Beyond Chunks: Answering Questions Across a Document Library by Reading Whole Sections and Checking Every Claim

Most systems that answer questions from an organization's documents cut each document into small chunks, fetch the few chunks that look most like the question, and ask a language model to write an answer from them. This works when the answer sits in one place. It works poorly when the answer is spread over several documents: a rule in one policy, changed by a memo, with an exception in a third file. We describe Chunkless, a question-answering engine that keeps each document's own structure (sections, tables, figures), records at upload time the facts each document states and how the documents relate to each other, reads whole sections instead of fragments, and checks each statement it shows against the exact text it cites. We test it on 34 questions that each need more than one document. The facts a complete answer must contain were written down before any run, and half of the questions were held out from all tuning. A blind judge scores the share of required facts each answer states. A standard chunk-based pipeline states 0.63 of them on the development questions and 0.59 on the held-out ones (0.70 and 0.65 with a reranker). Our everyday mode states 0.79 and 0.79 on a library of 32 documents and 1.5 million characters, and our slower research mode 0.87 and 0.81. A model given every section of a small library in a single call reaches 0.92 and 0.95, and the evidence our engine reads already holds 0.99 of the required facts, so most of the remaining loss happens when the answer is written, not when the text is found. We also report the changes that did not help, and three places where we had to correct our own earlier numbers.