G2 Hackathon

Problem Statement

G2 regularly updates its website with new products by creating new categories and refining existing ones. One crucial aspect of this process is ensuring that each product has a precise description and URL before it is added to the site. We are interested in automating the process of updating product descriptions in our database. We will provide you with a few product URLs, and your output will be a brief 3-4 lines description of each product.

Proposed Solution

We built an end-to-end solution that encompasses a G2-themed extension where the user can enter the URL and, in turn, get a product summary.

The high-level approach involves:

1. The user enters a product URL in the G2 extension.
2. The extension sends a POST request to the backend server with the provided URL.
3. The backend server uses a web scraper (Scrapy) to extract the relevant product information from the given URL.
4. The extracted data is then cleaned and processed to generate a concise 3-4 line product summary.
5. The generated summary is returned to the extension and displayed to the user.

How to Run

First you need to start a CORS server (We have used ngrok for that)

Open Command Prompt and run

ngrok http --host-header=rewrite 5000

Now, check the Forwarding Session Status and copy the .app url. This will be your <your_ngrok_url>

Clone the GitHub Repository

Open manifest.json and replace the existing value of permissions with [<your_ngrok_url>/summarize]

Also, replace the existing value of content_security_policy after 'self' with <your_ngrok_url>/summarize

Now, Open popup.js and after fetch inside paranthesis, replace the existing url with <your_ngrok_url>/summarize

Chrome extension settings > Toggle the Developer Mode > Load unpacked > Add the deadlineDashers_Track3 (which is the clone GitHub Repository)

If the steps are correctly followed, you should be able to see the G2 logo in your extensions bar.

cd G2_Backend
python app.py

Don't forget to add your own .env with OpenAI API key and the secret token in the backend folder.

If the server.py runs fine, you should be able to see Running on http://127.0.0.1:5000 in the terminal.

Note: When testing our app we noticed that the scrapy module threw some unexpected exceptions in some machines. If you encounter any such error: please restart the server i.e the flask app

Output

1

Our digital identity solutions utilize a fast, accurate trust engine to engage customers through secure channels and prevent fraud attacks. With our digital trust anchor and
consortium, we validate customers and ensure account integrity, providing a seamless and trustworthy experience for all transactions.

2

Litzias VCIO services offer a preventative approach to network management, with a dedicated tech support staff and a variety of network installation and maintenance options. Our virtual chief information officer services fulfill the role of a liaison with technology vendors, providing proactive monitoring and successful technology choices for your business's specific requirements.

3

Revolutionize your recruiting process with our advanced collaboration platform, designed to supercharge your communications and streamline interactions between hiring managers and candidates. Our cutting-edge technology liberates customers and candidates alike, providing real-time communication and transparency for seamless recruitment velocity.

4

Initas is a leading AI-powered marketing tool that streamlines content creation for small businesses. With our ready-to-use website and appointment payment systems, businesses can easily generate engaging social media posts and online content in a matter of seconds, eliminating the need for constant manual input. Our AI model takes care of the basics, allowing businesses to focus on their campaign goals and target audience.

5

Welcome to our boutique agency, where we provide bespoke PR services to our clients. With a strategic approach and creative problem-solving, we have built a reputation for delivering exceptional results for senior executives, founders, and high-net-worth individuals. Our team of professionals has a proven track record of success, as showcased in our case studies and client testimonials. Let us help you elevate your brand and reach

Summary of Main Novelties

• Full Stack Project: We've developed a comprehensive solution comprising frontend (HTML, CSS, JS) and backend (Python, Flask, Scrapy, KeyBERT, GPT, NLTK), along with a proxy server (ngrok).

• Scraping Capabilities: Leveraging Python Scrapy as our primary scraper, we extract information from various URLs, including relevant pages such as "About Us".

• Tokenization Techniques: Employing techniques like stop word removal, sentence, and word tokenization, as well as KeyBERT for keyword extraction from text, to streamline input tokenization.

• Summarization Process: Utilizing GPT 3.5 turbo, we query a language model to summarize the final content based on our structured input prompt. We also tried a mixture of experts model called Mistral 7B

• Prompt Engineered Output: Our final output is engineered based on the input prompt provided to the model, defining the structure of the prompt for generating the summary.

Added functionalities on the Backend (based on feedback from the judges)

• Error Handling done better: Non scrapable websites now return a suitable error message:

  

• Redis In-Memory Key-Value Store: We used the docker image of Redis and ran it as a local container. The Key-Value in memory DB is populated for fast fetching of intermediate results, for better processing. The summary value for each URL is stored as: [summary:{url} : ActualSummary ]

  

• Better KeyPhrases: We developed a voting mechanism to select the top 20 key phrases from sets generated by Llama3, Mistral, and KeyBERT. The algorithm preprocesses and deduplicates phrases, then embeds them using SentenceTransformer's 'all-MiniLM-L6-v2' model. It then computes a cosine similarity matrix between phrase embeddings, calculates relevance scores based on semantic similarity, and ranks phrases accordingly. This approach ensures the selection of frequent, semantically diverse phrases representative of the main concepts, thus combining multiple models' strengths!

• Summary Auto Evaluation Model developed!: Our Summary Generation relies on KeyPhrase extractor models, which extract top 20 KeyPhrases.The eval function compares the main model's KeyPhrase list against those from multiple extraction models (Llama3, Mistral, KeyBERT, TextRank, YAKE). It uses SentenceTransformer embeddings and cosine similarity with a threshold (0.40) to determine matches, accounting for semantic relationships. The final score for each URL is an average of matched phrases across all models. Our best Run achieved an accuracy of 72% across all URLs, with the best accuracy reaching 84%!

  ['https://www.telesign.com/products/trust-engine': 0.8380952380952381, 'https://www.litzia.com/professional-it-services/': 0.7523809523809524, 'https://www.chattechnologies.com/': 0.69, 'https://inita.com/': 0.638095238095238, 'https://aim-agency.com/': 0.6190476190476191]

  Score: 0.72