Ilia Karpov, Ivan Chepikov, Denis Losev, Alina Avanesyan, Maria Godunova, Dmitry Ilvovsky · ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 2026
DOI: 10.1145/3799682.3840242
Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).
GraphRAG pipelines build entity–relation graphs from text and use them for retrieval and question answering, but extraction noise, entity fragmentation, and unstable clustering can silently corrupt the evidence behind an answer. Correcting these errors is difficult to scale because their dominant forms depend on the corpus and repeated corrections remain manual. We present GraphCraft, an interactive system that records provenance-rich human corrections and compiles repeated corrections into reusable curation skills. Each skill identifies similar errors in subsequent corpus updates and proposes edits for curator review; proposals are previewed rather than applied automatically. The same journal links corrections and skill suggestions back to their graph context and source spans, while bi-temporal graph state distinguishes changes to the stored graph from changes in the underlying world. We evaluate a compiled skill on held-out ingestion batches and demonstrate the complete workflow from diagnosing an erroneous answer, through repairing its graph evidence, to reusing the correction on later data. GraphCraft is released as an open platform with offline deployment. Code and offline deployment: https://github.com/apsolutions/graphrag; demo video: https://vimeo.com/1141894235.
ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 0 citations
No comments yet — start the discussion below.