Dibas Kumar Borborah

Full Stack Engineer | AI/ML Enthusiast

Demo working preview

🚀 BroTube: YouTube on Steroids | Chat, Search, Sentiment Analysis & More

Welcome to BroTube — a next-gen YouTube experience where you can chat with transcripts, analyze sentiments of comments, search videos intelligently, and store knowledge for dynamic conversations!

In this blog post, Your Lordship 👑 unveils the full architecture and workflow of BroTube.

🔗 Project Links

⚡ Tech Stack Overview

  • Backend:

    • Python FastAPI (core backend service)
    • LangChain (orchestrating RAG pipelines)
    • Llama 3.3 (LLM for advanced reasoning)
    • YouTube API (fetching video details, transcripts, comments)
    • Redis (for caching results)
    • Sentiment Analysis (TOT + LangChain tools)

  • RAG Chat App:

    • Mastra AI (based on Hono.js server)
    • OpenAI Ada-002 (for embeddings)
    • GPT-4o and Mini models (for answers and small reasoning tasks)
    • PostgreSQL + pgvector (vector storage for embeddings)
    • Memory feature (persistent chat context)

  • Frontend:

    • Next.js 14
    • Shadcn UI (beautiful components)
    • TailwindCSS (styling)

  • DevOps:

    • Dockerized services
    • GitLab CI/CD Pipelines
    • EC2 Deployment (AWS)

🛠️ Workflow Deep Dive

1. Video Search (FastAPI)

  • User enters a query on the Next.js frontend.
  • Request is sent to the FastAPI backend, which:
    • Checks Redis cache for existing search results.
    • If found → returns cached results.
    • If not found → fetches fresh results from YouTube API, then caches them with a TTL.

2. Fetching and Storing Transcripts

  • When a user selects a video:
    • FastAPI backend fetches the video transcript from YouTube (or other transcript services).
    • First-time fetch → transcript is stored in PostgreSQL database.
    • Future accesses → transcript is directly retrieved from the database, avoiding redundant API calls.
Chat Layout

3. Comment Sentiment Analysis (FastAPI)

  • After fetching the video:

    • FastAPI also fetches top-level comments.
    • Comments are processed through a LangChain ToT pipeline using Llama 3.3.
    • Outputs for each comment:
      • Sentiment Label (Positive / Neutral / Negative)
      • Reasoning explanation (short text)

  • Results are sent back to the frontend for display.

    Comment Sentiments

4. Chatting with Video Transcripts (Mastra AI Server)

  • User opens the chat interface connected to the Mastra AI server.

  • The flow:

    • Transcript is chunked and embedded via OpenAI Ada embeddings.
    • Embeddings are stored in pgvector inside PostgreSQL.
    • When the user asks a question:
      • Relevant transcript chunks are retrieved (RAG Retrieval).
      • The context + user query is passed to GPT-4o or Mini model.
      • Answer is generated and returned to the frontend.

  • Memory Feature:

    • Past chats are maintained and reused to give continuity to conversations.

5. Frontend Magic ✨

  • Built with Next.js 14 + Shadcn UI for components.
  • Styled with TailwindCSS.
  • Features:
    • Video search and listing
    • Video player with synced transcripts
    • Sentiment analysis visualization for comments
    • Chat interface to interact with the transcript intelligently

6. DevOps: Deployment and Scaling

  • Backend (FastAPI) and chat server (Mastra AI) are Dockerized separately.
  • CI/CD Pipeline via GitLab:
    • Build → Test → Push Docker images → Deploy to EC2 instance.
  • Deployed with AWS EC2, Nginx reverse proxy, optional PM2 for node services.

🧠 Diagram (High-Level Architecture)

[User]
   ↓
[Next.js Frontend]
   ↓
[FastAPI Backend]
  ├──> YouTube API (for videos, comments)
  ├──> Redis (cache YouTube search)
  ├──> PostgreSQL (store transcripts)
  └──> Sentiment Analysis (LangChain + Llama 3.3)

[Mastra AI Chat Server]
  ├──> Fetch transcript chunks from pgvector
  ├──> Use GPT-4o/Mini for chat generation
  └──> Maintain memory for smooth conversations