Semantic-Driven Context Pruning for Arabic RAG: Quality-Preserving Compression Under Token Inflation

Abstract

Retrieval-Augmented Generation (RAG) for Modern Standard Arabic (MSA) is costly because Arabic needs more tokens than English for the same content, which enlarges the context the generator must hold. We evaluate sentence-level cross-encoder context pruning for Arabic RAG, adapting a DSLR-style split-score-reassemble pipeline into an open middleware (LSPM). Two tokenizers show that Arabic needs 1.76–1.79 times as many tokens as English for the same content (30 sentence pairs, p < 0.0001). On 140 Arabic reading-comprehension questions with a controlled six-document candidate pool, LSPM beats sentence-count-matched truncation at the most aggressive ratio tested (token-F1 0.652 vs. 0.537, Holm-adjusted p = 0.0067) and is statistically equivalent to the unpruned answer (TOST, ±0.05 F1 margin, p = 0.037). The advantage holds under token-budget matching, against LLMLingua-2, with a second generator, and on a fresh 140-question sample. Measured token reductions imply a KV-cache saving of 57–126 MiB per request under Llama-3.1-8B-Instruct’s published geometry; this is an analytical projection. Exploratory GPU serving experiments and an optional cache-occupancy ratio controller are released with the code, but they do not show a serving benefit, and we make no serving claim.