LLAMA 3 Qdrant Anomaly Detection Using Vector Search

Anomaly detection refers to identifying patterns or instances that deviate significantly from the norm within a dataset. It helps detect unusual behaviour or outliers.

Vectors and language models (LLMs) can aid in querying and discovering anomalies by:

• Representing data points as vectors, allowing efficient similarity comparisons.
• Leveraging LLMs to understand context and identify unexpected patterns in textual or sequential data.

Anomaly Detection

For this, we need a dataset with anomalies. Why not use a dataset that auto-updates live so that constant anomalies occur, which can then be automatically detected and presented in a readable format for the user to analyze and apply in the real world?

I will be using a lot of references to articles and assets present in the links below.

For this tutorial, I will be using [1].

https://money.rediff.com/news

And something special for you guys: Llama 3, which was just released a few days ago.

This Lllama 3–8B-Instruct can be set up using [2] and can be downloaded from [3].

How to Install and Run Llama2 Locally on Windows for Free

MaziyarPanahi/Meta-Llama-3-8B-Instruct-GGUF at main

Code

GitHub - heathbrew/Vector-Databases-can-help-with-Anomaly-Detection: Pinecone , reddiffmoney…

Git clone this repo [2] and follow along for the set-up.

Indexing Web Content to Create a Dataset

Let’s explore all the available content.

import requests
from bs4 import BeautifulSoup
import pandas as pd

# URL for Rediff Money news
url = "https://money.rediff.com/news"

# Fetch HTML content using requests
response = requests.get(url)

if response.status_code == 200:
    # Parse HTML content using BeautifulSoup
    soup = BeautifulSoup(response.content, "html.parser")
    print(soup)  # Just added print for demonstration, you can modify or remove this line as per your need

This will parse the HTML for the provided URL. In a production application, you can use an array of URLs (or a database of URLs) that you index regularly to keep fresh data in your vector store.

Now, let’s convert this to a usable format.

import requests
from bs4 import BeautifulSoup
import pandas as pd

# URL for Rediff Money news
url = "https://money.rediff.com/news"

# Fetch HTML content using requests
response = requests.get(url)

if response.status_code == 200:
    # Parse HTML content using BeautifulSoup
    soup = BeautifulSoup(response.content, "html.parser")
    
    # Find all news items
    news_items = soup.find_all("div", class_="rtnews_row_more")
    
    # Extract relevant information from each news item
    news_data = []
    for item in news_items:
        title = item.find("p").text.strip()
        link = item.find("a")["href"]
        summary = item.find("div").text.strip()
        published = item.find("span", class_="timeago").text.strip()
        news_data.append({
            "title": title,
            "summary": summary,
            "link": link,
            "published": published
        })
    
    # Create pandas DataFrame
    df = pd.DataFrame(news_data)
    
    # Save DataFrame to CSV
    df.to_csv("Dataset/financial_news.csv", index=False)
    
    print("Financial news scraped and saved to financial_news.csv")
else:
    print("Failed to fetch the webpage")

Now, I have the financial news indexed and saved to financial_news.csv.

Let’s view it.

import pandas as pd
# Read the CSV file into a DataFrame
df = pd.read_csv("Dataset/financial_news.csv")
# Display the first few rows of the DataFrame
df.head()

As you can see, using a live dataset instead of a pre-made one is much better because it allows the user to track the latest financial news and discover events that can positively or negatively impact a stock they are interested in. This can enable the user to make buy and sell decisions regarding that stock.

for column in df.columns:
    print(f"{column}: {df[column][0]}")

Creating a Qdrant Database

In the last project [4], I used Qdrant which is a locally supported vector store for RAG.

Steps to Monitoring DSPy-Qdrant Powered RAG with Prometheus or Grafana

For this project, I will once again use Qdrant, but this time for financial data, to create a vector store.

Docker Desktop must be installed.

In [5] I have included a file named Qdrant.ps1.

You can use this to pull the Qdrant image and then run it on port 6333. You will see the local UI, which displays the vector stores.

