Research map / Large language models / Retrieval and knowledge
SmartChunk Retrieval: Query-Aware Chunk Compression with Planning for Efficient Document RAG
Xuechen Zhang, Koustava Goswami, Samet Oymak, Jiasi Chen, Nedim Lipka
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the query-aware planner (ablation) | SmartChunk · Main QA benchmarks | monetary cost / latency: cost 0.096 -> 0.078, latency 1.94 -> 3.62 for full model | Improved |
| use trained compression encoder instead of direct encoding of chunks (ablation) | SmartChunk · Main QA benchmarks | QA accuracy: 0.427 -> 0.564 | Improved |
| use compression encoder instead of GPT summarization of each chunk (ablation) | SmartChunk · Main QA benchmarks | monetary cost: 0.204 -> 0.078 | Improved |
| use SmartChunk instead of single-level fixed-size chunking | RAG pipeline · Main QA benchmarks | QA accuracy: 0.285 -> 0.564 | Improved |
| use few-shot prompting for planner on out-of-domain data | SmartChunk · NewsQA | F1: 0.875 -> 0.906 | Improved |
| fine-tune planner with reasoning instead of frozen LLM (ablation) | planner · Planner evaluation | planning accuracy: 0.426 -> 0.820 | Improved |
| add reasoning to fine-tuned planner (vs SLM classifier) (ablation) | planner · Planner evaluation | planning accuracy: 0.724 -> 0.820 | Improved |
| train planner with STITCH instead of SFT+RL (ablation) | planner · Planner training (418k tokens) | planning accuracy: 0.544 -> 0.820 | Improved |
| use RL only without SFT or hints (ablation) | planner · Planner training | planning accuracy: 0.820 -> 0.356 | Hurt |
| sample reasoning traces from 6 diverse LLMs instead of one (ablation) | planner SFT · Planner trace diversity | planning accuracy: 53.1% -> 74.3% | Improved |
| use a larger single trace-generating model (ablation) | planner SFT · Planner trace diversity | planning accuracy: 53.1% -> 45.6% | Hurt |
| combine SmartChunk with late chunking and hybrid search | SmartChunk · QA benchmarks | QA accuracy | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- RAPTOR — RAPTOR: Recursive Abstractive Processing for Tree-Organized Retrieval (ICLR 2024) · better on 1 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- FreshStack: Building Realistic Benchmarks for Evaluating Retrieval on Technical Documents (NeurIPS 2025 · similarity 0.87)
- Accelerating Inference of Retrieval-Augmented Generation via Sparse Context Selection (ICLR 2025 · similarity 0.87)
- UDA: A Benchmark Suite for Retrieval Augmented Generation in Real-World Document Analysis (NeurIPS 2024 · similarity 0.87)
- Cooperative Retrieval-Augmented Generation for Question Answering: Mutual Information Exchange and Ranking by Contrasting Layers (NeurIPS 2025 · similarity 0.86)
- RAGGED: Towards Informed Design of Scalable and Stable RAG Systems (ICML 2025 · similarity 0.86)
- RAGRouter: Learning to Route Queries to Multiple Retrieval-Augmented Language Models (NeurIPS 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
