Start with the query that fails
A product search that cannot find a known item does not automatically need a new search vendor. First establish whether the failure comes from missing attributes, weak matching, incorrect permissions, or an index that has not caught up. Moving incomplete records into another system preserves the incompleteness and adds a synchronization problem.
Our decision rule is straightforward: keep search close to the database while its relevance and operational requirements remain manageable there. Introduce a dedicated service when the team needs search-specific capabilities it can evaluate and maintain. Catalogue size alone cannot make that decision. A specialist catalogue with complicated terminology can need more relevance work than a larger collection of straightforward records.
What the database already provides
Database search is not synonymous with scanning every title for a substring. PostgreSQL documents indexed full-text search, language configurations, and relevance ranking. That is a credible starting point for an internal application, a documentation library, or a catalogue with relatively predictable queries.
The attraction is operational proximity. The team already owns the database, its deployment process, and the application permissions. It can keep the searchable representation near the records it describes. That does not make access control automatic: filtering must still respect the viewer's permissions before results or snippets are returned.
Choose this path when an engineer can explain the ranking, investigate slow queries, and maintain the search configuration without turning every content correction into a development project. Reject it when required relevance changes repeatedly exceed the team's available attention.
What a managed service adds to the system
A dedicated search service can be evaluated for relevance controls, filtering, language handling, and operational tooling. Check the actual product and plan rather than assuming every hosted engine includes the same capabilities.
It also becomes another data surface. Decide which fields may leave the primary system, how removals propagate, and what happens during a failed indexing job. Search results should not continue exposing a product, document, or customer record after its access has been withdrawn.
For example, Algolia distinguishes administrative keys from restricted search keys. A browser-facing search key is not permission to publish an entire private dataset. The integration needs deliberately scoped access and a way to reconcile the index with the source of truth.
Compare work, not just invoices
Ask both implementations to answer the same collection of real queries. Include misspellings, exact identifiers, category terms, and requests that should return nothing. Have the people who understand the catalogue judge the results without knowing which engine produced them.
Then test a changed price, a removed item, and a permission change. A beautiful result list is not enough if it is stale. Record who can adjust relevance, how changes are reviewed, and how the previous configuration is restored.
Clodron's web platform work can involve either approach. The useful connection is between search behaviour and the underlying quality of product data, not a preferred vendor. A managed service cannot repair missing business meaning on its own.
Make the decision this week
Collect failed searches from support and analytics, then annotate the expected result and the reason it belongs there. Remove personal information before sharing the examples.
Run a small comparison against the existing database approach and a candidate service. Keep the interface identical. Evaluate result quality, freshness, permissions, operator effort, and failure recovery separately.
Stay with the database if it meets the agreed requirements with understandable maintenance. Choose the service if it resolves demonstrated gaps and the team accepts the additional indexing and access responsibilities. If neither works, inspect the data and query assumptions before buying another engine.