#Use this code to create a qdrant collection
from qdrant_client import QdrantClient
from qdrant_client.http.models import Distance, VectorParams
 
# Initialize Qdrant client
qdrant_client = QdrantClient(host='localhost', port=6333)
collection_name = "Finance Outliers"
 
# Specify the vectors' configuration
vectors_config = VectorParams(
    size=model.config.hidden_size,  # The size of your embeddings
    distance=Distance.COSINE  # The distance metric for the vector space
)
 
# Create or recreate the collection with the specified configuration
qdrant_client.recreate_collection(
    collection_name=collection_name,
    vectors_config=vectors_config,
    # Optionally, you can specify other parameters for the collection
)

Feeding the Dataset to Vector Store

Take a look at the general dataset:

import pandas as pd
# Load the CSV file into a pandas DataFrame
df = pd.read_csv("Dataset/financial_news.csv")
df.head()

Loading the Embedding Model

I will be using the MINI-LM12-V2 [4] to embed the news.

from sentence_transformers import SentenceTransformer
from transformers import AutoTokenizer, AutoModel
from pathlib import Path
 
def download_model_and_tokenizer(model_name, save_path):
 """
 Download and save both the model and the tokenizer to the specified directory.
 
 Parameters:
     model_name (str): Name of the model to download.
     save_path (str or Path): Path to the directory where the model and tokenizer will be saved.
 """
 # Create the save path if it doesn't exist
 save_path = Path(save_path)
 save_path.mkdir(parents=True, exist_ok=True)
  
 # Initialize tokenizer and model
 tokenizer = AutoTokenizer.from_pretrained(model_name)
 model = AutoModel.from_pretrained(model_name)
  
 # Save tokenizer
 tokenizer.save_pretrained(save_path)
  
 # Save model
 model.save_pretrained(save_path)
 
# Example usage
model_name = 'sentence-transformers/all-MiniLM-L12-v2'  # Model name to download
save_path = Path("MiniLM-L12-v2/")  # Path where model and tokenizer will be saved
download_model_and_tokenizer(model_name, save_path)

Load the model and the tokenizer.

from transformers import AutoTokenizer, AutoModel
 
def load_model_and_tokenizer(model_path):
 """
 Load the model and tokenizer from the specified directory.
 
 Parameters:
     model_path (str or Path): Path to the directory containing the saved model and tokenizer.
 
 Returns:
     tokenizer (transformers.PreTrainedTokenizer): Loaded tokenizer.
     model (transformers.PreTrainedModel): Loaded model.
 """
 model_path = Path(model_path)
 tokenizer = AutoTokenizer.from_pretrained(model_path)
 model = AutoModel.from_pretrained(model_path)
 return tokenizer, model
 
# Load the model and tokenizer
model_path = Path("MiniLM-L12-v2/")
tokenizer, model = load_model_and_tokenizer(model_path)

Merge the title and the summary.

df['news'] = df.apply(lambda row: row['title'] + ' ' + row['summary'], axis=1)
df.head() 

Embed the news.

import torch
 
#Mean Pooling - Take attention mask into account for correct averaging
def mean_pooling(model_output, attention_mask):
 token_embeddings = model_output[0] #First element of model_output contains all token embeddings
 input_mask_expanded = attention_mask.unsqueeze(-1).expand(token_embeddings.size()).float()
 return torch.sum(token_embeddings * input_mask_expanded, 1) / torch.clamp(input_mask_expanded.sum(1), min=1e-9)
 
def generate_embedding(text):
 # Tokenize input text
 encoded_input = tokenizer(text, padding=True, truncation=True, return_tensors='pt')
 # Compute token embeddings with model
 with torch.no_grad():
     model_output = model(**encoded_input)
 # Perform mean pooling
 sentence_embedding = mean_pooling(model_output, encoded_input['attention_mask'])
 # Convert to numpy for FAISS compatibility and ensure it's 2D
 return sentence_embedding.cpu().numpy().reshape(1, -1)
 
# Generate embeddings for the 'input' column
df['encoded_news'] = df['news'].apply(lambda x: generate_embedding(x)[0].tolist())
df.head()

