Cloud Search as a Service: Building Fast Website Search
Cloud search is a managed retrieval service whose provider operates indexing and query infrastructure while the customer owns connectors, document quality, permissions, relevance policy, and evaluation. It can reduce operational work, but it does not make search “automatic.” Choose it when those managed boundaries are worth the recurring cost and exit risk.
Map the service boundary before comparing features
List content sources, update rate, document volume, query rate, languages, regions, authorization model, latency objective, retention, and recovery expectations. Determine whether the provider pulls from supported connectors or accepts pushed documents, how deletes propagate, and whether the customer can rebuild from its own source of truth.
Products target different boundaries. Amazon Kendra emphasizes enterprise connectors and access-aware retrieval; Azure AI Search exposes indexes, lexical/vector queries, and enrichment; Google's Vertex AI Search offers managed discovery over configured data. Compare the current contracts, not a generic “AI search” label.
Keep ingestion and authorization verifiable
Use stable document IDs and record source version, connector checkpoint, index version, and delete state. Monitor freshness lag per source. Permissions must be carried into retrieval and tested with representative identities; post-filtering results after snippets or counts are returned can leak protected information.
Encrypt transport, minimize query logging, set retention deliberately, and document subprocessors and regions. Search queries often reveal customer names, incidents, medical concerns, or unreleased products. “Hosted in the EU” does not by itself establish a complete data-protection posture.
Evaluate relevance and total operating cost
Run the same judged queries against shortlisted services. Measure candidate recall, top-result relevance, permission correctness, freshness, p50/p95/p99 latency, error behavior, and supported observability. Calculate cost from indexed bytes or documents, replicas, ingestion, enrichment, vectorization, queries, egress, support, and the engineering needed for connectors and tuning.
Include failure drills: quota exhaustion, regional outage, stale connector, accidental full reindex, malformed document, and provider timeout. Define whether the application serves cached results, a lexical fallback, or a clear unavailable state.
Design the exit before onboarding
Retain original documents and ACLs outside the provider. Put a narrow, application-owned query and result model around the service, and export configuration such as synonyms and boosts. Avoid provider-specific fields in every UI client. A managed service is a sound choice when it wins a measured evaluation and the organization can leave it without reconstructing its corpus from search results.
Published · Updated