Hallucination guard prompt builder
LLMs are fluent enough to state a fabricated statistic as confidently as a true one. The most effective defense in a prompt is a set of explicit grounding rules: cite only what you were given, flag what you are unsure about, and refuse to invent specific numbers, dates, names, or quotes. This builder appends those rules to any prompt, tuned to your domain, your source situation, and how strict you want to be.
How it works
You paste your base prompt and tell the tool whether the model is working from provided sources, its own knowledge, or a mix. Each setting injects the right grounding rules — citation requirements for sourced answers, uncertainty labeling for knowledge-only answers, and clear separation rules for the mixed case. The strictness control then layers on harder constraints, up to tagging every sentence as supported or unverified and refusing to answer when the model cannot do so reliably.
The three source modes and what they change
Provided sources only is the mode for retrieval-augmented generation, where you pass context documents alongside the question. The appended rules instruct the model to cite the passage or source for every claim, explicitly say when the provided context does not contain the answer (rather than guessing), and refuse to use any knowledge from its training that is not corroborated by the provided material. This is the strictest form of grounding and the most effective defense against hallucination — when the model cannot drift to training data, the scope of possible fabrication shrinks dramatically.
Own knowledge only is the mode when you are asking a question without providing context documents. The grounding rules shift to uncertainty labeling: the model must flag when a claim is from memory and likely to be outdated, hedge properly when discussing events near its knowledge cutoff, and refuse to state a specific date, figure, name, or quote without a high degree of internal confidence.
Mixed is the mode when you have both provided context and may want the model to supplement it with general knowledge. The rules in this mode enforce a clear separation: claims from provided sources are cited, claims from model knowledge are labeled as such, and claims that mix both sources are broken apart so the reader can tell which part comes from which origin.
Why specific data points are the highest risk
Numbers, dates, names, and verbatim quotes are the categories of information most likely to be hallucinated confidently and the most damaging when wrong. A model inventing a rough general claim (“this is a common issue”) is less harmful than a model inventing a specific statistic (“this affects 43% of enterprises”) that will be cited, acted on, or published. The strictness settings in this builder are specifically calibrated to prevent these high-damage fabrications, even at the cost of producing a shorter or less confident answer.
Tips and notes
- Pair with retrieval. Grounding rules are strongest when you actually supply sources; “cite only provided sources” is empty if there are none.
- Use maximal strictness for high-stakes domains. Medical, legal, and financial answers benefit from sentence-level supported/unverified tagging.
- Give the model an out. Allowing “[not available]” or an explicit refusal is what stops the model from filling gaps with confident guesses.
- Verify the guardrails held. Spot-check outputs — if the model still invents specifics, raise strictness or tighten the source instructions.
- Grounding rules reduce hallucinations; they do not eliminate them. Use them as a first line of defense, not as a substitute for human verification of critical facts.