Decluttering Context with Query-Aware Snippets

Agentic search systems often retrieve more context than downstream models can usefully process, forcing search and answer generation modules to spend compute on text that is irrelevant, redundant, or distracting. We study query-aware snippets: extractive, query-conditioned compressions of retrieved documents that remove this clutter before downstream processing. Across five grounded QA benchmarks, snippets reduce downstream document tokens by nearly 80\% without quality regression on long, dense corpora, but hurt on short open-domain and code corpora where evidence is distributed or structurally dependent on context. We show that these failures are attributable to the compression step itself, including evidence fragmentation and gold-document dropout. Our main contribution is methodological: reference-free diagnostics such as self-containment, compression rate, drop rate, and readability make snippet quality measurable directly, rather than inferred only from noisy end-to-end answer scores.