CHOP: Chunkwise Context-Preserving Framework for RAG on Multi Documents

The core problem CHOP addresses is more specific than general hallucination: when a vector database holds multiple documents that are semantically close, standard retrieval conflates them, and the retrieved chunks lose the document-of-origin signal that would let the LLM reason correctly. Metadata prefixing is the mechanism doing the heavy lifting here, not just the chunk evaluation step.

This connects most directly to IG-Search, covered the day prior, which approaches retrieval quality from the reward-signal side, measuring how much a retrieved document actually improves model confidence. CHOP and IG-Search are attacking adjacent failure modes: IG-Search targets query formulation and search strategy, while CHOP targets what happens after retrieval, when the returned chunks are ambiguous or context-stripped. Together they sketch a fuller picture of where RAG pipelines currently break down. The broader archive here is largely focused on inference efficiency and agent behavior, so CHOP sits somewhat apart from those threads.

The real test is whether CHOP's metadata prefixing approach holds up on retrieval benchmarks that include adversarially similar documents, such as BEIR subsets with near-duplicate corpora. If independent replication on those benchmarks shows consistent gains, the CNM-Extractor is doing real work; if not, the improvement may be specific to the paper's own document sets.