Multi-hop question prompt builder
Some questions can only be answered by connecting facts that live in different places: who founded the company, where that person studied, what that university is known for. A single retrieved chunk never holds the whole answer. This builder produces a prompt that makes an LLM reason hop by hop across multiple documents, naming the source at every step, so the final answer is not just correct but traceable back to the evidence.
How it works
You paste the source documents (each with a short label), the question, a cap on the number of reasoning hops, and your preferred citation format. The tool builds a prompt that instructs the model to first identify which documents are relevant, then chain its reasoning one hop at a time — each step stating a fact, the document it came from, and what it lets the model conclude next — before assembling the final answer. Critically, the prompt tells the model to ground every fact strictly in the provided documents and to declare an unanswerable question rather than fill gaps with outside knowledge. The cited chain is what lets you audit and trust the result.
Tips and notes
Label your documents clearly — “Doc A: 2023 annual report” beats an unlabeled blob, because the citations are only useful if you can map them back. Set max hops to the smallest number that can plausibly answer the question; an over-generous cap invites the model to manufacture intermediate steps. Always read the chain, not just the conclusion: the value of this format is that a wrong or hallucinated fact shows up as a specific, checkable link rather than hiding inside a confident final sentence. If the model reports the documents do not contain the answer, that is a correct and useful result — it means you need better sources, not a different prompt.