Meridian's port-ops copilot must answer 'which berths are free right now for a 300-meter vessel,' which requires checking a live scheduling system that changes by the minute, not text sitting in a static document index. What does this requirement imply about the pipeline's design?

The correct design response is to give the agent a tool that queries the live scheduling system directly, because the requirement is fundamentally about current, minute-by-minute state, which no retrieval index built from static documents can represent no matter how well it is tuned. A static RAG index over historical circulars is insufficient on its own because circulars describe policy and past allocations, not the live occupancy of a specific berth right now; retrieval would return stale or irrelevant information for this exact question. Refusing the question as out of scope is the wrong call - it is a legitimate business question with a genuine answer, it simply cannot be answered by document retrieval alone, so the fix is architectural (add a tool), not a refusal policy. Swapping in a larger embedding model addresses semantic-match quality over the existing static corpus and has no bearing on the fact that the underlying data source (a live schedule) is not in that corpus at all - better embeddings cannot retrieve information that was never indexed. Recognizing when a question needs a live-system tool call versus static-document retrieval is central to selecting chain and tool components for a desired input/output.