A field guide to the vocabulary and the tools — and when each one actually earns its place
🧠ModelConcept
The trained neural network itself — a fixed set of weights that turns input into output. Stateless: no memory, no agency, no tools of its own.
Use when you need raw language/reasoning capability behind an API or app.
Not for expecting it alone to remember past sessions, browse the web, or take action — that requires scaffolding on top.
📚LLMConcept
"Large Language Model" — a model trained on huge amounts of text to predict and generate language.
Use when the task is language-shaped: writing, summarizing, explaining, coding, classifying text.
Not for tasks needing guaranteed exact facts or arithmetic without verification — pair it with a calculator/lookup tool instead of trusting raw output.
💬ChatbotConcept
An interface that wraps a model in back-and-forth conversation. You ask, it answers. No independent action.
Use when the job is knowledge transfer — you ask a question, you do the work yourself.
Not for multi-step tasks that need the AI to actually execute things (edit files, run code, call APIs) — that's an agent's job.
🤖AgentConcept
A model given tools, a loop, and a goal. It plans, acts, observes results, and repeats until the goal is met — instead of just answering.
Use when the task breaks into concrete tool calls with a checkable end state (fix the bug, deploy the change, gather data).
Not for a quick one-off question (slower, costlier than a chat reply) or high-stakes irreversible actions without a human checkpoint.
🦾BotConcept
The generic, older term for any automated program that performs actions — with or without AI. A scripted Slack reminder is a bot; so is an AI agent.
Use when describing simple, rule-based automation that reacts to fixed triggers.
Not for claiming intelligence it doesn't have — a rule-based bot with no model behind it isn't "AI," it's a script.
✍️PromptConcept
The text/instructions you feed the model to steer its output.
Use when refining wording, examples, and structure to get more reliable output (prompt engineering).
Not for fixing a fundamentally wrong model or missing knowledge — no prompt rescues a model that doesn't have the facts. Also never put secrets in a prompt; assume it's logged.
🔤TokenConcept
A chunk of text (roughly a word or word-piece) — the unit models actually process, and the unit most APIs bill by.
Use when estimating cost or how much text fits in a request.
Not a reliable 1:1 stand-in for "words," especially across languages — budget with real token counts, not word counts.
🪟Context WindowConcept
The total amount of text (in tokens) a model can "see" at once — prompt, conversation history, and output combined.
Use when deciding whether a document fits directly in the prompt.
Not for assuming unlimited memory — once content falls outside the window it's gone; for large corpora, use retrieval instead of stuffing everything in.
🎛️Fine-tuningConcept
Further training a base model on your own labeled data so it specializes in a tone, format, or domain.
Use when you have a high-volume, repetitive task and enough quality labeled examples to justify the cost.
Not for one-off tasks, small datasets, or knowledge that changes often — try prompting or RAG first; they're cheaper and faster to iterate on.
🔍RAGConcept
"Retrieval-Augmented Generation" — fetching relevant documents at query time and feeding them into the prompt, instead of retraining the model on that knowledge.
Use when you need current or proprietary knowledge, and want traceable citations.
Not for small, static knowledge that fits directly in a prompt, or latency-critical paths where an extra retrieval hop isn't affordable.
🧭EmbeddingConcept
A numeric vector representing the "meaning" of a piece of text or data, so similar meanings sit close together in vector space.
Use when building semantic search, recommendations, or de-duplication.
Not for exact/structured lookups (order IDs, prices) — a regular database index is faster and exact for that.
⚖️QuantizationConcept
Compressing a model's weights to lower precision (e.g. BF16 → Q4) to shrink memory and speed up inference, at some accuracy cost.
Use when running a model locally on limited VRAM/RAM, or serving cheaply at scale.
Not for tasks that lean on maximum precision — heavy multi-step reasoning or math can degrade more than casual chat.
🔓Open-Weight vs. Closed ModelConcept
Open-weight models (Llama, Mistral, Qwen) can be downloaded and run yourself. Closed models (GPT, Claude, Gemini) are only reachable through a provider's API.
