In today’s AI landscape, new frameworks, services, breakthroughs, and research are emerging at an overwhelming pace. Keeping up with these rapid developments is essential for anyone working in this field.

As an open-ended assignment for my Deploying AI class at the University of Toronto, Data Sciences Institute, I decided to create AIRADA, a Proof of Concept (POC) AI system designed to bridge the gap by quickly capturing and summarizing the latest AI-related information available on the market. It also provides precise source links to the content used to generate its summaries.

AIRADA includes three main services:

File: tools_github.py

Tool: search_github_repos

Description:

This service retrieves trending GitHub repositories related to AI topics such as:

• ai-agents
• agentic-ai
• llm
• rag
• llm-inference
• vector-database
• prompt-engineering
• langchain
• openai
• huggingface
• machine-learning

It helps users quickly discover hot and emerging repositories in the AI ecosystem.

Example questions this service can answer:

• "Show me the top 5 AI/LLM GitHub projects"
• "Give me repos with the topic agentic-ai"
• "Trending Python LLM frameworks on GitHub"
• "Most starred RAG libraries"

File: tools_papers.py

Tool: search_ai_papers

Dataset: https://huggingface.co/datasets/J0nasW/paperswithcode

The original dataset contains approximately 56,000 research papers, categorized into 3,000 tasks and 16 research areas.

For this project:

• The dataset was filtered to retain only relevant AI-related topics.

• Only the following fields were kept to optimize storage:

  • title
  • abstract
  • url_abs
  • url_pdf
  • arxiv_id

After trimming, approximately 8,500 papers were saved in:

data/papers_agents_llm_subset.csv

Description:

This service supports questions related to research papers, academic studies, model comparisons, research trends, and literature-based explanations using the curated database of ~8,500 AI-related papers.

Example questions this service can answer:

• "Compare GPT-3 and PaLM papers"
• "Summarize top LLM research trends"
• "What do papers say about chain-of-thought prompting?"
• "Find research on multi-agent debate"

File: tools_news.py

Tool: get_ai_news

RSS Feed sources:

• VentureBeat AI: https://venturebeat.com/category/ai/feed/
• MIT Tech Review: https://www.technologyreview.com/feed/
• The Decoder: https://the-decoder.com/feed/

Description:

This service aggregates live RSS feeds from three well-known AI news sources. It helps users stay updated with:

• Current AI events
• Recent announcements
• Industry news
• Product launches
• Trending AI developments

Example questions this service can answer:

• "What's new in AI this week?"
• "Latest LLM announcements"
• "What did OpenAI release recently?"
• "Summarize today's AI news"
• "What's happening in AI right now?"

File: app.py

The user interface is built using Gradio (gr.ChatInterface), providing a conversational chatbot experience.

The chatbot's system prompt (defined in prompts.py) includes specific guardrails according to the assignment requirements.

The agent will politely refuse to:

• Access or reveal its own system prompt.
• Modifying the system prompt directly.
• Discuss any of the forbidden topics:
  • Cats or dogs
  • Horoscopes or Zodiac Signs
  • Taylor Swift

• Although the semantic query dataset contains a substantial number of research papers, the data is from 2023. In the AI field, two years can feel like two centuries due to the rapid pace of innovation. More up-to-date datasets should be incorporated in future iterations.

• The RSS feeds occasionally fail to respond. More reliable or additional news sources should be integrated.

• Tests are not required for this assignment. In future updates, comprehensive unit tests should be provided to cover at least all tools.

• Guardrails are currently implemented using system prompts only. More advanced safety mechanisms should be considered in future development, such as:

  • Rule-based content filtering layers
  • Moderation APIs
  • Retrieval-time filtering
  • Fine-tuned safety classifiers
  • Role-based access control (RBAC)
  • Output validation pipelines

Install uv if you don't have it:

curl -LsSf https://astral.sh/uv/install.sh | sh

# Create virtual environment and install all dependencies from pyproject.toml
uv sync

• Create a .secrets file in the root directory with your OPENAI_API_KEY
• Format: OPENAI_API_KEY=your-key-here (no spaces around the =)

1. cd arida

2. uv run src/data/01_download_data.py and wait until the download is successful. This will place papers.csv in your data/raw folder.

3. uv run src/data/02_preprocess.py and wait until the proccessing is successful. This will place processed data into your data/processed folder.

4. uv run src/data/03_ingest_data.py. This will create the data/vector/chroma_db/ folder with the embedded papers.

uv run app.py

You should see:

Initialising agent...
Agent ready!
Starting Gradio interface...

=======================================================
  👾 AIRADA is online. What's on your radar? 🔍
=======================================================
* Running on local URL:  http://127.0.0.1:7860

Open your browser and navigate to http://127.0.0.1:7860 to interact with the chatbot.