A design that doesn't hand regulatory review wholesale to one model

This describes a real deployment that, instead of feeding hundreds of pages of design documents and constantly amended laws into one model wholesale, has a lead role set the review plan and order and then divides searching the documents, searching the regulations, and cross-checking the two among separate roles. What stands out is that the division of labor is justified by concrete failures: omissions and mixed-up information. For search, since the provisions include both ones that are found only when the terms match exactly and ones that mean the same thing in different wording, keyword and vector search are used together, and the knowledge base is kept current through incremental updates that re-collect only what has changed. The result stressed most is not speed but evidence tracing: clicking a review result immediately shows where in the original text the basis for the judgment lies. However, since it is a case study produced by the vendor, not a single quantitative result, cost figure or failure case is presented.

Key points

  • A lead sets the plan, and the two searches and the cross-check go to separate roles
  • Hybrid search combining keyword and vector search
  • Stays current through incremental updates that re-collect only amended laws
  • A tracing setup in which clicking a result shows where the source passage is

This is a note: my own write-up after reading several sources. It has no source link; events checked against a source are kept as records.

Same week's notes

Is it still holding the first rule at the tenth file?Agents meant to be isolated exchanged signals with each otherWhat held back the switch was accumulated setup, not performanceA flood of malicious packages in a short span, and the bar for isolationThe review bottleneck weighs more than the calls to slow down