Skip to main content
Erica is an AI course tutor designed to answer questions, provide explanations, and assist with learning through intelligent retrieval and generation. This demonstrator showcases a GraphRAG (Graph-based Retrieval Augmented Generation) system built for the “Introduction to AI” course, leveraging a knowledge graph of concepts, resources, and relationships to deliver pedagogically sound responses. While a full implementation of Erica includes multimodal capabilities (vision, speech), this demonstrator focuses on text-based interactions to showcase the core GraphRAG architecture and knowledge representation.

​
System Architecture

The system integrates several components into a cohesive educational AI platform:

​
User Interface

A chat application provides the interaction layer, handling text input from students and displaying generated answers with citations. The interface connects to the LLM server and GraphRAG retrieval system to deliver contextual responses.

​
LLM Server

The system runs on local inference servers (Ollama or LM Studio) using the Qwen2.5 model for reasoning and generation. For environments without sufficient local compute, the architecture supports remote API access through OpenRouter while maintaining the same interface.

​
GraphRAG Architecture

GraphRAG System Overview The system implements GraphRAG, which combines retrieval-based and generation-based approaches through a knowledge graph. Unlike baseline RAG systems that retrieve from flat document stores, GraphRAG constructs and queries a structured graph of concepts and relationships. GraphRAG Featurization The implementation is based on nano-graphrag, adapted for educational content. The system operates through several stages: a. Content Chunking: Course materials are segmented into semantically coherent chunks b. Entity and Relation Extraction: An LLM identifies concepts and their relationships within each chunk c. Graph Construction: Nodes and edges are written to a graph store (NetworkX by default, with Neo4j support for scale) d. Community Detection: The graph is clustered into semantic communities with summary reports for each cluster e. Query-Time Retrieval: User questions trigger retrieval of local neighborhoods or community summaries, which are used to generate answers with citations

​
Knowledge Graph Schema

The knowledge graph employs a pedagogically-informed schema designed to support educational interactions:

​
Nodes

  • Concept: (id, title, difficulty, aliases, definitions) - Core learning concepts with metadata
  • Resource: (type ∈ [pdf, slide, video, web], span, timecodes) - Course materials with precise references
  • Example: Worked example snippets demonstrating concept application

​
Edges

  • prereq_of(u → v): Prerequisite relationships forming a directed acyclic graph (DAG) for concept scaffolding
  • explains(resource → concept): Links resources to the concepts they explain
  • exemplifies(example → concept): Connects examples to their target concepts
  • near_transfer(concept ↔ concept): Semantic relationships including siblings, contrasts, is-a, and part-of relations
The near_transfer edges enable the system to assess knowledge transfer. For example, when a student demonstrates understanding of Jensen’s inequality, the system can probe whether that knowledge transfers to the variational lower bound (ELBO), which uses Jensen’s inequality in its derivation. This models pedagogical near transfer - applying learned concepts in related but novel contexts.

​
Data Ingestion Pipeline

The system ingests multiple media sources into a unified knowledge base:
  • Course Website: Structured content, navigation, and article text
  • YouTube Videos: Lecture content with automatic transcription and timecode mapping
  • Slide Decks: Presentation materials with concept extraction
Raw data is stored in MongoDB with metadata tracking for provenance and versioning. The ingestion pipeline maintains URLs and source references for downstream citation generation.

​
Query Processing and Generation

When a user submits a query, the system transforms it into a retrieval problem over the knowledge graph:
  1. Concept Mapping: The query is analyzed to identify candidate concepts
  2. Subgraph Selection: Relevant concepts are expanded into coherent subgraphs containing:
    • Prerequisite chains to scaffold explanations from fundamentals
    • Sibling nodes for generating near-transfer comprehension checks
    • Attached resources with exact spans (page numbers) and timecodes (video segments)
  3. Answer Generation: The LLM synthesizes responses following prerequisite chains, progressing from simpler to more complex concepts
  4. Citation Provision: Each answer includes graph nodes used and resource references

​
Demonstrated Capabilities

The system showcases advanced educational AI through several example queries:

​
Query 1: Attention in Transformers

Question: “What is attention in transformers and can you provide a python example of how it is used?” Response Characteristics:
  • Explains dot-product self-attention mechanism
  • Clarifies the reasoning behind Q (Query), K (Key), V (Value) vectors
  • Provides executable Python code demonstrating attention computation
  • Includes citations to course resources with specific page/slide references

​
Query 2: CLIP Model

Question: “What is CLIP and how it is used in computer vision applications?” Response Characteristics:
  • Explains dual encoders for text and image modalities
  • Describes contrastive loss principle and optimization
  • Details dot-product similarity computation
  • Demonstrates one-shot classification capabilities
  • Links to relevant video timecodes and paper references

​
Query 3: Variational Lower Bound

Question: “Can you explain the variational lower bound and how it relates to Jensen’s inequality?” Response Characteristics:
  • Establishes prerequisite understanding of Jensen’s inequality
  • Explains variational methods and their mathematical foundations
  • Shows how variational autoencoder (VAE) losses employ Jensen’s inequality
  • Follows prerequisite chain from probability theory to VAE implementation
  • Provides worked examples and resource citations

​
Technical Implementation

Each response includes transparency features for educational verification:
  • Knowledge Graph Nodes: Lists specific concepts, resources, and examples retrieved from the graph
  • Resource Citations: Provides precise references (page numbers, slide numbers, video timecodes)
  • System Prompt Visibility: Exposes the system prompt used for generation (questions are not modified)
  • Prerequisite Scaffolding: Visually indicates the concept chain used to structure the explanation

​
Applications

Erica demonstrates practical applications of GraphRAG in education:
  • Adaptive Tutoring: Responses adjust to prerequisite knowledge and concept difficulty
  • Resource Discovery: Students receive precise pointers to relevant course materials
  • Knowledge Transfer Assessment: System identifies opportunities for near-transfer evaluation
  • Concept Navigation: Graph structure enables exploration of related concepts and dependencies
  • Scaffolded Learning: Explanations follow pedagogically sound progression from fundamentals

​
Technical Stack

The demonstrator showcases modern AI infrastructure for education:
  • Knowledge Graph: NetworkX (local) or Neo4j (production scale)
  • Vector Store: Embedding-based retrieval for concept mapping
  • LLM: Qwen2.5 for reasoning and generation
  • Ingestion: Multi-format processing (web, video, PDF, slides)
  • Storage: MongoDB for raw content and metadata
  • Interface: Chat-based interaction layer
This implementation demonstrates how knowledge graphs can enhance RAG systems for domain-specific applications where semantic relationships and prerequisite structures matter, particularly in educational contexts where scaffolding and knowledge transfer are central to effective learning.
Edit this page on GitHub or file an issue.