Group repo-RAG evidence by canonical code object, not by fragment — Recall@4K jumps 39.6% → 63.7% while using 36% fewer tokens

arXiv 2607.26937·high signalXBlueskyLinkedInCopy link

VITAL-RAG (submitted July 29, 2026) names the failure mode most repo-level RAG has: fragments from the same function or class get treated as separate hits, so redundant renderings of one object eat multiple context slots and crowd out genuinely different code. The fix is to canonicalize evidence per code object, admit at most one companion fragment and only when it carries semantics the canonical view lacks, then render under both per-object and global token budgets. On RepoBench this lifted Recall@4K from 39.59% to 63.67% while cutting evidence tokens 35.63%, and it took the top raw Pass@1 on RepoExec — a rare case where the dedup and the budget both pull the same direction.