The five kinds
of data
Every answer an AI gives you is drawn from one of five sources. Knowing which one you are leaning on, and how far to trust it, is most of the game.
Scroll, click, or use the arrow keys to move through.Scroll, or use the arrows and dots to move through.
Five sources, five very different trust profiles. One at a time.
Training data
What the model already knows. You can't inspect it, can't change it, can't date it.
What the model learned during training. It is fixed: it cannot be inspected, updated, or dated. It supplies the model’s general knowledge, and it is also the source of out-of-date answers.
Your prompt and context
What you hand it directly: the question, the files, the pasted text. The AI always gets this.
The question, instructions, and files you provide in the session. The AI always gets all of it: everything you put here reaches the model, in full, every time. You control this layer completely and can see exactly what the model is working from. Thin context is the most common cause of poor answers.
Connected data
Your systems, reached through a connector or governed access. The AI chooses what to access.
Data reached live from your own systems, the CRM, the warehouse, the file store through a connector with permissions. Unlike your prompt, none of it arrives automatically: you grant the access, and the AI chooses what to reach for within it, so what it actually sees is the slice it decided to pull. Its reliability is the reliability of the source system, the scope of access you grant and the method of query.
Web data
Open retrieval. Anything goes.
Content retrieved from the open internet at answer time. Source and quality vary, and the model does not always show which pages it relied on. Useful for current information; verify it before you depend on it.
Memory
What persists across sessions. Typically managed by your agent and harness. Only as good as the tools.
Information carried across sessions, written and recalled by the agent and harness. Its quality depends on the tools that manage it: maintained well it compounds, unmaintained it accumulates errors.