Search Infrastructure for Video Catalogs: Faceted Indexing, Typo Tolerance, and Instant Results

Video catalogs die by their search box. When a library holds hundreds of thousands of titles, users who can’t find what they want in one query simply leave — and the gap between a 40ms search response and a 400ms one shows up directly in session depth metrics.

What a Video Search Index Actually Contains

Unlike document search, video catalog queries are shallow but broad: users type partial titles, performer names, categories, and rough descriptions, usually on mobile keyboards with autocorrect fighting them. The index therefore needs:

• Title fields — n-gram tokenized for instant prefix matching ("night" should hit while typing "nigh")
• Metadata facets — duration, resolution, upload date, category tags, language
• Popularity signals — view velocity and rating folded into ranking so relevant results aren’t buried under old noise

Engine	Latency @ 1M Docs	Typo Tolerance	Ops Complexity
Elasticsearch/OpenSearch	20–60 ms	Fuzzy + analyzer tuning	High (JVM cluster)
Typesense	10–30 ms	Built-in, per-field	Low (single binary)
Meilisearch	15–40 ms	Built-in	Low
Postgres FTS	50–200 ms	Manual (pg_trgm)	Lowest — already deployed

Typo Tolerance Is Non-Negotiable

Mobile typo rates on search queries run 8–15%. Levenshtein-distance fuzzy matching with max_fuzziness=2 recovers most of them — "streeming" still finds "streaming" — but unbounded fuzziness pollutes results. The production sweet spot: fuzziness on title tokens only, exact matching on facet filters.

{
  "searches": [{
    "collection": "videos",
    "q": "streeming",
    "query_by": "title,tags",
    "num_typos": 2,
    "sort_by": "_text_match:desc,view_velocity:desc",
    "facet_by": "category,resolution,duration_bucket"
  }]
}

Instant Search UX Pattern

The perceived speed comes from architecture, not raw latency:

1. Debounced keystroke queries at 120–150ms intervals against a lightweight edge endpoint — never let queries round-trip a distant origin.
2. Facet counts computed once per query and cached; rendering category counts alongside results costs almost nothing in inverted-index engines.
3. Prefetch top result — when the query settles, pre-warm the top result’s detail page so navigation feels instant.

“Search is the highest-intent interaction on any catalog. Every 100ms of query latency measurably reduces the probability the user issues a second query at all.”

Ranking formula details, index refresh cadence, and edge cache patterns are documented in our video catalog search architecture notes.