chat with AI Post-Upload**: - 🌍 Once your content is up in the app, ask away! Any question, any time. - 💡 Light up the conversation with Gemini AI. It is like having a chat with a wise wizard from the digital realm! """ ) st.sidebar.markdown( """ Enjoy exploring and have fun! App URL [here](https://huggingface.co/spaces/eagle0504/IDP-Demo)!😄🎉 """ ) if image is not None: # Display the captured image with st.expander("Expand/collapse the uploaded image:"): st.image(image, caption="Captured Image", use_column_width=True) # Convert the image to PIL format and resize pil_image = Image.open(image) resized_image = resize_image(pil_image) # Convert the resized image to base64 image_base64 = convert_image_to_base64(resized_image) col1, col2 = st.columns(2) # OCR by API Call of AWS Textract via Post Method if input_method == "Upload Image": with col1: st.markdown( "# OCR (Optical Character Recognition) - computer vision technology to locate letters/numbers from image." ) st.success("Running textract!") url = "https://2tsig211e0.execute-api.us-east-1.amazonaws.com/my_textract" payload = {"image": image_base64} result_dict = post_request_and_parse_response(url, payload) output_data = extract_line_items(result_dict) df = pd.DataFrame(output_data) # Using an expander to hide the json with st.expander("Show/Hide Raw Json"): st.write(result_dict) # Using an expander to hide the table with st.expander("Show/Hide Table"): st.table(df) # Using an expander to hide the table st.success("Bounding boxes drawn!") with st.expander("Show/Hide Annotation"): try: image = Image.open(image) # Draw bounding boxes and labels image_with_boxes = draw_bounding_boxes_for_textract( image.copy(), result_dict ) # Display annotated image st.image( image_with_boxes, caption="Annotated Image", use_column_width=True, ) except: st.warning("Check textract output!") if api_key: with col2: # Make API call st.markdown( "# Gemini (Generative AI) - read the image content in a general form" ) st.success("Running Gemini!") with st.spinner("Wait for it..."): response = call_gemini_api(image_base64, api_key) with st.expander("Raw output from Gemini"): st.write(response) try: text_from_response = response["candidates"][0]["content"]["parts"][0][ "text" ] with st.spinner("Wait for it..."): st.write(text_from_response) except: st.warning("Please check the Gemini API as we do not have response from it.") # Display the response try: st.sidebar.success( "Check the box if you want to see a sample retrieved (we had a template of keys, in practice this depends on the stakeholders) information to download (only use this if this is a document-based task)! 👇" ) use_retrieval_tech = st.sidebar.checkbox( "Retrieve information!", value=False, ) if use_retrieval_tech: st.markdown( "# Information Retrieval - use Gemini to extract the values for the keys required by the stakeholders" ) with st.spinner("Processing csv to download..."): try: keys = ["First Name", "Last Name", "Policy Number"] values = [] for k in keys: updated_text_from_response = call_gemini_api( image_base64, api_key, prompt=f""" What is {k} in this document? Just answer the question directly with a word or two, don't say a complete sentence. If there is any special characters, rewrite it w """, ) value = updated_text_from_response["candidates"][0][ "content" ]["parts"][0]["text"] values.append(value) # Create dataframe to download sample_payload_output = pd.DataFrame( {"Key": keys, "Values": values} ) # Display table with st.expander("Inspect table (before download)"): st.table(sample_payload_output) # Convert DataFrame to CSV csv = sample_payload_output.to_csv(index=False) # To convert to CSV and encode for the download link csv = sample_payload_output.to_csv(index=False).encode( "utf-8" ) # Streamlit download button st.download_button( label="Download data as CSV", data=csv, file_name="data.csv", mime="text/csv", ) except: st.warning("Please verify document source.") st.sidebar.success( "Check the box if you want to chat with Gemini (do this if you want Gemini to answwer your questions)! 