# contract-rag

> Reproducible before/after benchmark: cleaning dirty contract PDFs for RAG.

## Pages

- [Why your contract RAG returns garbage — and a reproducible fix](https://contractrag.com/benchmark.html): A one-command, credential-free before/after benchmark: cleaning dirty contract PDFs lifts data-quality and extraction accuracy on a committed synthetic set.
- [为什么你的合同 RAG 检索全是乱码——一个可复现的解决方案](https://contractrag.com/benchmark.zh.html): 一条命令、无需凭据的前后对比基准测试：清洗脏合同 PDF 可在已提交的合成数据集上显著提升数据质量与抽取准确率。
- [Kleister-NDA, measured: honest extraction numbers on a public benchmark](https://contractrag.com/kleister-nda.html): Point-in-time results on the public Kleister-NDA benchmark (real SEC EDGAR NDAs): rule-based field-F1 0.523 → 0.697 with source-accuracy 1.0, and why server-side schema-constrained decoding eliminated 30% structured-output failures.
- [Kleister-NDA 实测：公开基准上的诚实抽取数字](https://contractrag.com/kleister-nda.zh.html): 在公开的 Kleister-NDA 基准（真实 SEC EDGAR 保密协议）上的定点实测结果：规则抽取字段 F1 0.523 → 0.697，来源归因准确率 1.0；以及服务端 Schema 约束解码为何消除了 30% 的结构化输出失败。
- [OCR confidence can't detect omissions: what a near-perfect quality score hides](https://contractrag.com/ocr-omission.html): Measured on 85 real degraded SEC filings: 7.7% of 491 expert-labeled facts appear nowhere in the OCR output, while the document-level quality score reads 0.998 — and no OCR confidence threshold can flag the loss, because an omitted fact produces no block at all.
- [OCR 置信度检测不出内容遗漏：近乎满分的质量分数掩盖了什么](https://contractrag.com/ocr-omission.zh.html): 在 85 页真实退化的 SEC 文件上实测：491 个专家标注事实中有 7.7% 在 OCR 输出里完全消失，而文档级质量分数却高达 0.998——且任何 OCR 置信度阈值都无法标记这种丢失，因为被遗漏的事实根本不产生任何文本块。