Pushing the Data to Qdrant Vector Store

Querying Qdrant

Query the anomalies in any way you want.

from langchain.vectorstores import Qdrant
 
url = "http://localhost:6333"  # URL where the Qdrant service is running
collection_name =  "Finance Outliers"  # Name of the collection in Qdrant
 
# Initialize the Qdrant client with the specified URL
client = QdrantClient(
    url=url,
    prefer_grpc=False  # Indicates whether to use gRPC for communication
)
 
logging.info(f"QdrantClient initialized: {client}")  # Prints the client information
logging.info(f"#################################")  # Prints a separator line
 
# Create a Qdrant object with the specified client, embeddings, and collection name
# Initialize the Qdrant vector store from langchain
db = Qdrant(
    client=client,
    embeddings=df['encoded_news'].tolist(),  # Use the generated embeddings
    collection_name=collection_name
)
 
logging.info(f"Qdrant vector store initialized: {db}") # Prints the database object information

2024-04-26 16:27:55 - INFO - QdrantClient initialized: <qdrant_client.qdrant_client.QdrantClient object at 0x0000025DABB59D60>
2024-04-26 16:27:55 - INFO - #################################
D:\Desktop\Superteams AI\Task 2 Vector Databases can help with Anomaly Detection\Vector-Databases-can-help-with-Anomaly-Detection\venv\Lib\site-packages\langchain_community\vectorstores\qdrant.py:150: UserWarning: `embeddings` should be an instance of `Embeddings`.Using `embeddings` as `embedding_function` which is deprecated
  warnings.warn(
2024-04-26 16:27:55 - INFO - Qdrant vector store initialized: <langchain_community.vectorstores.qdrant.Qdrant object at 0x0000025DAD359F70>

def similarity_search_with_score(query, k=2):
    query_embedding = generate_embedding(query)[0].tolist()
    search_results = qdrant_client.search(
        collection_name=collection_name,
        query_vector=query_embedding,
        limit=k,
        with_payload=True,
        with_vectors=False
    )
    return search_results
 
query = "Stock market crashes due to unexpected event"
search_results = similarity_search_with_score(query=query, k=5)
 
for result in search_results:
    doc_id = result.id
    score = result.score
    payload = result.payload  # The payload should contain your text or a reference to it.
 
    # Assuming the payload contains a field 'input' where the text is stored
    doc_content = payload.get('output', 'No content available')
 
    # Print the similarity score and document content
    logging.info({"score": score, "doc_id": doc_id, "content": doc_content})

