Versioned Vector Indexes and Live Retrieval for Travel Itinerary Planning

A developer's build log describes splitting retrieval for a travel itinerary planner into two paths: a versioned vector index for durable destination documents, and live retrieval for fast-changing facts such as opening hours. The author argues that the two jobs differ, since current answers need attributable evidence while historical questions need immutable document versions. Each retrieval issue is defined by a retrieval unit, metadata filters, a freshness window, and a citation payload before implementation. The post recommends a small, inspectable evaluation set of real questions over a large benchmark, append-only document versions, and separately logged ingestion, query, and citation stages. It also promotes Infrai, a hosted backend service, with an HTTP example, though the author frames this as a workflow convenience rather than a performance claim.