Introduction
Tagwise is a straightforward and effective AI-powered web application developed as an internship project to automatically categorize bookmarked links.
You can checkout project at Mantis Intern's Github.
This article clearly discusses the project's technical infrastructure, methodologies, and developed solutions.
Project Objective
Modern internet users frequently bookmark hundreds of links, but manually organizing these links is often time-consuming. Tagwise aims to automate this task, quickly and accurately categorizing bookmarks from a single URL input.
Technologies Used
Backend Framework: Django, Django REST Framework
Database: PostgreSQL (psycopg2)
HTTP Requests: htmx
HTML Parsing: BeautifulSoup4
Web Automation: Selenium, webdriver_manager
Artificial Intelligence: OpenAI GPT-4o, Google Gemini API
YouTube Integration: yt-dlp, youtube-transcript-api
Vector Store & Chatbot: LangChain, FAISS (used only for chatbot functionality)
System Workflow and Process
URL Processing
Users enter only the URL. The content from the URL is retrieved in HTML format using htmx. HTML content is parsed into the title, description, and main content using BeautifulSoup4.Special Process for YouTube Links
For YouTube links, video titles and descriptions are fetched using yt-dlp. If available, transcripts (subtitles) are retrieved using youtube-transcript-api. The gathered content is then sent to the AI for categorization.Alternative Content Capture (Selenium)
For sites where HTML content cannot be fetched or parsed, a screenshot of the page is captured using Selenium. This screenshot is sent as visual data to the AI model for category determination. Additionally, if the site lacks a thumbnail (og:image), the Selenium screenshot is automatically used as a thumbnail.
Category Assignment Approach
The categorization process is entirely performed through large language models (LLMs). Prompt engineering methods send content directly to OpenAI GPT-4o or Google Gemini API, automatically determining the category. Technologies such as vector store, RAG, or embeddings are not used in the category determination process.
Chatbot Feature and Vector Store Usage
The project also includes a chatbot feature allowing users to query their bookmark archives in natural language. This chatbot operates by converting bookmark content into embeddings via LangChain, which are then stored in a FAISS vector store. When a user query is received, relevant content is retrieved using the Retrieval-Augmented Generation (RAG) methodology, and presented to the user. These vector store and embedding operations are exclusively for chatbot functionality and are not involved in the categorization process.
Challenges Encountered and Solutions
Fetching HTML Content: Selenium screenshot solutions were employed for content that could not be directly fetched with htmx and BeautifulSoup4.
YouTube Transcript Absence: Categorization was conducted solely based on video titles and descriptions when transcripts were unavailable.
Thumbnail Absence: Selenium screenshots were utilized as thumbnails when og:image or similar visuals were missing.
Conclusion
Tagwise offers a simple yet efficient solution for automatically categorizing bookmarks quickly. The project was developed as part of an internship. No further development is planned for the time being.
Feel free to reach out with your questions and comments!