2024-04-26 16:33:48 - INFO - HTTP Request: POST http://localhost:6333/collections/Finance%20Outliers/points/search "HTTP/1.1 200 OK"
2024-04-26 16:33:48 - INFO - {'score': 0.296198, 'doc_id': 0, 'content': "Kotak Mahindra Bank's loan, deposit growth may be impacted after RBI curbs Kotak Mahindra Bank's loan, deposit growth may be impacted after RBI curbs\n         Kotak Mahindra Bank’s loan and deposit growth are likely to be affected after the Reserve Bank of India (RBI) asked the private-sector lender not to take on board new customers through the bank’s online and mobile banking channels and not to issue any new credit cards, according to analysts.   Photograph: Adnan Abidi/Reuters The bank’s share price fell 10.85 per cent on Thursday to close the day at Rs 1,643 on the BSE. The RBI’s action came after market hours on ...\nRediff.com, 1 hour(s) ago\nAlso from:"}
2024-04-26 16:33:48 - INFO - {'score': 0.28810832, 'doc_id': 9, 'content': 'Sensex revisits 74K; Nifty climbs 168 points Sensex revisits 74K; Nifty climbs 168 points\n         Rising for the fifth straight session, equity benchmark Sensex rallied nearly 500 points to reclaim the 74,000 mark while the Nifty closed above the 22,550 level on Thursday, driven by heavy buying in banking, financial and metal stocks.   Photograph: Shailesh Andrade/Reuters Recovering after a sell-off in early trade, the 30-share BSE Sensex climbed 486.50 points or 0.66 per cent to settle at 74,339.44. During the day, it surged 718.31 points or 0.97 per cent to 74,571.25. The NSE Nifty ...\nRediff.com, 23 hour(s) ago\nAlso from: Rediff.com'}
2024-04-26 16:33:48 - INFO - {'score': 0.26815844, 'doc_id': 6, 'content': 'Good Q4, commentary perk up ICICI Lombard General Insurance stock Good Q4, commentary perk up ICICI Lombard General Insurance stock\n            Investment yields could be around 8.1 per cent in FY25 rising to 8.5 per cent in FY26   Photograph: Courtesy, ICICI Lombard ICICI Lombard General Insurance Company reported financial improvement and optimistic commentary in Q4FY24. It reported 17 per cent year-on-year (YoY) growth in Gross Written Premium (GWP) and 115 bps improvement in the Combined Ratio (COR) in FY24, and improved COR guidance with COR going from 104.5 per cent in FY23 to 103.3 per cent in FY24, 102.4 per cent in FY25 and ...\nRediff.com, 5 hour(s) ago\nAlso from:'}
2024-04-26 16:33:48 - INFO - {'score': 0.23023885, 'doc_id': 4, 'content': "Tech Mahindra jumps over 12% in opening trade Tech Mahindra jumps over 12% in opening trade\n         Equity benchmark indices climbed in early trade on Friday, extending their rally for the sixth day running, on heavy buying in Tech Mahindra and firm trends in Asian markets.   Photograph: Adnan Abidi/Reuters The 30-share BSE Sensex climbed 176.47 points to 74,515.91 in early trade. The NSE Nifty went up by 50.05 points to 22,620.40. \xa0 From the Sensex basket, Tech Mahindra jumped over 12.50 per cent after the IT services company's CEO outlined an ambitious three-year roadmap to accelerate ...\nRediff.com, 5 hour(s) ago\nAlso from:"}
2024-04-26 16:33:48 - INFO - {'score': 0.21434154, 'doc_id': 1, 'content': 'Above-normal monsoon likely to ease food prices: FinMin Above-normal monsoon likely to ease food prices: FinMin\n         With the prediction of an above normal monsoon in 2024, the government is expecting food prices to come down, the finance ministry’s monthly economic report for March has said.   Photograph: Amit Dave/Reuters The report, released on Thursday, said robust foreign inflows and comfortable trade deficits were expected to keep the rupee within a comfortable range. “Further easing of food prices is on the anvil as IMD (India Meteorological Department) has predicted above-normal rainfall ...\nRediff.com, 3 hour(s) ago\nAlso from:'}

Modify This Function to Give Contacted String

def similarity_search_with_score(query, k=2):
    query_embedding = generate_embedding(query)[0].tolist()
    search_results = qdrant_client.search(
        collection_name=collection_name,
        query_vector=query_embedding,
        limit=k,
        with_payload=True,
        with_vectors=False
    )
 
    # Extract the document content from the payload and include it in the results
    results_with_content = []
    for result in search_results:
        doc_id = result.id
        score = result.score
        payload = result.payload  # The payload should contain your text or a reference to it.
       
        # Extract the document content from the payload
        doc_content = payload.get('output', 'No content available')
 
        results_with_content.append((score, doc_content))
 
    # Sort the results based on the similarity score in descending order
    sorted_results = sorted(results_with_content, key=lambda x: x[0], reverse=True)
 
    # Concatenate the content of the top k results
    concatenated_content = ' '.join([content for _, content in sorted_results[:k]])
 
    return concatenated_content
 
query = "Stock market crashes due to unexpected event"
outlier_paragraph = similarity_search_with_score(query=query, k=5)
 
