The Grounded-Answer Prompt

stuffing docs into context

Part of: FAQ Support Bot

The FAQ is numbered. Now wire it into a real prompt: the smallest version of the bot that answers a question using only the FAQ text you hand it. This is "stuffing" in the literal sense. You put the whole document into the system prompt. What context stuffing means Context stuffing is the simplest way to ground a model in your own data. Paste the entire source document into the prompt, then ask the question. No search step, no database, no embeddings. You hand over everything and let the model read it. This only works because your FAQ is small enough to fit in one call, which is the scope of this project. Building the system prompt The system prompt carries two things: the numbered FAQ text, and strict instructions about how to use it. The FAQ lives in system because it's the standing reference material for every question in the session, not a one-off message. The user's actual question goes in messages. It's the same split you've used in every project: config on one channel, live input on the other. Why the wording of the rules matters "Answer using the FAQ" alone is too soft. The model still reaches for its general knowledge when the FAQ is thin. Two things tighten it. First, "ON

Challenge: Stuff the Context Budget