RAG检索失败根源：文档切分导致语义断裂与验证方案
Your RAG can't answer. It's probably not the model's fault
直击RAG落地痛点，提供了一套低成本、高可复用的文档切分验证工作流，工程师可直接参考其规则设计优化现有Pipeline。
I was debugging a contract Q&A RAG recently. One question kept failing: When can the Contractor terminate this Agreement? The retriever returned plausible-looking chunks. The model tried hard. The answer was still wrong
我最近在调试一个合同问答的 RAG（检索增强生成）系统。有一个问题一直失败：承包商何时可以终止本协议？检索器返回了看似合理的文本块。模型也尽力了，但答案仍然是错误的。
Then I opened the source document and found this -- the splitter had cut as single sentence into two chunks:
然后我打开了源文档，发现了这个问题——分割器将单个句子切分成了两个文本块：
Chunk 1: if the Client breaches any material obligation,
文本块 1：如果客户违反任何重大义务，
Chunk 2: the Contractor may terminate this Agreement.
文本块 2：承包商可以终止本协议。
Both chunks passed my chunk-size check. Both got embedded.Neither one, alone, can answer the question
这两个文本块都通过了我的文本块大小检查。两者都被嵌入。但单独任何一个都无法回答这个问题。
The failure wasn't in the model or the embeddings. It happened two steps earlier - when the document was cut into pieces
问题不出在模型或嵌入上。它发生在更早的两个步骤之前——当文档被切成碎片时。
I kept running into the same class of problem across different documents, so I built a library to catcch it, rag-verifier runs quality gates at two points in the pipeline
我在不同文档中反复遇到同一类问题，因此我构建了一个库来捕获这些问题。rag-verifier 在管道的两个关键点运行质量门控。
Before ingestion: is this chunk self-sufficient?
摄入前：这个文本块是否自足？
A rule layer(pure regex, zero LLM cost) scans every chunk for structural defects: a conditional sentence cut in half, a chunk that start with 'Therefore' or 'However' with its premise left behind, a template fragment like 'yyyy' left blank
规则层（纯正则表达式，零 LLM 成本）扫描每个文本块的结构缺陷：被切断一半的条件句、以“因此”或“然而”开头但其前提条件被留下的文本块、像 'yyyy' 这样留空的模板片段。
When the rule layer can't decide -- say the grammar looks fine but the meaning might still depen on context -- an LLM semantic layer takes over. If the LLM is down, it returns uncertain. It never fakes a pass
当规则层无法决定时——例如语法看起来没问题，但含义可能仍然依赖上下文——LLM 语义层接管。如果 LLM 不可用，它返回不确定。它从不伪造通过结果。
After generation: does the citation actually support the claim?This one is sneakier. THE AI quotes something that is verbatim from the source -- but it's only the result clause of a conditional.
生成后：引用是否真正支持该主张？这一点更隐蔽。AI 引用了源文档中的逐字内容——但它只是条件句的结果部分。
The trigger condition got cut away, The quote sounds right, but citing it alone misleads the reader
触发条件被切除了。引用听起来正确，但单独引用它会误导读者。
There's a deterministic check for this. if the citation sits at the result side of a conditional in the source, an the condition isn't inside the citation, it's flagged immediately -- no LLM call needed.
对此有一个确定性检查。如果引用位于源文档中条件句的结果侧，且条件不在引用内部，则立即标记——无需调用 LLM。
Tested on 1388 chunks across 20 real documents, with a rule-layer false-positive rate around 0.5%
在 20 个真实文档的 1388 个文本块上进行了测试，规则层的误报率约为 0.5%。
GitHub: github.com/hlc0000/rag-verifier
if you have a chunk that looks fine but you RAG can't answer from it -- drop it in the comments. I'II run it through and show you exactly what it catches
如果你有一个看起来没问题但你的 RAG 无法从中回答问题的文本块——请在评论中贴出来。我会将其放入系统中运行，并展示它具体捕获了什么。
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