Staged retrieval and chunk boundaries for a travel itinerary planner

The piece argues that retrieval quality and latency trade off against each other in itinerary planning, since large chunks add noise while small ones can separate facts from their conditions. It recommends staged retrieval with explicit collections, bounded queries, traceable source context, and chunk boundaries tied to traveler decisions. A ferry-timetable alerting example illustrates why an answer must include the sailing, dates, route, restriction, and source together. The author offers a shortlist of vector and search services (Pinecone, Weaviate, Elasticsearch, Infrai) but stresses these are not benchmarks and must be tested on one's own corpus. The proposed pipeline has four observable stages (page ingestion, chunk construction, bounded retrieval, cited answer), with metadata filters and snapshot-based freshness to keep changes auditable.