👇" ) use_gemini_to_chat = st.sidebar.checkbox( "Chat with Gemini (about the data)!", value=False, ) if use_gemini_to_chat: # Text input for the question input_prompt = st.text_input( "Type your question here:", ) # Display the entered question if input_prompt: updated_text_from_response = call_gemini_api( image_base64, api_key, prompt=input_prompt ) if updated_text_from_response is not None: # Do something with the text updated_ans = updated_text_from_response["candidates"][0][ "content" ]["parts"][0]["text"] with st.spinner("Wait for it..."): st.write(f"Gemini: {updated_ans}") else: st.warning("Check gemini's API.") except: st.write("No response from API.") else: st.write("API Key is not set. Please set the API Key.") # YOLO if image is not None: st.sidebar.success( "Check the following box to run YOLO algorithm if desired (only do this if the task at hand is an object detection task)! 👇" ) use_yolo = st.sidebar.checkbox("Use YOLO!", value=False) if use_yolo: yolo_option = st.selectbox( "Which YOLO algorithm would you like?", ("hustvl/yolos-small", "eagle0504/detr-finetuned-balloon-v2"), ) else: yolo_option = None # Load YOLO pipeline if yolo_option == "hustvl/yolos-small": yolo_pipe = pipeline("object-detection", model="hustvl/yolos-small") elif yolo_option == "eagle0504/detr-finetuned-balloon-v2": yolo_pipe = pipeline( "object-detection", model="eagle0504/detr-finetuned-balloon-v2" ) else: yolo_pipe = None if yolo_pipe is not None: # Process image with YOLO image = Image.open(image) with st.spinner("Wait for it..."): st.success("Running YOLO algorithm!") predictions = yolo_pipe(image) st.success("YOLO running successfully.") # Draw bounding boxes and labels image_with_boxes = draw_boxes(image.copy(), predictions) st.success("Bounding boxes drawn.") # Display annotated image st.image(image_with_boxes, caption="Annotated Image", use_column_width=True) # File uploader widget if uploaded_file is not None: # Select token size: st.sidebar.success("Note: 1 Token ~ 4 Characters.") token_size = st.sidebar.slider( "Select a token size (when we scrape the document)", 5, 150, 45 ) top_n_content = st.sidebar.slider( "Select top n content(s) you want to display as reference", 3, 30, 5 ) # To read file as bytes: bytes_data = uploaded_file.getvalue() st.success("Your PDF is uploaded successfully.") # Get the file name file_name = uploaded_file.name # Save the file temporarily with open(file_name, "wb") as f: f.write(uploaded_file.getbuffer()) # Display PDF # displayPDF(file_name) # Read file reader = PdfReader(file_name) pdf_texts = [p.extract_text().strip() for p in reader.pages] # Filter the empty strings pdf_texts = [text for text in pdf_texts if text] st.success("PDF extracted successfully.") # Split the texts character_splitter = RecursiveCharacterTextSplitter( separators=["\n\n", "\n", ". ", " ", ""], chunk_size=1000, chunk_overlap=0 ) character_split_texts = character_splitter.split_text("\n\n".join(pdf_texts)) st.success("Texts splitted successfully.") # Tokenize it st.warning("Start tokenzing ...") token_splitter = SentenceTransformersTokenTextSplitter( chunk_overlap=5, tokens_per_chunk=token_size ) token_split_texts = [] for text in character_split_texts: token_split_texts += token_splitter.split_text(text) st.success("Tokenized successfully.") # Generate a random number between 1 billion and 10 billion. random_number: int = np.random.randint(low=1e9, high=1e10) # Generate a random string consisting of 10 uppercase letters and digits. random_string: str = "".join( np.random.choice(list(string.ascii_uppercase + string.digits), size=20) ) # Combine the random number and random string into one identifier. combined_string: str = f"{random_number}{random_string}" # Add to vector database embedding_function = SentenceTransformerEmbeddingFunction() chroma_client = chromadb.Client() chroma_collection = chroma_client.create_collection( combined_string, embedding_function=embedding_function ) ids = [str(i) for i in range(len(token_split_texts))] chroma_collection.add(ids=ids, documents=token_split_texts) st.success("Vector database loaded successfully.") # User input query = st.text_input("Ask me anything!", "What is the document about?") results = chroma_collection.query(query_texts=[query], n_results=top_n_content) retrieved_documents = results["documents"][0] results_as_table = pd.DataFrame( { "ids": results["ids"][0], "documents": results["documents"][0], "distances": results["distances"][0], } ) # API of a foundation model output = rag(query=query, retrieved_documents=retrieved_documents) st.write(output) st.success( "Please see where the chatbot got the information from the document below.👇" ) with st.expander("Raw query outputs:"): st.write(results) with st.expander("Processed tabular form query outputs:"): st.table(results_as_table) if __name__ == "__main__": main()