在文档处理流程中，数据提取的准确性往往需要多层验证。1Unsiloed用于从文档中提取结构化字段并定位其在页面上的位置，而TypeSafe Jev则通过类型化问题检查来验证提取数据的合理性。1这两个工具的结合能够有效识别OCR失败和数据错误。
Unsiloed为每条提取的记录提供两个置信度分数：extraction_score和grounding_score，两者均在0到1之间。1以药物清单为例，硫酸西替利嗪的剂量被提取为"1 tablet"，其extraction_score为0.74，grounding_score为0.76。1TypeSafe Jev的请求分为状态和决策两部分，其中状态包含提取的记录和上下文，决策部分提出相关问题。1该服务支持三种答案类型：Noul（概率）、Choice（选项）和Score（评分）。1在布洛芬的案例中，Unsiloed读取的剂量为"10 tablets, three times daily"，两个置信度分数均达到0.99，但Jev的合理性检查返回0.04，表明数据存在问题。1
从成本角度看，TypeSafe Jev的价格为每十亿输入令牌$42（截至2026年9月30日），处理100,000份文档约增加$7.61的成本。1TypeSafe Jev的API端点为https://api.typesafe.ai/v1/systemone。[1](#source-1)
A technical approach combining Unsiloed and TypeSafe Jev enables comprehensive document data extraction and validation.1 Unsiloed specializes in extracting structured fields and locating their positions within documents, while Jev performs type-based reasonableness checks on the extracted data.1 This combined methodology addresses common challenges such as OCR failures and data errors, as demonstrated through pharmaceutical document examples.1
The Jev request framework consists of two components: state information containing extracted records and context, and decision elements posing validation questions.1 Jev supports three answer types—Noul for probability assessments, Choice for multiple-choice responses, and Score for numerical ratings.1 Unsiloed provides two confidence metrics for each extraction: extraction_score and grounding_score, both ranging from 0 to 1.1 For instance, when extracting a cetirizine sulfate dosage of "1 tablet," Unsiloed reported an extraction_score of 0.74 and grounding_score of 0.76.1 In another case, despite Unsiloed reading an ibuprofen dosage as "10 tablets, three times daily" with both confidence scores at 0.99, Jev's reasonableness check returned a score of 0.04, indicating potential validity concerns.1
The TypeSafe Jev service is priced at $42 per billion input tokens as of September 30, 2026.1 Processing 100,000 documents under this pricing model adds approximately $7.61 in costs.1 The Jev API endpoint is available at https://api.typesafe.ai/v1/systemone.[1](#source-1)
评论
还没有评论，欢迎留下第一条。