role-fullstack:search-implementation

Search Implementation

When to use

• Adding full-text search to a product catalog or content site
• Building autocomplete or typeahead input
• Implementing faceted filtering with sidebar checkboxes and result counts
• Syncing a database with a search index (initial bulk load + incremental updates)
• Tuning relevance ranking after users complain results are wrong
• Choosing between self-hosted (Meilisearch, Typesense) and managed (Algolia)

Core principles

1. Engine matches scale and budget — pg_trgm for small datasets with no extra infra; Meilisearch/Typesense for fast typo-tolerant search on a budget; Algolia when managed SaaS and rich UI components justify the cost
2. Denormalize for search — flatten related data into the search document at index time; joins at query time kill latency
3. Incremental sync, not full re-index — hook into Prisma middleware, DB triggers, or a CDC pipeline; bulk re-indexing on every change is the first thing that breaks under load
4. Debounce is not optional — 200-300ms debounce on autocomplete input; without it you hit rate limits and burn the user's battery
5. Encode filters in the URL — active facets belong in useSearchParams; shareable and bookmarkable search state is a product feature, not a nice-to-have

Reference Files

• references/engine-selection-indexing.md — engine comparison table with latency and hosting characteristics, four-step indexing strategy, incremental sync options (CDC, hooks, job queue), Meilisearch settings and search code example
• references/ui-patterns-relevance.md — debounced useSearch hook with TanStack Query, keepPreviousData pattern, faceted search UI with URL sync, relevance tuning controls (field boosts, typo tolerance, synonyms, stop words, custom ranking), common pitfalls