Course Finder Agent

The Problem: LLM Hallucinations

Generative AI models, while powerful, sometimes "hallucinate" – providing inaccurate information or links to resources (like online courses or YouTube videos) that don't actually exist. This can be frustrating when trying to find reliable learning materials.

---

The Solution: Real-Time Search and Synthesis

This project tackles the hallucination problem by building a LangGraph agent that finds and ranks online learning resources based on real-time data. Instead of relying solely on the LLM's internal knowledge, it actively searches the web and YouTube for relevant courses.

---

Things I learned...

I wanted to make this project to improve my ability to build an advanced AI agent , This project gave me a better understanding of how to use langchain. I also learned using API services from third party providers . structuring a project with organized .py files and also improved cli commands relating to git

---

How it Works

The agent follows these steps when you provide a course topic (e.g., "javascript course"):

1. Query Input: Takes your desired course topic as input.
2. Multi-Source Search:
   • Uses the Bright Data SERP API to perform a Google Search for relevant courses.
   • Uses the Bright Data YouTube Scraper (simulated or real, via Youtube_api) to find relevant YouTube video courses/tutorials.
3. Data Extraction & Analysis:
   • Employs a Large Language Model (LLM, specifically gpt-4o-mini via LangChain) with structured output capabilities (with_structured_output).
   • Parses the Google search results and YouTube video data (including titles, descriptions, transcripts) to extract key details (course name, platform, topics, duration, cost) using predefined Pydantic models (OutputStructure, Analysis).
4. Synthesis & Ranking:
   • A final LLM step analyzes the extracted course lists from both Google and YouTube.
   • It selects the top 3 courses from each source.
   • It ranks these 6 courses based on relevance, content quality, and clarity, providing a justification for each rank.
5. Formatted Output: Presents the final ranked list in a clean, readable format using a Pydantic model (FinalSynthesis, RankedCourse).

---

Key Features

• LangGraph Orchestration: Uses LangGraph to define and run the multi-step agent workflow.
• Real-Time Data: Leverages Bright Data APIs for up-to-date search and video information, minimizing hallucinations.
• Structured Output: Utilizes Pydantic models and LangChain's with_structured_output to ensure reliable data extraction from the LLM.
• Multi-Source Synthesis: Combines and ranks results from both Google web search and YouTube for a comprehensive overview.
• Clear Ranking & Justification: Provides a ranked list with explanations for why each course was chosen.

---

Technology Stack

• Programming Language: Python
• Orchestration: LangGraph
• LLM Interaction: LangChain, langchain-openai (using gpt-4o-mini)
• Data Modeling: Pydantic
• Web Search: Bright Data SERP API (via web_actions.py)
• YouTube Scraping: Bright Data Scraper API (via web_actions.py)
• Environment Management: python-dotenv

---

Setup and Running

• ** You can clone this git repo but you should also add an .env file and include an OPENAI_API_KEY and a BRIGHTDATA_API_KEY from the relavent websites