weights
The numbers that make up the model file, left by training. They do not change when a prompt is sent.
context window
How much of the current conversation the model can hold at once. Earlier turns past that limit are gone.
hallucination
A fluent span with no source in the weights, the window, or a fetch. The sentence can still be well-formed.
In: 01 What the model knows and doesn't 05 Provenance: retrieved, computed, generated
retrieval
Fetching text — a PDF, a search hit, a row — and putting it in the window for the model to read. The model does not go get it.
RAG
A common way to do retrieval automatically: search, then put the hits into the prompt. The model is still reading. It is not acting on an external system.
tool call
The model stops generating text and invokes a function that runs somewhere else. That function may only look, or it may change a record.
In: 03 Tools: letting it do something 06 It looked up the flight. It booked the flight.
MCP
An agreed shape of plug for tools and data sources. It is the connector, so a tool built once can be reused. It is not itself a capability.
In: 04 MCP: the standard plug 03 Tools: letting it do something
provenance
Which spans were fetched, which were calculated, and which were generated. Those are different kinds of claim even when the tone is the same.
agent
The same model, allowed to choose the next tool call for a while without a person between steps. Looking something up is not the same as doing it.
In: 06 It looked up the flight. It booked the flight. 03 Tools: letting it do something
data center
The building: rows of machines, a large electrical load, and a large share of floor space whose job is moving heat outside.