List of MicroSims for the CKG Benchmark

Interactive diagrams and workflows that visualize the three knowledge retrieval architectures compared in this benchmark.

• Architecture Comparison

  

  Interactive side-by-side diagram of all three retrieval pipelines (RAG, GraphRAG, CKG) in a single view. Hover any component for a description of its role.

• Corpus Provenance

  

  Upstream pipeline that produces the inputs to the three retrieval architectures. The learning-graph.csv seeds the textbook markdown corpus, which RAG and GraphRAG consume; CKG consumes the CSV directly.

• CKG Workflow

  

  Build-time and query-time pipeline for Compact Knowledge Graph retrieval. Parses a pre-authored DAG directly and extracts the relevant subgraph by BFS or DFS traversal, yielding roughly 150 to 400 tokens of context per query.

• GraphRAG Workflow

  

  Build-time and query-time pipeline for GraphRAG. Uses an LLM to extract an entity graph and community summaries from source text, then routes each query to local or global retrieval for roughly 2,000 to 8,000 tokens of context.

• RAG Workflow

  

  Build-time and query-time pipeline for Retrieval-Augmented Generation. Chunks markdown text, embeds each chunk, and retrieves the top-five most similar chunks per query, yielding roughly 2,400 tokens of context.