# Print the concatenated content
logging.info({"concatenated_content": outlier_paragraph})

2024-04-26 16:36:50 - INFO - HTTP Request: POST http://localhost:6333/collections/Finance%20Outliers/points/search "HTTP/1.1 200 OK"
2024-04-26 16:36:50 - INFO - {'concatenated_content': "Kotak Mahindra Bank's loan, deposit growth may be impacted after RBI curbs Kotak Mahindra Bank's loan, deposit growth may be impacted after RBI curbs\n            Kotak Mahindra Bank’s loan and deposit growth are likely to be affected after the Reserve Bank of India (RBI) asked the private-sector lender not to take on board new customers through the bank’s online and mobile banking channels and not to issue any new credit cards, according to analysts.   Photograph: Adnan Abidi/Reuters The bank’s share price fell 10.85 per cent on Thursday to close the day at Rs 1,643 on the BSE. The RBI’s action came after market hours on ...\nRediff.com, 1 hour(s) ago\nAlso from: Sensex revisits 74K; Nifty climbs 168 points Sensex revisits 74K; Nifty climbs 168 points\n            Rising for the fifth straight session, equity benchmark Sensex rallied nearly 500 points to reclaim the 74,000 mark while the Nifty closed above the 22,550 level on Thursday, driven by heavy buying in banking, financial and metal stocks.   Photograph: Shailesh Andrade/Reuters Recovering after a sell-off in early trade, the 30-share BSE Sensex climbed 486.50 points or 0.66 per cent to settle at 74,339.44. During the day, it surged 718.31 points or 0.97 per cent to 74,571.25. The NSE Nifty ...\nRediff.com, 23 hour(s) ago\nAlso from: Rediff.com Good Q4, commentary perk up ICICI Lombard General Insurance stock Good Q4, commentary perk up ICICI Lombard General Insurance stock\n            Investment yields could be around 8.1 per cent in FY25 rising to 8.5 per cent in FY26   Photograph: Courtesy, ICICI Lombard ICICI Lombard General Insurance Company reported financial improvement and optimistic commentary in Q4FY24. It reported 17 per cent year-on-year (YoY) growth in Gross Written Premium (GWP) and 115 bps improvement in the Combined Ratio (COR) in FY24, and improved COR guidance with COR going from 104.5 per cent in FY23 to 103.3 per cent in FY24, 102.4 per cent in FY25 and ...\nRediff.com, 5 hour(s) ago\nAlso from: Tech Mahindra jumps over 12% in opening trade Tech Mahindra jumps over 12% in opening trade\n         Equity benchmark indices climbed in early trade on Friday, extending their rally for the sixth day running, on heavy buying in Tech Mahindra and firm trends in Asian markets.   Photograph: Adnan Abidi/Reuters The 30-share BSE Sensex climbed 176.47 points to 74,515.91 in early trade. The NSE Nifty went up by 50.05 points to 22,620.40. \xa0 From the Sensex basket, Tech Mahindra jumped over 12.50 per cent after the IT services company's CEO outlined an ambitious three-year roadmap to accelerate ...\nRediff.com, 5 hour(s) ago\nAlso from: Above-normal monsoon likely to ease food prices: FinMin Above-normal monsoon likely to ease food prices: FinMin\n         With the prediction of an above normal monsoon in 2024, the government is expecting food prices to come down, the finance ministry’s monthly economic report for March has said.   Photograph: Amit Dave/Reuters The report, released on Thursday, said robust foreign inflows and comfortable trade deficits were expected to keep the rupee within a comfortable range. “Further easing of food prices is on the anvil as IMD (India Meteorological Department) has predicted above-normal rainfall ...\nRediff.com, 3 hour(s) ago\nAlso from:"}

RAG Using Llama 3

Using the concatenated result, generate a paragraph on anomalies that allows you to read everything at once. Article [5] explains how to set up Llama 2 on your local system and run it. The code in [2] that you just git cloned includes the PS1 script to run it. The code in [2] contains a file called llama3backend, which makes llama3 a simple function call away. This file automatically calls llama2 GGUF, which is a quantized version of llama3.

query = "Stock market crashes due to unexpected event ? Summarize this in one paragraph "
from llama3backend import generate_text
RAG_answer = generate_text(str(query + outlier_paragraph)[:512])

