Architecture Diagrams
A collection of Mermaid diagram sources (.mmd) covering machine-learning system design, retrieval-augmented generation, deployment architecture, ETL workflows, and game-state machines.
ML System Architecture
ch1-system-architecture.mmd — End-to-end ML System
flowchart TD: Raw Data Sources → Feature Pipeline → Vector Store
Training Data → Base Model → Inference Pipeline → API Endpoints → Users
Top-down flowchart of a complete ML system: feature ingestion, vector storage, model fine-tuning, inference serving, and API delivery.
ch3-etl-workflow.mmd — ETL Workflow
Extract-transform-load pipeline connecting raw data sources through transformation stages to a downstream data store.
ch4-rag-architecture.mmd — RAG Architecture
Retrieval-augmented generation pipeline: document chunks are embedded and stored in a vector database; at query time the query is embedded, relevant context is retrieved, and an LLM generates the final response.
ch5-training-flow.mmd — Model Training Flow
Training pipeline stages from raw dataset preparation through model training, evaluation, and checkpoint export.
ch9-advanced-rag.mmd — Advanced RAG
Extended RAG architecture incorporating re-ranking, hybrid retrieval, or multi-hop reasoning steps beyond the basic embed-retrieve-generate loop.
Deployment and Operations
ch10-deployment-arch.mmd — Deployment Architecture
Production topology: Client → Load Balancer → API Gateway → Auth Service; Inference Service queries a Vector Store and an LLM Service; results are logged to a Monitoring DB.
ch11-cicd-pipeline.mmd — CI/CD Pipeline
Left-to-right flowchart: Code push triggers a Git repo, CI pipeline builds a Docker image, image is pushed to a registry, and the model service is deployed with metrics and alerting.
Game State Machines
pushcases.mmd — Sokoban Push Cases
State diagram modelling the two push variants in a Sokoban-style puzzle: hitting a box centre triggers a single-box push; hitting a box edge triggers a double-box push. Each path resolves to Success or Blocked.
wideflow.mmd — Move Validation Flow
Decision tree for processing a robot move: wall-collision check, box-contact check, box-type branch (single vs double), push-success check, and final state update or invalid-move resolution.
Viewing the Diagrams
The .mmd files can be rendered with the Mermaid CLI:
npx -p @mermaid-js/mermaid-cli mmdc -i ch1-system-architecture.mmd -o ch1-system-architecture.svg
Or viewed directly on mermaid.live by pasting the file contents.