{"paper":{"title":"Correctness-Aware Repository Filtering Under Maximum Effective Context Window Constraints","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"A size-based filter using only file metadata cuts tokens in LLM repository contexts by 80 to 89 percent and raises task accuracy from 25 to 72 percent.","cross_cats":["cs.AI"],"primary_cat":"cs.SE","authors_text":"Shweta Mishra","submitted_at":"2026-05-14T04:37:39Z","abstract_excerpt":"Context window efficiency is a practical constraint in large language model (LLM)-based developer tools. Paulsen [12] shows that all tested models degrade in accuracy well before their advertised context limits the Maximum Effective Context Window (MECW) which makes context construction a quality problem, not just a cost one. Modern software repositories routinely contain large non-code artifacts compiled datasets, binary model weights, minified JavaScript bundles, and gigabyte-scale log files that overflow the context window and push out task-relevant source code. We present a correctness-awa"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Across 10 real open-source repositories (22,046 files, 5 languages), the proposed SizeFilter at θ=1 MB achieves 79.6% (±13.2%) mean token reduction at 0.30 ms overhead; the HybridFilter achieves 89.3% (±9.0%) the lowest variance of any filter evaluated. A limited-scope evaluation (18 tasks, CodeLlama-7B-Instruct) yields 72% file-level accuracy under filtering versus 25% at baseline; hallucination frequency declines from 61% to 17%.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That file size (or a quick token-density estimate) is a reliable proxy for task relevance and that discarding large files will not remove critical code needed for the downstream LLM task.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"A pre-execution size filter cuts repository tokens by 80-89% at sub-millisecond cost and raises file-level accuracy from 25% to 72% in a small CodeLlama evaluation.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"A size-based filter using only file metadata cuts tokens in LLM repository contexts by 80 to 89 percent and raises task accuracy from 25 to 72 percent.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"119aaa3deabf2a57b79966bdcec4d0d12bae2cdf19241d9b23c1de425c6b719d"},"source":{"id":"2605.14362","kind":"arxiv","version":1},"verdict":{"id":"b14afb6f-df27-4f91-bb44-8eafdc3cd999","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T02:35:17.174378Z","strongest_claim":"Across 10 real open-source repositories (22,046 files, 5 languages), the proposed SizeFilter at θ=1 MB achieves 79.6% (±13.2%) mean token reduction at 0.30 ms overhead; the HybridFilter achieves 89.3% (±9.0%) the lowest variance of any filter evaluated. A limited-scope evaluation (18 tasks, CodeLlama-7B-Instruct) yields 72% file-level accuracy under filtering versus 25% at baseline; hallucination frequency declines from 61% to 17%.","one_line_summary":"A pre-execution size filter cuts repository tokens by 80-89% at sub-millisecond cost and raises file-level accuracy from 25% to 72% in a small CodeLlama evaluation.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That file size (or a quick token-density estimate) is a reliable proxy for task relevance and that discarding large files will not remove critical code needed for the downstream LLM task.","pith_extraction_headline":"A size-based filter using only file metadata cuts tokens in LLM repository contexts by 80 to 89 percent and raises task accuracy from 25 to 72 percent."},"references":{"count":24,"sample":[{"doi":"","year":2020,"title":"Retrieval-Augmented Generation for Knowledge- Intensive NLP Tasks,","work_id":"77d5a28c-994d-439b-bd97-353caa70cee5","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":1901,"title":"Language Models are Few-Shot Learners,","work_id":"4d251c53-e06a-439e-9ad8-594a06476d99","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2021,"title":"Evaluating Large Language Models Trained on Code","work_id":"042493e9-b26f-4b4e-bbde-382072ca9b08","ref_index":3,"cited_arxiv_id":"2107.03374","is_internal_anchor":true},{"doi":"","year":2013,"title":"The Tail at Scale,","work_id":"ca1c484b-8973-4fe0-a7e7-472cf3adff50","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"LLMLingua: Compressing Prompts for Accelerated Inference,","work_id":"f9ffd372-daf4-48ed-99ae-be4ee2bd1a44","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":24,"snapshot_sha256":"0f506273cb2088a4d6ce53484091ed3ded8e84a29e1a2ca37ed399e422e1ddfd","internal_anchors":4},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}