SCAR: Semantic Continuity-Aware Retrieval for Efficient Context Expansion in RAG

2026-06-15 · Source: Computation and Language · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing, Information Retrieval · Depth: Expert, quick

Summary

SCAR (Semantic Continuity-Aware Retrieval) is an adaptive retrieval policy designed to mitigate boundary fragmentation in Retrieval-Augmented Generation (RAG) systems, a common issue with fixed-length chunking that degrades recall. SCAR selectively expands neighboring chunks by balancing query-neighbor relevance with a structural continuity penalty, employing a relative expansion threshold for scale-invariant decisions across embedding models without recalibration. Evaluated across four diverse corpora, SCAR achieved 92.8% recall on boundary-fragmented queries using only 7.84 chunks, representing a 22.9% reduction compared to static windowing's 10.16 chunks. This chunk reduction is highly significant (p<0.0001). The policy successfully transfers across text-embedding-3-large, BGE-large-en-v1.5, and zembed-1 models. Furthermore, RAGAS evaluation on the 10-K corpus confirmed SCAR preserves generation faithfulness while reducing context tokens by 27.1%.

Key takeaway

For RAG developers optimizing context window efficiency and retrieval accuracy, SCAR presents a compelling adaptive retrieval strategy. By dynamically expanding context based on semantic continuity and relevance, you can achieve 92.8% recall on fragmented queries while significantly reducing context tokens by 27.1% and overall chunk count by 22.9%. This approach offers a robust way to enhance RAG system performance and reduce operational costs, especially given its transferability across various embedding models.

Key insights

SCAR adaptively expands RAG context by balancing semantic relevance and structural continuity to improve recall and efficiency.

Principles

• Adaptive context expansion reduces token overhead while maintaining recall.
• Balancing query-neighbor relevance with structural continuity improves retrieval.
• Scale-invariant decision rules enhance transferability across embedding models.

Method

SCAR selectively expands neighboring chunks by weighing query-neighbor relevance against a structural continuity penalty, using a relative expansion threshold tied to the retrieved chunk's query-relevance.

In practice

• Implement adaptive chunk expansion in RAG pipelines.
• Prioritize semantic continuity in context retrieval.
• Reduce RAG inference costs by optimizing context length.

Topics

• Retrieval-Augmented Generation
• Context Expansion
• Semantic Continuity
• Information Retrieval
• Embedding Models
• Chunking Strategies

Best for: AI Architect, AI Engineer, Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• Embeddings Aren’t Magic: The Predictable Failure Modes of RAG Retrieval — Embeddings provide synonym-tolerant search but predictably fail on structural issues like OOV terms and logical operations.
• Rethinking the Necessity of Adaptive Retrieval-Augmented Generation through the Lens of Adaptive Listwise Ranking — Adaptive retrieval's role shifts from noise filtering for weaker LLMs to efficiency optimization for stronger ones.
• Semantic Chunking vs Fixed Chunking: Why Your RAG’s Retrieval Quality Starts Before the Query — Effective RAG retrieval hinges on semantic chunking, not just fixed-size splits, to ensure coherent context.
• Long Context Is Not Replacing RAG - It Is Forcing Us to Rethink Context Engineering — Long context models enhance, rather than replace, RAG, demanding advanced context engineering.
• Retrieval-Augmented Generation as a Bridge Between Classical NLP Pipelines and Modern LLM Systems — Retrieval-augmented generation (RAG) addresses the limitation of static knowledge in large language models (LLMs) by connecting them to external, up-to-date knowledge sources.
• Beyond Semantic Search: The Complete Guide to Advanced RAG with Milvus | Sagar Patil — Advanced RAG requires multi-stage retrieval, combining hybrid search with techniques like query rewriting and reranking for production reliability.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computation and Language.