Glossary
Words we actually use
Short original definitions. If a vendor uses the same word differently, trust their docs for their product.
- Model
- The trained statistical engine. Products wrap models with chat UI, tools, and billing. A model name is not a product name.
- Token
- A chunk of text the model reads or writes. Billing and context limits are usually counted in tokens, not words.
- Prompt
- The instructions and material you give the model for one request, including hidden system instructions added by a product.
- Context window
- How much token space one request can hold: instructions, conversation, retrieved text, and the reply.
- Hallucination
- A fluent statement that is not supported by the input or by a checkable source. Treat it as a default risk, not a rare bug.
- Open-weight
- Model files you can download (license permitting) and run yourself. Open-weight is not the same as “free to use for any purpose.”
- Inference
- Running a trained model to get an output. Training creates the weights. Inference is the everyday call.
- Agent
- Software that loops: it may call tools, read results, and continue. An agent is a workflow, not a personality.
- MCP
- Model Context Protocol — an open way to connect assistants to tools and data sources. The protocol is not a single vendor’s app.
- RAG
- Retrieval-augmented generation: fetch relevant snippets first, then have the model write with those snippets in context.