Step‑Back Prompting first generates an abstract, higher‑level "step‑back" question from the original query, retrieves context for both, and then uses the broader context to answer the specific question. This improves retrieval for narrow or specialized queries.
Step‑Back Prompting addresses the problem of overly specific questions that may not match relevant documents directly. It prompts the LLM to generate a more abstract or general version of the question (the “step‑back” question). Then it retrieves documents for both the original and the step‑back questions. The broader context helps the LLM reason about the specific case, leading to more accurate and grounded answers, especially in technical or niche domains.[reference:6][reference:7]
Highly specific technical questions (e.g., physics, medicine, law) where exact wording may not match the corpus.
Queries that require reasoning about broader principles before applying to a specific case.
Fact‑verification tasks where the original question is too narrow to retrieve sufficient context.
Educational or tutorial scenarios where abstract concepts need to be established first.