Tag Vocabulary

The complete controlled vocabulary for Vault_2026. See Conventions §12 for rules and application guidelines.

Key rules: kebab-case only · ≤4 tags per note · no structural redundancy · no tool names · index files always tags: []

---

Dimension 1 — Concept Type

Tags that classify what kind of knowledge a note contains.

Tag	Definition	Use when	Not for
algorithm	A named, step-by-step computational procedure	A note defines or analyzes a specific algorithm (backpropagation, k-means, ARIMA)	Generic topic notes; notes named after the algorithm already convey this
pattern	An architectural or design pattern	A note describes a reusable solution structure (RAG pipeline, feature store, training pipeline pattern)	One-off implementations; notes that just happen to follow a pattern
theory	Mathematical or statistical theory	A note proves or formally defines a theoretical concept (PAC learning, VC dimension, convexity)	Intuitive overviews; applied engineering notes
workflow	An operational process or practice	A note describes a repeatable process (experiment tracking, CI/CD pipeline, data labeling workflow)	Single-tool reference notes; conceptual introductions

---

Dimension 2 — ML Lifecycle Stage

Tags that mark where in the ML/AI production lifecycle a note is relevant. Do not apply if the note’s folder already encodes the stage (e.g., don’t add training to notes already inside 05_ml_engineering/05_model_development/).

Tag	Definition	Use when	Not for
data	Data collection, ingestion, preprocessing, validation, augmentation	A note outside the data engineering sublayer that is primarily about data concerns	Notes inside 05_ml_engineering/02_data_engineering/
feature-engineering	Feature creation, encoding, selection, feature stores	A note outside the feature engineering sublayer that covers feature transformation	Notes inside 05_ml_engineering/04_feature_engineering/
training	Model fitting, optimization, regularization, hyperparameter tuning	A foundations or modeling note that is specifically about the training process (not just theory)	Notes inside 05_ml_engineering/05_model_development/
evaluation	Metrics, benchmarking, cross-validation, calibration, model validation	Any note that is primarily about measuring model quality	Generic overview notes
deployment	Serving, inference, containerization, rollout strategies, model compression	Notes about putting models into production	Notes inside 05_ml_engineering/06_deployment_and_serving/
monitoring	Drift detection, alerting, observability, retraining triggers	Notes about tracking model health post-deployment	Notes inside 05_ml_engineering/07_monitoring_and_observability/

---

Dimension 3 — Domain / Modality

Tags that mark what kind of data or problem domain a note addresses. Do not apply if the folder already encodes the domain (e.g., don’t add nlp to notes inside a transformer or NLP sublayer).

Tag	Definition	Use when	Not for
llm	Large language model systems	A note is about LLM-specific concerns (prompting, fine-tuning, inference, evaluation of LLMs) and is not already in 06_ai_engineering/ — or is in 06_ai_engineering/ but also highly relevant from an MLOps/engineering angle	Generic NLP notes; notes already inside 06_ai_engineering/ where llm is the default domain
nlp	Natural language processing tasks not specifically LLM-centric	Classical NLP, tokenization theory, text preprocessing, sentiment analysis, named entity recognition	LLM-specific notes (use llm)
vision	Computer vision, image understanding	Notes about CNNs, object detection, image segmentation, visual models	Notes inside a CV sublayer already
time-series	Temporal data, forecasting, sequential modeling of time	ARIMA, Prophet, state-space models, walk-forward validation	General sequence models (use nlp or algorithm)
tabular	Structured/relational data, classical ML	Tree ensembles, logistic regression, feature engineering for tabular data	Deep learning notes on tabular data (use algorithm + tabular)
multimodal	Cross-modal systems (vision+language, audio+text, etc.)	CLIP, BLIP, vision-language models, audio captioning	Single-modality notes
recommendation	Collaborative filtering, ranking, retrieval for recommender systems	Matrix factorization, two-tower models, CTR prediction	General retrieval notes (use retrieval)

---

Dimension 4 — Capability / Task Type

Tags that mark what the system or model does.

