The port-operations copilot needs to answer questions scoped to a specific port and a specific date range, such as 'what circulars were active at the Rotterdam terminal in March.' Beyond the chunked notice text itself, what must the data-preparation pipeline also produce for this requirement to be satisfiable?

Structured metadata columns such as port and effective-date, written alongside the text chunks, are correct because Vector Search's metadata-filtering capability is exactly the mechanism designed to scope queries to a specific port and date range, and that filtering only works if those fields exist as structured columns in the underlying Delta table rather than being buried inside free-text prose. Relying on the embedding of the notice text alone is wrong because semantic similarity search finds text that means something similar to the query, but it does not reliably guarantee an exact, hard scope like 'only Rotterdam' or 'only March' - a semantically similar circular from a different port could easily out-rank a less similar but geographically correct one without an explicit filter. A separate embedding model trained exclusively on port names is wrong because embedding models represent semantic meaning broadly; training or choosing a model narrowly around port-name recognition is not how structured, exact-match filtering is implemented in this system, and it would not give reliable exact scoping the way a metadata column does. Increasing chunk size so the port name is more likely to appear in the text is a weak, unreliable workaround compared to a dedicated metadata field - it depends on the port name happening to be mentioned in that particular chunk's text and still requires the retriever to treat it as filterable, which raw embedded text does not support the way a structured column does. The exam-relevant point is that precise scoped filtering requires structured metadata captured during preparation, not larger chunks or embedding tricks.