embedding.py

from PyPDF2 import PdfReader
import requests
import json
import os
import concurrent.futures
import random
from langchain_google_genai import ChatGoogleGenerativeAI
from langchain_community.document_loaders import WebBaseLoader
from langchain_community.document_loaders import PyPDFLoader
from langchain.text_splitter import RecursiveCharacterTextSplitter
import google.generativeai as genai
from langchain_core.messages import HumanMessage
from io import BytesIO
import numpy as np
import re
import torch
from transformers import AutoTokenizer, AutoModel
import numpy as np
import onnxruntime as ort
# import torch._dynamo
import time
# torch._dynamo.config.suppress_errors = True

from search import search_images

gemini = ChatGoogleGenerativeAI(model="gemini-1.0-pro-001",google_api_key='AIzaSyCo-TeDp0Ou--UwhlTgMwCoTEZxg6-v7wA',temperature = 0.1)
gemini1 = ChatGoogleGenerativeAI(model="gemini-1.0-pro-001",google_api_key='AIzaSyAtnUk8QKSUoJd3uOBpmeBNN-t8WXBt0zI',temperature = 0.1)
gemini2 = ChatGoogleGenerativeAI(model="gemini-1.0-pro-001",google_api_key='AIzaSyBzbZQBffHFK3N-gWnhDDNbQ9yZnZtaS2E',temperature = 0.1)
gemini3 = ChatGoogleGenerativeAI(model="gemini-1.0-pro-001",google_api_key='AIzaSyBNN4VDMAOB2gSZha6HjsTuH71PVV69FLM',temperature = 0.1)

vision = ChatGoogleGenerativeAI(model="gemini-1.5-flash",google_api_key='AIzaSyCo-TeDp0Ou--UwhlTgMwCoTEZxg6-v7wA',temperature = 0.1)
vision1 = ChatGoogleGenerativeAI(model="gemini-1.5-flash",google_api_key='AIzaSyAtnUk8QKSUoJd3uOBpmeBNN-t8WXBt0zI',temperature = 0.1)
vision2 = ChatGoogleGenerativeAI(model="gemini-1.5-flash",google_api_key='AIzaSyBzbZQBffHFK3N-gWnhDDNbQ9yZnZtaS2E',temperature = 0.1)
vision3 = ChatGoogleGenerativeAI(model="gemini-1.5-flash",google_api_key='AIzaSyBNN4VDMAOB2gSZha6HjsTuH71PVV69FLM',temperature = 0.1)

tokenizer = AutoTokenizer.from_pretrained('dwzhu/e5-base-4k',trust_remote_code = True)
# model = AutoModel.from_pretrained('dwzhu/e5-base-4k',trust_remote_code = True)
model_path = "model_opt2_QInt8.onnx"

session = ort.InferenceSession(model_path)
# model = torch.compile(model)
# model.to('cpu')  # Ensure the model is on the CPU

from transformers import PreTrainedTokenizerFast

class TokenBasedTextSplitter:
    def __init__(self, tokenizer_path='tokenizer.json', chunk_size=2000, chunk_overlap=50):
        self.tokenizer = PreTrainedTokenizerFast(tokenizer_file=tokenizer_path)
        self.chunk_size = chunk_size
        self.chunk_overlap = chunk_overlap

    def split_text(self, text):
        tokens = self.tokenizer.tokenize(text)
        chunks = []

        for i in range(0, len(tokens), self.chunk_size - self.chunk_overlap):
            chunk = tokens[i:i + self.chunk_size]
            chunks.append(self.tokenizer.convert_tokens_to_string(chunk))
        
        return chunks
    


genai.configure(api_key="AIzaSyAtnUk8QKSUoJd3uOBpmeBNN-t8WXBt0zI")

def pdf_extractor(link):
    text = ''

    try:
        # Fetch the PDF file from the URL
        response = requests.get(link)
        response.raise_for_status()  # Raise an error for bad status codes

        # Use BytesIO to handle the PDF content in memory
        pdf_file = BytesIO(response.content)

        # Load the PDF file
        reader = PdfReader(pdf_file)
        for page in reader.pages:
            text += page.extract_text()  # Extract text from each page

    except requests.exceptions.HTTPError as e:
        print(f'HTTP error occurred: {e}')
    except Exception as e:
        print(f'An error occurred: {e}')
    
    return text

def web_extractor(link):
    text = ''

    try:
        loader = WebBaseLoader(link)
        pages = loader.load_and_split()

        for page in pages:
            text+=page.page_content
    except:
        pass
    
    return text

def imporve_text(text):

    prompt = f'''
    Please rewrite the following text to make it short, descriptive, concise, and of high quality.
    Ensure that all essential information is retained. 
    Focus on improving clarity, coherence, and word choice without altering the original meaning.

