A new GraphRAGProv preprint links database-generated answers to the exact passages used in their computation. Its demonstration reveals how an answer can be correct despite a mistaken interpretation or an incomplete query. Researchers examined 400 questions in a supervised setting, giving the system access to expected answers during construction.
GraphRAGProv, an architecture described in a single new arXiv preprint, makes the computation behind an answer inspectable. Language models extract relationships from text and turn questions into database queries; the database produces the answer. Each contributing statement keeps a link to the exact wording from which it was extracted. Separate occurrences remain distinguishable even when different passages yield the same relationship.[1]
Researchers obtained full source traces for 311 of 400 questions. In one example, the requested country was right, although the nationality adjective concerned the movie, not the person directing it. Another query returned the right year without establishing the required connection between a game and its publisher. The system exposes which statements the database combined, allowing those mistaken interpretations and missing relationships to be located.[1]
The demonstration uses 200 questions each from 2WikiMultiHopQA and MuSiQue, two tests requiring information to be connected across passages. Graph and query construction receive expected answers and reasoning evidence. The authors therefore frame the experiment as an inspection of constructed computations, rather than evidence of success on unseen questions. Their interactive workbench also allows changes to the question or query while displaying the resulting answer and source links.[1]
