NOTES / READING LIST AND GLOSSARY
Glossary
Short plain-language definitions of terms used across these notes, each pointing to the note that explains it.
UPDATED 18 NOV 2026#glossary
Definitions are my own short versions; the linked notes give detail and sources.
- Agent: a system where the model directs its own steps and tool use.1 What is an agent.
- Workflow: LLM calls arranged along a fixed, predefined path.1 Workflows vs agents.
- Harness: everything around the model: tools, loop, memory, permissions. Execution loop.
- Context window: the text a model can read in one call. Context as a budget.
- RAG: retrieving documents and giving them to the model to answer from.2 RAG overview.
- Embedding: a vector representing meaning, used for similarity search. Embeddings.
- BM25: a keyword-based ranking function. BM25.
- KV cache: stored attention keys and values that avoid recomputation. KV cache.
- Quantisation: storing weights in fewer bits. Quantisation.
- LoRA: fine-tuning with small low-rank adapter matrices.3 LoRA and QLoRA.
- Eval: a repeatable test of system quality. Why evals.
- LLM-as-judge: using a model to grade outputs. LLM as judge.
- Prompt injection: untrusted text that tries to steer the model. Prompt injection.
- Guardrail: a check on inputs or outputs. Guardrails.
- MCP: an open protocol for connecting models to tools and data. MCP.
- De-identification: removing identifiers from health data. Privacy and regulation.
Sources
- [1]Building effective agents — Schluntz and Zhang · Anthropic · 2024PARAPHRASED
- [2]Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks — Lewis et al. · 2020PARAPHRASED
- [3]LoRA: Low-Rank Adaptation of Large Language Models — Hu et al. · 2021PARAPHRASED