Shadowed in Vector Space: Zero-Label Detection and Resolution of Stale Version Collisions in Retrieval-Augmented Generation
Retrieval-Augmented Generation (RAG) systems deployed over dynamic enterprise corpora suffer from a silent reliability failure: when source documents undergo factual revisions (e.g., policy updates, specification changes, or deadline adjustments), residual outdated chunks (O) frequently coexist with updated chunks (R) in the retrieval index due to incremental upserts, secondary FAQ summaries, or delayed shard purging. Recent empirical benchmarks, including HoH (Ouyang et al., 2025) and StaleBench (Singh, 2026), demonstrate that retrieving R and O simultaneously degrades downstream Large Language Model (LLM) accuracy by over 20% and leaves roughly 50% of generated answers stale due to generator position bias. However, existing countermeasures either require ground-truth oracle labels for post-hoc auditing or rely on supervised classifiers over globally pooled sentence embeddings (ConflictRAG; Wang et al., 2026) that dilute localized m-token factual edits by a factor of O(m/L) across chunk length L, necessitating expensive LLM fallbacks (>800 ms latency). In this paper, we formalize the O(1/L) Pooling Dilution Theorem, proving why standard dense bi-encoders, diversity filters such as Maximal Marginal Relevance (MMR), and pooled-embedding conflict classifiers fail under fine-grained temporal updates. To overcome this bottleneck, we introduce the Stale Collision Score (SCS*) and Collision Graph Pruning (CGP), a training-free, zero-label PyTorch layer operating at the retriever boundary. By decomposing the retriever's unpooled token hidden states into a Global Shared Frame and a Query-Conditioned Orthogonal Residual Span, SCS* isolates localized numeric, temporal, and entity mutations while normalizing benign unit-converted equivalences (e.g., 14 days vs. two weeks). Furthermore, we expose and empirically validate the MMR Diversity Paradox: under tight retrieval score ties, standard MMR diversity filtering actively increases Stale-Answer Rates (SAR) up to 85.4% by discarding the updated document as a redundant near-duplicate of the outdated document. Evaluated across 2,400 multi-version corpus chunks, three dense retriever backbones (BGE-base, E5-base, MiniLM-L6), and live Qwen2.5-1.5B-Instruct generation on an NVIDIA A100 GPU, SCS*+CGP achieves 0.966-1.000 AUROC, lifts live LLM fresh answer accuracy from 43.8% to 97.5% (reducing SAR from 56.2% to 0.0%), and executes 3.6x faster than pairwise DeBERTa-v3 NLI cross-encoders and 18x faster than LLM fallbacks. An independent re-run on different hardware (Appendix A) reproduces the qualitative findings, but shows lower detection AUROC when domain-specific regular expressions are removed. Code and logs: https://github.com/soham2710/scs-rag-collision-pruner
Authors
- Botmartz AI Solutions Pvt Ltd
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-15
- DOI
- https://doi.org/10.5281/zenodo.24243870
- Primary Topic
- Topic Modeling
- Type
- preprint