2024–04–26 16:45:20 - INFO - {'RAG_answer': " debit cards or other payment instruments.\nThe RBI has taken this step as a precautionary measure to maintain financial stability in the country. The RBI has also asked Kotak Mahindra Bank to review its lending policies and ensure that they are in line with the RBI's guidelines.\nIn conclusion, Kotak Mahindra Bank's loan and deposit growth may be impacted after the Reserve Bank of India (RBI) asked the private-sector lender not to take on board new customers through the bank's online and mobile banking channels and not to issue any new credit cards, debit cards or other payment instruments. The RBI has taken this step as a precautionary measure to maintain financial stability in the country. The RBI has also asked Kotak Mahindra Bank to review its lending policies and ensure that they are in line with the RBI's guidelines.\nIn conclusion, Kotak Mahindra Bank's loan and deposit growth may be impacted after the Reserve Bank of India (RBI) asked the private-sector lender not to take on board new customers through the bank's online and mobile banking channels and not to issue any new credit cards, debit cards or other payment instruments. The RBI has taken this step as a precautionary measure to maintain financial stability in the country. The RBI has also asked Kot"}

Anomaly Detection

This method is used for larger datasets, so we will be using [6]. We will be loading it

News Category Dataset

import pandas as pd
# Load the CSV file into a pandas DataFrame
df = pd.read_json("Dataset/News_Category_Dataset_v3.json" , lines=True)
df.head()

Take the first 1000 rows.

# Select the first 1000 rows of the DataFrame
df = df.head(1000)

Create embeddings like before.

# Generate embeddings for the 'input' column
df['encoded_news'] = df['news'].apply(lambda x: generate_embedding(x)[0].tolist())
df.head()

Apply t-SNE for Dimensionality Reduction

import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
import seaborn as sns
from sklearn.manifold import TSNE
 
X = np.array(df['encoded_news'].tolist())
# Increase the perplexity value (default is 30)
tsne = TSNE(random_state=0, n_iter=1000, perplexity=500)
tsne_results = tsne.fit_transform(X)
 
df_tsne = pd.DataFrame(tsne_results, columns=['TSNE1', 'TSNE2'])
df_tsne['Class Name'] = df['category']  # Using 'title' as a placeholder for 'Class Name'
df_tsne['news'] = df['news']
# df_tsne['encoded_news'] = df['encoded_news']
df_tsne.head()

# Plot t-SNE results
fig, ax = plt.subplots(figsize=(8, 6))
sns.set_style('darkgrid', {"grid.color": ".6", "grid.linestyle": ":"})
sns.scatterplot(data=df_tsne, x='TSNE1', y='TSNE2', hue='Class Name', palette='Set2')
sns.move_legend(ax, "upper left", bbox_to_anchor=(1, 1))
plt.title('Scatter plot of news using t-SNE')
plt.xlabel('TSNE1')
plt.ylabel('TSNE2')
plt.show()

Outlier Detection

# Function to get centroids of each class
def get_centroids(df_tsne):
    centroids = df_tsne.drop(columns=['news']).groupby('Class Name').mean()
    return centroids

centroids = get_centroids(df_tsne)

# Function to detect outliers
def calculate_euclidean_distance(p1, p2):
    return np.sqrt(np.sum(np.square(p1 - p2)))

def detect_outlier(df, emb_centroids, radius):
    outlier_indices = []
    for idx, row in df.iterrows():
        class_name = row['Class Name']
        dist = calculate_euclidean_distance(np.array([row['TSNE1'], row['TSNE2']]),
                                            np.array([emb_centroids.loc[class_name, 'TSNE1'],
                                                      emb_centroids.loc[class_name, 'TSNE2']]))
        if dist > radius:
            outlier_indices.append(idx)
    return outlier_indices
# Assuming df_tsne and centroids are already defined DataFrames
range_ = np.arange(0.01, 1.0, 0.05).round(decimals=2).tolist()
outliers_list = []

for i in range_:
    outliers = detect_outlier(df_tsne, centroids, i)
    outliers_list.append(outliers)

