QASC boosts RAG retrieval by 18-27% with query-adaptive chunking

Mudit Rastogi's new paper, "Query-Adaptive Semantic Chunking for Retrieval-Augmented Generation: A Dynamic Strategy with Contextual Window Expansion," tackles the long-standing chunking problem in RAG pipelines. Traditional fixed chunking splits documents uniformly, ignoring semantics and user intent, leading to a precision-recall trade-off. Semantic and agentic chunking improve on this but still don't factor in the specific query at segmentation time. QASC closes that gap by embedding the user's query directly into the chunking process. It works through three novel mechanisms: cosine similarity scoring between sentence and query embeddings to pick seed sentences, contextual window expansion around those seeds to keep coherence, and chunk-level score aggregation to ensure whole-chunk relevance.

Rastogi evaluated QASC on a dataset of 100 technical documents across 200 queries spanning four distinct types, comparing against fixed chunking at five granularities, recursive splitting, semantic chunking, and agentic chunking. The results are impressive: QASC achieves an F1-score of 0.85, representing a relative improvement of 18-27% over fixed chunking and 8-12% over the best semantic and agentic alternatives. Ablation studies confirm each of the three components contributes meaningfully. Human evaluation by three annotators (with a Cohen kappa of 0.82) further validates that QASC produces chunks that are more relevant and coherent than any existing method. For professionals building production RAG systems, this means more accurate retrievals, less hallucination, and a smarter way to handle diverse queries without manual tuning.