    text = {text}
    '''

    model = random.choice([gemini,gemini1,gemini2,gemini3])
    result = model.invoke(prompt)

    return result.content

def feature_extraction(tag, history , context):

    prompt = f'''
    You are an intelligent assistant tasked with updating product information. You have two data sources:
    1. Tag_History: Previously gathered information about the product.
    2. Tag_Context: New data that might contain additional details.
    Your job is to read the Tag_Context and update the relevant field in the Tag_History with any new details found. The field to be updated is the {tag} FIELD.
    Guidelines:
    - Only add new details that are relevant to the {tag} FIELD.
    - Do not add or modify any other fields in the Tag_History.
    - Ensure your response is in coherent sentences, integrating the new details seamlessly into the existing information.
    Here is the data:
    Tag_Context: {str(context)}
    Tag_History: {history}
    Respond with the updated Tag_History.
    '''

    model = random.choice([gemini,gemini1,gemini2,gemini3])
    result = model.invoke(prompt)

    return result.content

def feature_extraction_image(url):
    text = ' '
    model = genai.GenerativeModel('gemini-1.5-flash-001')
    try:
        res = model.generate_content(['Describe this image to me',url])
        text = res.text

    except:
        pass
    return text

def detailed_feature_extraction(find, context):

    prompt = f'''
    You are an intelligent assistant tasked with finding product information. You have one data source and one output format:
    1. Context: The gathered information about the product.
    2. Format: Details which need to be filled based on Context.
    Your job is to read the Context and update the relevant field in Format using Context.
    Guidelines:
    - Only add details that are relevant to the individual FIELD.
    - Do not add or modify any other fields in the Format.
    - If nothing found return None.
    Here is the data:
    The Context is {str(context)}
    The Format is {str(find)}
    '''

    model = random.choice([gemini,gemini1,gemini2,gemini3])
    result = model.invoke(prompt)

    return result.content

def detailed_history(history):

    details = {
    "Introduction": {
        "Product Name": None,
        "Overview of the product": None,
        "Purpose of the manual": None,
        "Audience": None,
        "Additional Details": None
    },
    "Specifications": {
        "Technical specifications": None,
        "Performance metrics": None,
        "Additional Details": None
    },
    "Product Overview": {
        "Product features": None,
        "Key components and parts": None,
        "Additional Details": None
    },
    "Safety Information": {
        "Safety warnings and precautions": None,
        "Compliance and certification information": None,
        "Additional Details": None
    },
    "Installation Instructions": {
        "Unboxing and inventory checklist": None,
        "Step-by-step installation guide": None,
        "Required tools and materials": None,
        "Additional Details": None
    },
    "Setup and Configuration": {
        "Initial setup procedures": None,
        "Configuration settings": None,
        "Troubleshooting setup issues": None,
        "Additional Details": None
    },
    "Operation Instructions": {
        "How to use the product": None,
        "Detailed instructions for different functionalities": None,
        "User interface guide": None,
        "Additional Details": None
    },
    "Maintenance and Care": {
        "Cleaning instructions": None,
        "Maintenance schedule": None,
        "Replacement parts and accessories": None,
        "Additional Details": None
    },
    "Troubleshooting": {
        "Common issues and solutions": None,
        "Error messages and their meanings": None,
        "Support Information": None,
        "Additional Details": None
    },
    "Warranty Information": {
        "Terms and Conditions": None,
        "Service and repair information": None,
        "Additional Details": None
    },
    "Legal Information": {
        "Copyright information": None,
        "Trademarks and patents": None,
        "Disclaimers": None,
        "Additional Details": None

    }
}

    for key,val in history.items():

        find = details[key]

        details[key] = str(detailed_feature_extraction(find,val))

    return details


def get_embeddings(link,tag_option): 

        print(f"\n--> Creating Embeddings - {link}")

        if tag_option=='Complete Document Similarity':
            history = { "Details": "" }

        else:
            history = {
                    "Introduction": "",
                    "Specifications": "",
                    "Product Overview": "",
                    "Safety Information": "",
                    "Installation Instructions": "",
                    "Setup and Configuration": "",
                    "Operation Instructions": "",
                    "Maintenance and Care": "",
                    "Troubleshooting": "",
                    "Warranty Information": "",
                    "Legal Information": ""
                }

