← AI Feed
AI Feed

The regulator wrote down its own five steps

A supervisory authority published the test behind a fine, a cost platform started importing value figures out of the databases that hold them, and the man who sells structured search said the chunking era failed for want of structure.

How we Organise

The value figure now comes out of the database that recorded it

Vantage added BigQuery and ClickHouse as sources for its business metrics on 17 September. A customer supplies SQL, or the id of a saved query, and the platform pulls the available history and refreshes it daily. Until now a firm keeping subscribers or AI invocations in either store had to build a pipeline or export a spreadsheet, and Vantage says both fell behind. The imported figures feed cost per subscriber, tokens per Jira ticket and gross margin.

A client asks where a saving came from, and the only answer that helps is a row they can open themselves. Our practice puts it flatly:

Every value row records the measurement in the source system that produced it.

The import lands as a daily copy rather than a live read, so somebody could fairly say the trust problem has moved one hop instead of going away. It has moved to a place where a query settles it.

How we Build

The chunking era failed because the material had no structure to search

Jon Bratseth, who runs Vespa, argues that an agent should be handed the query surface a professional analyst uses: field selection, proximity, date ranges, grouping, a choice of ranking profiles. Search boxes were kept crude for people, because people will not use finer control. An agent has no such reluctance. It will string several queries together to build an overview, form a hypothesis and check a detail. The 2024 belief, that chopping text into chunks and retrieving by nearest neighbour was enough, did not work.

The complaint we hear is that the search is bad. It is almost always the material:

Retrieval is a structure problem, not a search problem.

Bratseth sells the structure he recommends and publishes no evaluation of his own, and that limits how far the post carries. It is still a search vendor saying similarity alone was never the answer.

How we Assure

A supervisory authority published the test behind a fine

The European Data Protection Board adopted guidelines on 21 September setting out the five steps a data-protection authority works through before imposing a fine. Can the infringement carry one, under the regulation or national law. Who is liable, which turns on who was bound by the provision breached. Was it intentional or negligent. What are the aggravating and mitigating factors. Would a fine be effective, proportionate and dissuasive. Fourteen worked examples come with it, and consultation runs to 13 November.

Where two people can both decide, neither has to. We ask for the opposite:

Every decision right the firm has written down carries a test and exactly one decider.

The Board is designing its own process rather than a firm’s, and the text can still move before November. Step two is the part to read. Liability is settled on who was bound by the provision breached, before any figure is discussed.