Skip to content

README.md

Latest commit

 

History

History
139 lines (88 loc) · 3.45 KB

README.md

File metadata and controls

139 lines (88 loc) · 3.45 KB
Raw

V2 — Academic RAG Chatbot for Media Literacy

Built by Asvix

V2 is an AI-driven academic assistant designed to help teachers and students understand and teach Media Literacy more effectively. Powered by a hybrid architecture—open-source LLMs, RAG pipelines, Neo4j knowledge graphs, and multimodal capabilities—V2 provides deeply contextual responses, pedagogically structured explanations, and exam-oriented learning tools.

Vision

To make Media Literacy education accessible, accurate, and easily teachable through an AI-powered companion that supports lesson planning, conceptual understanding, and interactive learning.

Key Features

Dual User Modes

1. Teacher Mode (Primary Focus)

  • Structured responses in curated sections:

    • Methodology
    • Conceptual Understanding
    • Classroom Flow
    • Assessment Ideas

  • Uses the knowledge graph to surface prerequisite topics and related concepts.

  • Designed to support lesson preparation and pedagogical clarity.

2. Student Mode

  • Conversational QA experience.
  • Clear, concise academic explanations.
  • Multilingual output support.

Core Architecture

V2 is built on a multi-layered academic retrieval system consisting of:

1. Open-Source LLM

  • Scalable, cost-effective, and less dependent on deprecated APIs.
  • Candidates: LLaMA, Deepseek, Qwen, etc.
  • Served through an inference engine (vLLM/TGI).

2. Vector Database

Used for semantic search across media-literacy-related documents.

  • Option: Qdrant or Pinecone
  • Stores embeddings with detailed metadata.

3. Knowledge Graph (Neo4j)

Neo4j powers contextual understanding by structuring relationships between:

  • Subjects
  • Topics & Subtopics
  • Media literacy concepts
  • Resources
  • Question banks
  • Prerequisite relationships

This enhances teaching guidance and interconnected explanations.

4. Multi-source Dataset

Composed of:

  • PDF-based unstructured academic text
  • SQL database containing structured academic content, question banks, and metadata

Both datasets are unified into a RAG-ready searchable format.

5. RAG Pipeline

  • Query → Embedding → Vector Search
  • Expand context using Neo4j for concept relationships
  • Incorporate SQL-derived structured insights
  • Prompt templates vary based on Teacher/Student mode

Project Roadmap

Phase 1 – Data Engineering

  • PDF ingestion, cleaning, chunking
  • SQL → structured text conversion
  • Initial embedding + vector DB population
  • Neo4j schema + first KG build

Phase 2 – Backend (RAG + Dual Mode Behaviour)

  • Retrieval fusion: Vector DB + Neo4j + SQL
  • Teacher Mode vs Student Mode response engine
  • FastAPI backend endpoints

Phase 3 – Streamlit MVP

  • Role-based UI
  • Interactive chat interface
  • Teacher-optimized response formatting

Phase 4 – Premium Layer

  • Preparation mode
  • Speech & image support
  • User progress tracking

Phase 5 – Full Frontend Deployment

  • Custom React/Next.js frontend
  • Supabase for auth, user roles, subscription logic
  • Cloud-hosted backend infrastructure

Phase 6 – Iteration

  • Logging, evaluations
  • KG refinement
  • Dataset growth
  • Performance tuning

Contributing

We are working with the Asvix intern team. Internal contributors should follow:

  • Branch naming conventions
  • PR-based review workflows
  • Documentation updates with each new module

External contributions will be opened in future phases.