        # Extract Text -----------------------------
        print("Extracting Text")
        if link[-3:] == '.md' or link[8:11] == 'en.':
            text = web_extractor(link)
        else:
            text = pdf_extractor(link)

        # Create Chunks ----------------------------
        print("Writing Tag Data")


        if tag_option=="Complete Document Similarity":
            history["Details"] = feature_extraction("Details", history["Details"], text[0][:50000])
            
        else:
            chunks = text_splitter.create_documents(text)

            for chunk in chunks:

                with concurrent.futures.ThreadPoolExecutor() as executor:
                        future_to_key = {
                            executor.submit(
                                feature_extraction, f"Product {key}", history[key], chunk.page_content
                            ): key for key in history
                        }
                        for future in concurrent.futures.as_completed(future_to_key):
                            key = future_to_key[future]
                            try:
                                response = future.result()
                                history[key] = response
                            except Exception as e:
                                print(f"Error processing {key}: {e}")
            
        print("Creating Vectors")
        genai_embeddings=[]
            
        for tag in history:
            result = genai.embed_content(
                    model="models/embedding-001",
                    content=history[tag],
                    task_type="retrieval_document")
            genai_embeddings.append(result['embedding'])


        return history,genai_embeddings

def get_embed_chroma(link):

    print(f"\n--> Creating Embeddings - {link}")

    # Extract Text -----------------------------
    if link[-3:] == '.md' or link[8:11] == 'en.':
        text = web_extractor(link)
    else:
        text = pdf_extractor(link)
    print("\u2713 Extracting Text")

    # Create Chunks ----------------------------

    text = re.sub(r'\.{2,}', '.', text)
    text = re.sub(r'\s{2,}', ' ', text)
    text = re.sub(r'\d{7,}', '', text)

    text = re.sub(r'\n{2,}', '\n', text)


    chunks = text_splitter_small.split_text(text)
    # print(chunks[:2])
    print("\u2713 Writing Tag Data")

    # Creating Vector
    embedding_vectors=[]
    # textual_data = []
    print("\u2713 Creating Vectors")


    # batch_size = 1
    # # Process chunks in batches
    # for i in range(0, len(chunks), batch_size):
    #     batch = chunks[i:i + batch_size]

    #     # texts = [text for text in batch]
    #     # print(texts)

    # t1 = time.time()
    for chunk in chunks:
        # Tokenize the input text
        inputs = tokenizer(chunk, return_tensors="np", padding=True, truncation=True)
        
        # Convert inputs to int64
        input_ids = inputs['input_ids'].astype(np.int64)
        attention_mask = inputs['attention_mask'].astype(np.int64)
        token_type_ids = inputs.get('token_type_ids', np.zeros_like(input_ids)).astype(np.int64)  # Some models might not use token_type_ids

        # Create the input feed dictionary
        input_feed = {
            'input_ids': input_ids,
            'attention_mask': attention_mask,
            'token_type_ids': token_type_ids
        }

        # Get the model's outputs
        outputs = session.run(None, input_feed)

        # Convert the outputs to numpy and process as needed
        last_hidden_state = np.array(outputs[0])
        embeddings = last_hidden_state.mean(axis=1).tolist()
        embedding_vectors.extend(embeddings)
        # textual_data.a(text)
    
    # t2 = time.time()
    # print(t2-t1)
    return chunks , embedding_vectors


def get_image_embeddings(Product):
    image_embeddings = []
    
    links = search_images(Product)
    with concurrent.futures.ThreadPoolExecutor() as executor:
        descriptions = list(executor.map(feature_extraction_image, links))
    
    for description in descriptions:
        result = genai.embed_content(
                model="models/embedding-001",
                content=description,
                task_type="retrieval_document")
        
        image_embeddings.append(result['embedding'])
    # print(image_embeddings)
    return image_embeddings , links
            
global text_splitter
global data
global history

text_splitter = RecursiveCharacterTextSplitter(
    chunk_size = 10000,
    chunk_overlap  = 100,
    separators = ["",''," "]
)

# text_splitter_small = RecursiveCharacterTextSplitter(
#     chunk_size = 2000,
#     chunk_overlap  = 100,
#     separators = ["",''," "]
# )

text_splitter_small = TokenBasedTextSplitter(chunk_size=500, chunk_overlap=50)
# chunks = splitter.split_text(text)

if __name__ == '__main__':
    print(get_embed_chroma('https://www.galaxys24manual.com/wp-content/uploads/pdf/galaxy-s24-manual-SAM-S921-S926-S928-OS14-011824-FINAL-US-English.pdf'))
    # print(get_image_embeddings(Product='Samsung Galaxy S24'))