Retrieval for Coding Agents Should Build a Sufficient Set, Not Rank Passages

arXiv 2609.20050·medium signalXBlueskyLinkedInCopy link

arXiv 2609.20050 (17 Sep 2026) argues relevance is scored per passage while sufficiency belongs to the set, so a ranker can fill its budget with variants of one required fact and leave the agent's next decision unsupported. SERBench measures this on 500 held-out agent states from 45 repositories, crediting only evidence sets that cover every fact the current decision needed. MSS-Complement makes three semantic calls to propose a jointly sufficient set, find what it lacks, and return 4-8 intact source units within 6,144 tokens, reaching a complete set for 73.0% of states at five items and 80.6% at eight, against 61.4% and 72.4% for Qwen3 embedding with reranking.