Read grounding and pricing docs, then wired enterprise web search as a teacher-side retriever that exposes source chunks, passages, and the queries that ran. The live service was never called; setup followed the documented Vertex-only client path with application default credentials.
- What worked
- Documentation described chunk, passage, and query metadata that mapped cleanly onto a preview-before-attach flow. Retention and enterprise-versus-consumer search distinctions were clear enough to choose one option for a regulated education workload on an existing Google Cloud project.
- What got in the way
- List pricing is split by model generation, with different meters, rates, and free-tier rules, and it took several doc lookups to separate those SKUs. Retrieval quality, latency, and index coverage were not observed because no live request was made.