Tag	Definition	Use when	Not for
classification	Supervised classification	Notes about classifiers, classification loss, ROC/AUC, multiclass strategies	Regression notes
regression	Supervised regression	Notes about regression models, MSE/MAE, prediction intervals	Classification notes
clustering	Unsupervised grouping	k-means, DBSCAN, GMM, hierarchical clustering	Dimensionality reduction (unless combined)
generation	Generative models and text/image/audio generation	LLMs, diffusion models, VAEs, GANs, seq2seq generation	Discriminative models
retrieval	Search, nearest-neighbor, RAG, vector databases	Embedding search, FAISS, Chroma, BM25, hybrid search, reranking	Pure generation without retrieval
reasoning	Chain-of-thought, planning, agents, step-by-step inference	ReAct agents, DSPy, tool-calling, multi-step reasoning	Straightforward inference
forecasting	Time-series prediction	ARIMA, Prophet, temporal models predicting future values	General regression (use regression)
anomaly-detection	Outlier and novelty detection	Isolation Forest, one-class SVM, statistical anomaly tests	Classification of rare events (use classification)
interpretability	Explainability, feature attribution, fairness, model transparency	SHAP, LIME, PDP, fairness metrics, model cards	Evaluation notes (use evaluation)

---

Dimension 5 — Infrastructure / System Concern

Tags that mark cross-cutting production system concerns.

Tag	Definition	Use when	Not for
mlops	ML production lifecycle operations	Notes about the operational concerns of running ML systems (pipelines, registries, orchestration) that appear outside 05_ml_engineering/	Notes inside 05_ml_engineering/ (structural redundancy)
llmops	LLM-specific production operations	Notes about operating LLM systems in production (model gateways, prompt management, LLM observability) that appear outside 06_ai_engineering/	Notes inside 06_ai_engineering/08_architecture_and_feedback/
distributed	Multi-GPU, multi-node, parallelism, sharding	Notes about distributed training or inference	Single-GPU notes
quantization	Model compression via reduced precision (INT8, INT4, GGUF, AWQ, GPTQ)	Notes about quantization methods and trade-offs	General compression/pruning (use deployment)
fine-tuning	Adapting pretrained models: LoRA, PEFT, RLHF, SFT, DPO, GRPO	Notes about fine-tuning methods and workflows	Pretraining from scratch; general transfer learning
safety	Alignment, guardrails, content moderation, red-teaming, constitutional AI	Notes about making AI systems safe and aligned	General evaluation (use evaluation)

---

Synonym → Canonical Mapping

Use this when migrating old tags:

Old tag(s)	Canonical tag	Reason
finetuning, sft, lora, qlora, peft, rlhf, dpo, grpo, adapter, adapters	fine-tuning	All refer to the same lifecycle concern
serving, model-serving, ml-serving, inference-api, production-ml	deployment	All refer to production serving
metrics, benchmarks, benchmarking, offline-evaluation, llm-evaluation, model-validation, llm-eval	evaluation	All about measuring quality
concept-drift, data-drift, distribution-shift, drift, drift-detection	monitoring	All about production health
vector-store, vector-database, vector-db, vector-search, similarity-search	retrieval	All vector search patterns
ranking, reranking, re-ranking	retrieval	Ranking is part of the retrieval pipeline; use retrieval
deep_learning, deep-learning, neural-networks, neural_networks	remove (structural)	Redundant with 01_foundations/06_deep_learning_theory/
mlops, ml-ops, ml-pipelines, ml-systems, ml-lifecycle	mlops	When outside 05_ml_engineering/; otherwise remove
computer_vision	vision	Canonical form
sequence_models, rnn, lstm, gru, BPTT	remove or nlp	Structural (inside transformer/RNN sublayers)
transformers, transformer, attention, bert, gpt	remove (structural)	Inside transformer sublayers already
python, go, typescript, javascript	remove	Inside language-specific sublayers
end-to-end	remove	Folder 08_implementations/02_end_to_end_examples/ encodes this
production, production-ml	remove or deployment	Usually structural redundancy