Retail

Semantic Search for a Small Catalog: When It Is Overkill

September 2026 · ISTRALLEN Team

The honest answer for a few hundred SKUs

If your catalogue is a few hundred products, semantic search is probably not the thing to build first. A well-configured keyword search with a maintained synonym list, clean category structure, and a good set of filters will handle most queries, cost almost nothing to run, and be live in days. The case for an embedding pipeline gets stronger as the catalogue and the query variety grow, not before.

For a few hundred SKUs, semantic search is usually overkill. A maintained synonym list, consistent filterable attributes, typo and plural handling, and a useful zero-results page fix most “search is broken” complaints in days, at near-zero running cost. Add an embedding pipeline only when the signals change — the catalogue grows past a few thousand SKUs, a supplier feed outpaces tagging, queries turn mostly descriptive rather than named, or the synonym list becomes unmaintainable — and even then keep it minimal: embeddings in the catalogue database, one re-embedding worker, no dedicated vector store or second-pass reranker. Re-check the signals every couple of quarters; deciding once and never revisiting is how a store ends up with search that no longer fits it. Two exceptions pull the decision earlier: a specialist shop whose shoppers describe needs rather than name products, and a multilingual small catalogue, where synonym lists hit their limits fast.

An embedding-based search adds real moving parts: an embedding model call per product and per query, a vector index to maintain, a re-embedding job when products change, and often a reranking round-trip. Our semantic search project took on all of that deliberately, at 120k SKUs fed by multiple suppliers with a four-person team drowning in synonym spreadsheets. A small catalogue rarely has that problem.

Where keyword search already wins

For a small, well-understood range, most searches are for things the shopper can already name — a product type, a brand, a colour. Keyword matching plus synonyms handles those directly, and filters do the rest. The long-tail natural-language queries that semantic search shines on are a smaller share of a smaller catalogue.

Fix the cheap things first

Before reaching for embeddings, the higher-return work is usually: a synonym list for the terms your shoppers actually use, consistent attributes so filters work, sensible handling of typos and plurals, and a good zero-results page. These fix most "search is bad" complaints on a small store.

The signals that change the answer

Semantic search starts to earn its place when the catalogue grows past a few thousand SKUs, when suppliers add products faster than anyone can tag them, when a large share of queries are descriptive rather than nominal, or when the merchandising team cannot keep the synonym list current. Those are the conditions to watch for.

If you do build it, keep it simple

A small catalogue does not need a dedicated vector database or a heavy reranking stack. Embeddings in the same database as the catalogue, a single re-embedding worker, and vector search without a second-pass reranker is often enough, and it can be added to later.

A worked example

A store with 400 products keeps getting "search is broken" complaints. An audit finds the real causes: no synonym for a term half the customers use, three colour attributes spelled inconsistently so the colour filter misses items, and a zero-results page that is a dead end. Fixing those three things — a day of work — resolves most of the complaints. An embedding pipeline would have taken weeks and addressed a smaller share of the actual queries.

Revisit the decision on a schedule

The right answer changes as the catalogue grows. Put a reminder to re-check the signals every couple of quarters: catalogue size, the share of descriptive versus nominal queries, how much effort the synonym list takes to maintain, and whether new markets or languages have been added. Deciding once and never revisiting is how a store ends up with search that no longer fits it.

Where this stops being right

A small catalogue with unusually varied language — a specialist shop whose shoppers describe needs rather than name products — may benefit from semantic search sooner than size alone suggests.
Multilingual small catalogues hit the limits of synonym lists quickly, which can justify embeddings earlier.
If your team keeps the synonym list current with little effort, that is a sign you do not need to replace it yet.

FAQ

When is semantic search overkill for a store? When the catalogue is small, mostly searched by name, and well served by synonyms and filters. The operational cost of an embedding pipeline outweighs the benefit at that scale.

What should a small store fix first instead? Synonyms for real shopper terms, consistent filterable attributes, typo and plural handling, and a useful zero-results page. These resolve most search complaints cheaply.

What signals mean it is time to reconsider? Catalogue growth past a few thousand SKUs, a supplier feed outpacing tagging, mostly descriptive queries, or a synonym list the team can no longer maintain.

ISTRALLEN builds semantic search where it pays off, and will tell you when clean filters and synonyms are the better spend — see AI for Retail.

See it in production

AI for Retail → Semantic search case study →

← All articles