Prompt Grounding
Constraining a model's output to a defined, verified source of truth, rather than open-ended generation.
Prompt grounding means constraining what a model is allowed to draw on when it generates, tying its output to a defined, verified source of truth instead of letting it generate freely from whatever it absorbed in training. The source of truth is typically inserted directly into the model's context, so it's answering from that specific material rather than from general pattern-matching over similar text it's seen before.
In SwishX's pipeline, the brand dossier is that source. A request's context is harnessed by merging it with the dossier's approved claims and regulatory ruleset before generation happens, and every claim the resulting prompt implies is then checked against that same set before anything renders. A claim with no matching entry in the dossier has nothing to be grounded in, which is the reason it gets rejected rather than generated and caught later.