部落格
來自不光鮮那一半的記錄
關於解析難搞的文件、把權限放在該生效的地方、以及讓答案可重現的工程筆記。
Reproducing last quarter's number after the definition changed
The report said 41.2 percent in April. The same question returns 39.8 today, and nothing is wrong with either figure. Whether you can explain the gap in ten minutes or ten days depends on one architectural choice.
Row-level rules belong in the query, not in the prompt
If the only thing stopping an AI from returning another region's salary data is an instruction telling it not to, you do not have access control. You have a suggestion.
Why merged cells break retrieval, and what to do about it
A merged cell that flattens into the wrong row does not produce an error. It produces a confident, wrong number that looks exactly like a right one, and it is the most common silent failure in document AI.
Analysis rules are knowledge objects, not prompt text
Suppress cells under n=30. Control for age before reporting anything by marital status. Every senior analyst carries a hundred rules like these, and almost every AI analytics tool asks you to paste them into a prompt and hope.
The two halves of enterprise data AI, and why almost nobody joins them
Document AI reads your contracts but cannot query your warehouse. Analytics AI queries your warehouse but has never read the contract. The questions that actually matter to a business sit across the join.