Loop Engineering for Listing Questions: When the Answer Is Every Passage, Not the Top One

Standard RAG pipelines often fail silently on "listing questions" that ask for every instance of something, like all policy exclusions. These systems retrieve the top few relevant passages, but the full answer is often scattered throughout the document, leading to incomplete yet confident-looking lists. To overcome this, a system must first detect the user's intent to create a list, often signaled by words like "all," "list," or "every." Instead of simple similarity search, specialized retrieval strategies are then used, such as extracting information from the document's structure, finding repeating text patterns, or running an iterative semantic search. This process continues in a loop until a completeness signal, like a count mentioned in the text, confirms the entire list has been found.