# Combine all outlier indices into a single list
all_outliers = [idx for sublist in outliers_list for idx in sublist]

# Update the 'Outlier' column in df_tsne
df_tsne['Outlier'] = df_tsne.index.isin(all_outliers)
df_tsne.head()

# Filter out rows where 'Outlier' is False
df_tsne = df_tsne[df_tsne['Outlier'] == True]
num_outliers = [len(outliers) for outliers in outliers_list]
import matplotlib.pyplot as plt

# Plot range_ and num_outliers
fig = plt.figure(figsize=(14, 8))
plt.rcParams.update({'font.size': 12})
plt.bar(list(map(str, range_)), num_outliers)
plt.title("Number of outliers vs. distance of points from centroid")
plt.xlabel("Distance")
plt.ylabel("Number of outliers")
for i in range(len(range_)):
    plt.text(i, num_outliers[i], num_outliers[i], ha='center')

plt.show()

def get_outlier_texts(df, class_name):
    # Filter the DataFrame to get outliers of the specified category
    outliers = df[(df['Class Name'] == class_name) & df['Outlier']]
    
    # Extract the outlier texts
    outlier_texts = outliers['news'].tolist()
    
    return outlier_texts

# Example usage:
outlier_texts = get_outlier_texts(df_tsne, 'TECH')
for idx, text in enumerate(outlier_texts, start=1):
    print(f"Outlier {idx}: {text}\n")

Outlier 1: Twitch Bans Gambling Sites After Streamer Scams Folks Out Of $200,000 One man's claims that he scammed people on the platform caused several popular streamers to consider a Twitch boycott.

Outlier 2: TikTok Search Results Riddled With Misinformation: Report A U.S. firm that monitors false online claims reports that searches for information about prominent news topics on TikTok are likely to turn up results riddled with misinformation.

Outlier 3: Citing Imminent Danger Cloudflare Drops Hate Site Kiwi Farms Cloudflare CEO Matthew Prince had previously resisted calls to block the site.

Outlier 4: Instagram And Facebook Remove Posts Offering Abortion Pills Facebook and Instagram began removing some of these posts, just as millions across the U.S. were searching for clarity around abortion access.

Outlier 5: Google Engineer On Leave After He Claims AI Program Has Gone Sentient Artificially intelligent chatbot generator LaMDA wants “to be acknowledged as an employee of Google rather than as property," says engineer Blake Lemoine.

Outlier 6: Facebook Is Still Allowing Mug Shots Even Though They Can Ruin Lives When an individual’s mug shot goes viral on Facebook, they are often subjected to extreme harassment and struggle to find stable housing and employment.

Outlier 7: Ex-Twitter CEO Dings Elon Musk For Attacks On Twitter's Top Lawyer A one-sided feud between Musk and Vijaya Gadde has turned even uglier.

Outlier 8: Investor Sues Elon Musk Over His Delayed Twitter Filing Marc Rasella says he sold shares of Twitter at “artificially deflated prices,” unaware that Musk had made a large purchase in the social media platform.

def get_outlier_paragraph(df, class_name):
    # Filter the DataFrame to get outliers of the specified category
    outliers = df[(df['Class Name'] == class_name) & df['Outlier']]
    
