实测10款文档解析器：合同问答准确率差异显著
I tried 10 document parsers on 1,500 contract questions. Sharing what I learned
RAG落地必读的工程复盘，用真实数据揭示了文档预处理中容易被忽视的细节，直接指导解析策略选择。
We ran 10 document-parsing configurations across 94 redlined contracts, then gave the output to the same AI agent to answer 1,500 questions.
我们在94份带有修订痕迹的合同上运行了10种文档解析配置，然后将输出结果交给同一个AI代理来回答1500个问题。
A few things stood out:
有几件事引人注目：
Clean Markdown can still be completely wrong
干净的Markdown格式仍然可能完全错误
Say a contract has this:
假设合同中有以下内容：
ten three five business days
ten three five business days（十个三五个工作日）
Some parsers return `ten three five business days`. All the words are there, but the agent can no longer tell which term was actually agreed.
某些解析器返回`ten three five business days`。所有单词都在，但代理已无法分辨实际商定的条款是什么。
When we removed all edit markings, the agent got 28.8% of the questions right. When every deletion was clearly preserved, it got 91.9%. These are reference runs, not vendor scores.
当我们移除所有编辑标记时，代理答对了28.8%的问题。当明确保留所有删除内容时，答对率为91.9%。这些是参考运行结果，而非供应商评分。
The right setting can matter more than changing vendors
正确的设置可能比更换供应商更重要
Datalab’s normal conversion endpoint scored 62.4%. Its tracked-changes endpoint scored 78.6%.
Datalab的标准转换端点得分为62.4%，其跟踪更改端点得分为78.6%。
Same vendor, 16.2 points higher just from using the correct endpoint.
同一供应商，仅因使用正确的端点就提高了16.2分。
Scanned PDFs weren’t always harder
扫描版PDF并不总是更难处理
This was the most surprising bit for us. We turned the same contracts into image-only PDFs and reran everything.
这是让我们最惊讶的部分。我们将相同的合同转换为纯图像PDF并重新运行了一切。
GPT-6 Astra went from 63.3% to 91.3%, while LlamaParse agentic dropped from 80.0% to 67.5%.
GPT-6 Astra从63.3%提升至91.3%，而LlamaParse agentic则从80.0%降至67.5%。
Not sure what to make of it but clearly text PDFs and scanned PDFs can take very different processing paths. Testing only clean PDFs probably won’t tell you how your production documents will behave.
不确定该如何解读，但显然文本PDF和扫描PDF可以采取截然不同的处理路径。仅测试干净PDF可能无法告诉你生产环境中文档的实际表现。
We've published more insights here: https://openbenchmarks.com/document-processing, along with full results, methodology, costs, latency, dataset and open-source runner.
我们在此发布了更多见解：https://openbenchmarks.com/document-processing，以及完整结果、方法论、成本、延迟、数据集和开源运行程序。
更进一步：量化金融体系
看懂新闻只是起点——沿量化金融路径，把它变成能交付的工程能力