Go open-weight when you need data to stay on your own hardware, offline capability, or no per-token bill.
Go closed when you want the strongest available capability and don't want to run or maintain infrastructure yourself.
👻HallucinationConcept
When a model confidently states something false or made-up, with no signal that it's guessing.
Reduce it with RAG/citations, lower temperature, and a human or automated check on high-stakes claims.
Don't treat raw ungrounded LLM output as a sole source of truth for legal, medical, or financial facts.
🔐System PromptConcept
Hidden instructions set before the conversation starts, defining the model's persona, rules, and boundaries.
Use it to keep behavior consistent across a whole product or session.
Not a hard security boundary — a determined user can attempt to override it via prompt injection, so don't rely on it alone to gate sensitive actions.
🦙OllamaLocal Runtime
A CLI/API for downloading and running open-weight LLMs on your own machine with one command.
Use for private or offline inference, fast local prototyping, avoiding per-token cost.
Not for serving many concurrent users at production scale, or when you need frontier-level capability that open-weight models don't yet match.
🖥️LM StudioLocal Runtime
A desktop app with a GUI for browsing, downloading, and chatting with local open-weight models — Ollama with a point-and-click front end.
Use for non-technical users who want a local chat UI without touching a terminal.
Not for scripting or automating at scale — its CLI/API story is thinner than Ollama's.
💬ChatGPTChatbot
OpenAI's consumer chatbot, built on GPT models, with optional browsing, image generation, and custom GPTs.
Use for general Q&A, writing help, brainstorming, quick research.
Not for autonomous multi-file coding or system changes — the chat itself doesn't act on your machine (their separate agent/Codex-style products do).
✨claude.aiChatbot
Anthropic's consumer chatbot, built on Claude models — strong at long-document analysis and careful, nuanced writing.
Use for reasoning through long or complex material, drafting, careful analysis.
Not for real-time facts without web search enabled, or actions outside the chat window — it talks, it doesn't act on your systems.
👨💻Claude Code / Copilot / CursorCoding Agent
AI coding assistants wired into your editor or terminal with access to your files, tests, and often git — able to read, edit, and run code across a whole repo.
Use for multi-file refactors, debugging with real test feedback, repo-aware changes.
Not for non-coding tasks, or letting it push/deploy to production without human review on anything that matters.
🤗Hugging FaceModel Hub
A hub for open model weights, datasets, and the libraries (like transformers) used to run and fine-tune them.
Use for discovering/downloading open models, running research experiments, sharing demos via Spaces.
Not a production inference platform by itself — plan your own serving layer for real traffic.
🔗LangChain / LlamaIndexFramework
Libraries for chaining LLM calls together and building RAG pipelines or agents, with pre-built integrations for common tools and data sources.
Use for quickly prototyping a multi-step pipeline using existing integrations.
Not for a single simple API call — the abstraction is overhead you don't need; many production teams drop it later for direct API calls when they need tighter control.
🗂️Vector DatabasesInfrastructure
Databases built to store and search embeddings by similarity — Pinecone, Chroma, Weaviate, pgvector, and similar.
Use for semantic search or RAG over large document sets.
Not for small datasets that fit in memory (skip the infra), or exact/structured queries — a normal database index handles those better.
🎨Midjourney / Stable Diffusion / DALL·EImage Gen
Image-generation models that turn text prompts into original images.
Use for concept art, mockups, marketing visuals, creative exploration.
Not for exact likenesses of real people, reproducing protected brand assets, or technical diagrams that need to be precisely accurate.
🌀Autonomous Agent FrameworksFramework
Tools like AutoGPT-style loops that let a model set its own sub-goals and run with minimal supervision toward an open-ended objective.
Use for research demos and exploring what open-ended autonomy can do.
Not for production without strict guardrails — prone to unproductive loops and runaway API costs without a human checkpoint or hard step limit.
⚖️ Local (Ollama/LM Studio) vs. Cloud (ChatGPT/Claude/API)