NaceAI 用小模型自校验 OCR，Parse Index 得分 0.834 超 GPT-6 Astra
NaceAI 搞了个会自查的文档读取器，用小模型 Drex 对比图像和文本，只改可疑部分。Parse Index 得分 0.834 反超 GPT-6 Astra，价格却是它的十分之一，做批量 OCR 的团队值得看看这套架构。
NaceAI 发布了一个文档读取系统，核心是用小模型 Drex 对 OCR 结果自我校验：先把 PDF 或扫描件转成文本，再由 Drex 逐块比对文本与原始页面图像，只把不匹配的部分送回修正。整个服务定价为每 1000 页 7 美元。
技术上，Drex 只回答窄问题，比如某行 OCR 文本是否与图像一致，单次回答约 50 毫秒、每 100 万次回答约 2 美元。由于只重写被怀疑的文本，每页修正时间中位数下降 28% 到 38%，修正失败时回退到原始读取结果。传统 OCR 会压平表格、丢失勾选框，而 GPT-6 Astra 等前沿模型在大批量场景下成本过高。
基准上，该系统在 Parse Index 上得分 0.834，高于 GPT-6 Astra (low) 的 0.813，成本为每 1000 页 7 美元，对比后者的 70.57 美元，约为十分之一。对需要大批量处理文档的企业来说，这套“小模型做裁判”的架构在准确率和成本之间给出了新的取舍方案。
really cool launch.
Brilliant use of Decision models.
@NaceAI built a document reader that proofreads itself. After it turns a PDF or scan into text, a small model called Drex compares each block with the original page image and sends only the mismatches back for correction, all for $7 per 1,000 pages.
Standard OCR flattens tables and drops ticked boxes, while frontier models help but get expensive at volume.
Drex instead answers narrow questions, such as whether an OCR line matches the image, in about 50 milliseconds for about $2 per million answers.
Because only doubted text gets rewritten, median correction time per page fell 28% to 38%, and failed fixes revert to the original reading.
The system scores 0.834 on the Parse Index, beating GPT-6 Astra (low) at 0.813 while costing $7 per 1,000 pages against $70.57.
https://t.co/wnYo5UWUwz
This report is categorized under AITOP's "Products" desk, focusing on US frontier AI labs and Silicon Valley technology leaders. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows and US-China foundation model benchmarks and commercialization comparison. Key entities and technical topics under continuous tracking include NaceAI, OCR, Drex, GPT-6 Astra.
Source attribution: curated from rohanpaul_ai (@rohanpaul_ai).
