Adding a file or a document to your conversation gives the model access to it, but it doesn't automatically set the limits. There is a difference between context and grounding that really matters here. Context is just a a vision. When you paste a contract, a report or a policy document into a chat, the model can see it. But it's also using everything it was trained on, patterns from before it was ever released. And without grounding, it blends both. If something is missing from your document, the model doesn't stop.
It quietly fills the gaps using its training. And the output sounds exactly like it uses just the data from the attached document. You cannot hear where the information came from. Grounding is what adds a boundary. It is an instruction to AI model to stay within a defined source. In order to do this, together with the document, you should just give a couple of additional sentences. The first one. Answer only based on the information from the document attached to this conversation. And the second one. If the answer is not in the text, say you don't know.
The first one creates the boundary. The second changes the behavior when something is missing. Because by default, the model always tries to produce an answer. Without that second line, it quietly fills the gaps from its training. With it, it can just stop. Grounding is not about making AI correct. It is about making the answer traceable to a source and also making sure that AI doesn't add additional information from training. One important note to make here. If your document is too large, the model may not see all of it at once.
And if your document is not real text but an image like a scanned PDF, or maybe if it contains a lot of diagrams or tables, the model may be at least partially blind to it. Only what the model can actually see can be used as a boundary. Grounding is not just giving AI a document. It's telling AI where the boundaries are and making sure that those boundaries are respected. This was a note on AI. Thank you for watching and I'll see you in the next one.
