The tempting shortcut
A familiar assumption deserves scrutiny: if every citation is accurate, the answer must be adequate. For "What is the deadline in this agreement?", finding and inspecting the relevant clause addresses a bounded fact. But "What risks recur across our project archive?" makes a different promise. Accurate examples can support individual claims while leaving most projects unexamined. Citation correctness therefore cannot, by itself, establish that the overview represents the collection.
Make the coverage claim visible
The GraphRAG paper addresses questions spanning an entire corpus, where a few similar passages may provide inadequate coverage. Broader RAG research distinguishes retrieval, augmentation and generation choices. These ideas support matching the method to the question, without establishing that every overview needs a graph. Define the selected projects, then disclose which documents were available. Unindexed material and access restrictions limit the answer's reach; community summaries cannot repair evidence that never entered the collection.
A hypothetical boundary case
Suppose an archive contains only three short project reports, all accessible and small enough to inspect together. The recurring-risk question still concerns the whole collection, but direct examination could provide the necessary coverage without a graph. This hypothetical case adds a routing rule: first ask whether the entire relevant evidence can be inspected within the practical budget. If yes, compare all reports directly. If no, consider broader aggregation and check its coverage. Global scope alone does not determine the machinery.
Test the promise separately
For a deadline answer, verify the clause against its source. For an overview, compare the claimed collection with the material examined, then check balance and omitted themes. Keep these evaluations separate: one success score can conceal an accurate but incomplete synthesis. ABOLTUS is in development; these examples are editorial design analysis, not reported product results or available features. The research basis here is limited to reviewed excerpts, including abstracts or introductions. Introduce global summaries when real questions and coverage checks justify them.