    # Extract the outlier texts
    outlier_texts = outliers['news'].tolist()
    
    # Concatenate all outlier texts into one paragraph
    outlier_paragraph = ' '.join(outlier_texts)
    
    return outlier_paragraph

# Example usage:
outlier_paragraph = get_outlier_paragraph(df_tsne, 'TECH')
print(outlier_paragraph)

Twitch Bans Gambling Sites After Streamer Scams Folks Out Of $200,000 One man's claims that he scammed people on the platform caused several popular streamers to consider a Twitch boycott. TikTok Search Results Riddled With Misinformation: Report A U.S. firm that monitors false online claims reports that searches for information about prominent news topics on TikTok are likely to turn up results riddled with misinformation. Citing Imminent Danger Cloudflare Drops Hate Site Kiwi Farms Cloudflare CEO Matthew Prince had previously resisted calls to block the site. Instagram And Facebook Remove Posts Offering Abortion Pills Facebook and Instagram began removing some of these posts, just as millions across the U.S. were searching for clarity around abortion access. Google Engineer On Leave After He Claims AI Program Has Gone Sentient Artificially intelligent chatbot generator LaMDA wants “to be acknowledged as an employee of Google rather than as property," says engineer Blake Lemoine. Facebook Is Still Allowing Mug Shots Even Though They Can Ruin Lives When an individual’s mug shot goes viral on Facebook, they are often subjected to extreme harassment and struggle to find stable housing and employment. Ex-Twitter CEO Dings Elon Musk For Attacks On Twitter's Top Lawyer A one-sided feud between Musk and Vijaya Gadde has turned even uglier. Investor Sues Elon Musk Over His Delayed Twitter Filing Marc Rasella says he sold shares of Twitter at “artificially deflated prices,” unaware that Musk had made a large purchase in the social media platform.

RAG Using Llama 3

query = "write a summary for this ? "
print(str(query + outlier_paragraph)[:512])

write a summary for this ? Twitch Bans Gambling Sites After Streamer Scams Folks Out Of $200,000 One man’s claims that he scammed people on the platform caused several popular streamers to consider a Twitch boycott. TikTok Search Results Riddled With Misinformation: Report A U.S. firm that monitors false online claims reports that searches for information about prominent news topics on TikTok are likely to turn up results riddled with misinformation. Citing Imminent Danger Cloudflare Drops Hate Site Kiwi Fa

from llama3backend import generate_text
RAG_answer = generate_text(str(query + outlier_paragraph)[:512])
RAG_answer

‘… (read more) …r, Which Was Linked To The Christchurch Mosque Massacre. Cloudflare, a content delivery network (CDN), has dropped its support for hate site Kiwi Farms, which was linked to the Christchurch mosque massacre. The decision comes after Cloudflare faced intense pressure from human rights groups and other organizations to sever ties with the hate site. In a statement, Cloudflare said that it had “re-evaluated” its relationship with Kiwi Farms and had decided to terminate its support for the site. The company said that it would continue to provide services to other websites and organizations that promote hate speech or other forms of discrimination. (read more) …r, Which Was Linked To The Christchurch Mosque Massacre. Cloudflare, a content delivery network (CDN), has dropped its support for hate site Kiwi Farms, which was linked to the Christchurch mosque massacre. The decision comes after Cloudflare faced intense pressure from human rights groups and other organizations to sever ties with the hate site. In a statement, Cloudflare said that it had “re-evaluated” its relationship with Kiwi Farms and had decided to terminate its support for the site. The company said that it would continue to provide services to other websites and organizations that’

References

[1] “Realtime Market News. Live BSE, NSE, Stock Prices, Expert Stock Advice, Share Market Updates : Rediff.com.” Accessed: Apr. 18, 2024. [Online]. Available: https://money.rediff.com/news

[2] A. Pranav, “How to Install and Run Llama2 Locally on Windows for Free,” Medium. Accessed: Apr. 18, 2024. [Online]. Available: https://medium.com/@AyushmanPranav/how-to-install-and-run-llama2-locally-on-windows-for-free-05bd5032c6e3

[3] “MaziyarPanahi/Meta-Llama-3–8B-Instruct-GGUF at main.” Accessed: Apr. 26, 2024. [Online]. Available: https://huggingface.co/MaziyarPanahi/Meta-Llama-3-8B-Instruct-GGUF/tree/main

[4] A. Pranav, “Steps to Monitoring DSPy-Qdrant Powered RAG with Prometheus or Grafana,” DevOps.dev. Accessed: Apr. 18, 2024. [Online]. Available: https://blog.devops.dev/steps-to-monitoring-dspy-qdrant-powered-rag-with-prometheus-or-grafana-b642335cbd50

[5] “heathbrew/Vector-Databases-can-help-with-Anomaly-Detection: Pinecone , reddiffmoney financial dataset.” Accessed: Apr. 18, 2024. [Online]. Available: https://github.com/heathbrew/Vector-Databases-can-help-with-Anomaly-Detection

[6] “News Category Dataset.” Accessed: Apr. 18, 2024. [Online]. Available: https://www.kaggle.com/datasets/rmisra/news-category-dataset