Search Infrastructure: How Ecommerce Search Engines Are Built and Scaled
Search infrastructure is the technical architecture underneath a storefront's search bar — indexing, query processing, and ranking — and how well it's built determines both how relevant results are and how the system holds up as a catalog and traffic grow.
The basic pipeline
Catalog data gets indexed — transformed into a structure the search engine can query quickly. A shopper's query gets processed: parsed, and possibly rewritten or embedded. The index gets queried for candidate matches, and those candidates get ranked before being returned.
Indexing for keyword vs. semantic search
Keyword indexes are typically inverted indexes, mapping terms to the documents containing them. Semantic search requires a vector index instead, storing embeddings and supporting similarity search. Many hybrid systems maintain both simultaneously and combine results from each.
Scaling for high SKU counts
As catalog size grows, both index size and query latency become real constraints. Infrastructure choices that work fine for a few thousand SKUs often need real re-architecture at hundreds of thousands.
Build vs. layer
Infrastructure changes don't have to mean replacing an existing search stack outright — a layer can sit on top of an existing engine and handle specific failure cases, deferring a full rebuild until it's actually justified by scale or requirements.
Common questions
Do I need to replace my existing search engine to improve search infrastructure?
No — infrastructure improvements can be layered on top of an existing system to address specific problems, rather than requiring a full replacement.
What changes as a catalog scales?
Index size and query latency become real constraints, and infrastructure decisions that work at a few thousand SKUs often need re-architecture at much larger scale.
What's the difference between a keyword index and a vector index?
A keyword index maps terms to documents for exact or fuzzy text matching. A vector index stores embeddings and supports similarity search for semantic matching — hybrid systems typically maintain both.