Skip to main content

Knowledge Engine (KE)

Overview​

The Knowledge Engine is the foundational layer of the EduVision architecture. It serves as the system's "Brain," responsible for ingesting, structuring, and retrieving educational content. It transforms unstructured data (PDFs, text, slides) into a structured Knowledge Graph (KG) that can be navigated by the other engines.

Core Capabilities​

1. Semantic Search (RAG)​

The KE utilizes a vector database (e.g., ChromaDB, FAISS) to perform semantic searches. When a student asks a question, the KE retrieves the most relevant "chunks" of information based on embedding similarity, not just keyword matching.

  • Embedding Model: sentence-transformers/all-MiniLM-L6-v2 (optimized for speed/accuracy trade-off).
  • Vector Database: Used for dense retrieval.

2. Knowledge Graph Construction​

Beyond simple text chunks, the KE builds a graph of concepts.

  • Nodes: Concepts (e.g., "Loop", "Variable", "Recursion").
  • Edges: Relationships (e.g., "requires", "is-a", "part-of").

This graph structure allows the Pedagogy Engine to understand prerequisites. For example, knowing that "Recursion" requires "Functions" allows the system to enforce learning paths.

3. Automatic Chunking​

The engine intelligently splits long documents into pedagogically sound units.

  • Strategy: Recursive Character Splitting + Semantic Boundary Detection.
  • Metadata: Each chunk is tagged with its source, difficulty level, and associated learning objectives.

Technical Implementation​

class KnowledgeEngine:
    def __init__(self, vector_store, graph_store):
        self.vector_store = vector_store
        self.graph_store = graph_store

    def ingest(self, document):
        chunks = self.chunker.split(document)
        embeddings = self.embedder.embed(chunks)
        self.vector_store.add(chunks, embeddings)
        self.graph_builder.extract_entities(chunks)

Integration with Other Engines​

  • Pedagogy Engine: Queries the KE to find the next concept in the dependency graph.
  • Tutor Engine: Uses retrieved context from the KE to ground its responses (reducing hallucinations).
  • Assessment Engine: Uses KE to generate relevant questions based on the specific content taught.