Glossary
All key concepts you’ll encounter in the local AI world.
Levels:survival →
comfortable reading →
advanced beginner.
[🇷🇺 Russian version: glossary.ru.md]
Level 1: Survival (10 Terms)
Must know to understand 90% of articles and conversations.
| # | Term | Explanation | Related Section |
|---|---|---|---|
| 1 | LLM (Large Language Model) | Large Language Model — a program that generates and understands text. ChatGPT, Qwen, Llama — all are LLMs. | what-is-ai.md |
| 2 | Prompt | Your request to the model. What you type in chat. “Write a Python function” — that’s a prompt. | ../agents/prompting.md |
| 3 | Token | “LEGO brick of text”. A word can be 1-3 tokens. Model works with tokens, not letters. | how-models-work.md |
| 4 | Parameters (B) | Model’s “neurons”. 7B = 7 billion parameters. More = smarter, but heavier. | what-is-ai.md |
| 5 | Inference | Process of getting answer from model. You asked → model answered — that’s inference. | what-is-ai.md |
| 6 | Open-source / Open Model | Model whose files you can download and run locally. Qwen, Llama, Mistral — open-source. | what-is-ai.md |
| 7 | Ollama | Main program for running models locally. One command — and model works. | ../local-models/getting-started.md |
| 8 | Quantization | Model compression to save memory. Q4 = model weighs 4x less, quality almost same. | ../local-models/quantization.md |
| 9 | GGUF | Model file format for local running. All models on Ollama — in GGUF. | ../local-models/running-models.md |
| 10 | Local | Everything runs on your computer, no internet, data never leaves. | cloud-vs-local.md |
Level 2: Comfortable Reading (15 Terms)
Needed to understand technical articles and documentation.
| # | Term | Explanation | Related Section |
|---|---|---|---|
| 11 | Context Window | How many tokens model “remembers” at once. 8K = ~5000 words. More = better, but needs more memory. | how-models-work.md |
| 12 | Temperature | Model “creativity” setting. 0 = conservative (for code), 1 = creative (for writing). | how-models-work.md |
| 13 | System Prompt | Instruction for model on how to behave. E.g.: “You are an experienced Python developer. Answer concisely and to the point.” | ../agents/prompting.md |
| 14 | Fine-tuning | Additional training of ready model on your data. Cheaper than training from scratch. | Rare locally, mentioned in what-is-ai.md |
| 15 | GPU / Video Card | Hardware accelerating model work. Optional but desirable. Apple Silicon — also GPU. | hardware-guide.md |
| 16 | VRAM | Video card memory. Main resource for large models. More VRAM = larger model you can run. | hardware-guide.md |
| 17 | Hallucination | When model confidently outputs incorrect fact. Not maliciously — just “imagined” most probable continuation. | how-models-work.md |
| 18 | API | Programmatic interface — way to call model from code. Ollama compatible with OpenAI API. | ../local-models/advanced-setup.md |
| 19 | RAG (Retrieval-Augmented Generation) | “Search + answer”: model first searches info in your documents, then answers based on it. | ../use-cases/rag.md |
| 20 | Embedding | Numerical representation of text meaning. Allows searching similar texts by meaning, not words. | ../use-cases/rag.md |
| 21 | FIM (Fill-in-the-Middle) | Code autocomplete. Model looks at code before and after cursor and inserts what should be in middle. | ../use-cases/coding.md |
| 22 | Tool Calling (Function Calling) | Model’s ability to call functions. Allows model to run code, search web, work with files. | ../agents/ollama-for-agents.md |
| 23 | Self-attention / Attention | Mechanism allowing model to connect words in sentence. Thanks to it, model understands “it” refers to “animal”. | how-models-work.md |
| 24 | Transformer | Neural network architecture all modern LLMs are built on. Invented by Google in 2017. | how-models-work.md |
| 25 | num_ctx | Ollama parameter setting context window size. num_ctx: 8192 = 8K context. |
../local-models/memory-and-context.md |
Level 3: Advanced Beginner (10 Terms)
Will be needed when you start configuring agents and digging deeper.
| # | Term | Explanation | Related Section |
|---|---|---|---|
| 26 | Prompt Engineering | Art of crafting requests properly to get best answers from model. | ../agents/prompting.md |
| 27 | Tokens per Second (t/s) | Model speed. 10+ t/s — comfortable for chat, 30+ — for autocomplete. | ../local-models/benchmarks/apple-silicon.md |
| 28 | ReAct (Reasoning + Acting) | Agent pattern: “think → act → observe result → think again”. | ../agents/patterns.md |
| 29 | AI Agent | Model that decides which tools to call and in what order to complete a task. | ../agents/README.md |
| 30 | Multi-agent | Multiple agents working together: one plans, second writes code, third tests. | ../agents/multi-agent.md |
| 31 | Top-k, Top-p | Additional parameters affecting model creativity (like temperature). | ../agents/prompting.md |
| 32 | KV Cache | Cache model uses to speed up responses. Takes lots of memory with large context. | ../local-models/memory-and-context.md |
| 33 | MoE (Mixture of Experts) | Model architecture where part of “experts” activates only for relevant task. Saves resources. | ../local-models/catalog.md |
| 34 | Vector Database | Embedding storage for fast semantic search. Used in RAG. | ../use-cases/rag.md |
| 35 | Sliding Window Attention | Old context window type (limited length, Infinity). Rarely used in modern models. | — |
How to Remember All Terms
Don’t try to memorize everything at once. Read handbook sections in order — each term will appear in context and stick naturally.
Tip: when you encounter an unfamiliar term in an article — come back here. Over time, all will be memorized.
What’s Next
| If You Want To | Go To |
|---|---|
| Frequently Asked Questions | faq.md |
| Step-by-step Learning Plan | learning-path.md |
| Install Ollama | ../local-models/getting-started.md |
| Back to Navigation | README.md |
In section: what-is-ai · how-models-work · cloud-vs-local · hardware-guide · glossary · faq · learning-path · setup-windows · setup-linux
Related sections: Local Models · AI Agents · Use Cases
Navigation: ← Zero Level · ↑ Back to main · 🇷🇺 Русский