Search systems

Search that only matches text literally misses inflections, compound words and typos, and has no idea which hit is the right one. The database you already have can often do more: PostgreSQL, for example, has built-in full-text search with Swedish support and ranking. So start by finding out what it actually does today, and test with real searches before deciding whether the index stays there or moves to a separate search service.

Three reasons search misses

1. Exact matching only. Someone searching for “skruvdragare” (screwdriver) should also hit “skruvdragaren” and “skruv dragare”. Swedish inflections and compounds need analysis that understands the language, and it has to be tested on your data, not just switched on.
2. No relevance ranking. Results sorted by item number only help someone who already knows the item number. For everyone else, the best match should come first, and then it is a choice what carries weight: stock status, sales or something else that matters to you.
3. Filters that mirror the database, not the user. Facets built on how the data happens to be structured don’t help someone looking for “fits my model” or “in stock near me”. Filters should answer how people actually search.

How we build it

A search index, in the database or in a separate search service depending on what the tests show, with Swedish language support, synonyms and typo tolerance. Relevance ranking lives in configuration rather than code, so it can be adjusted when the business sees the wrong things ending up on top. A changed weighting still needs testing against real searches before it goes out, but it doesn’t need a new release of the system.

We also measure what people search for and what returns nothing. The zero-results list holds typos, synonyms you are missing, products you don’t stock and some junk. Go through it regularly and you see which of them are worth acting on.

Search builds on the same data model as databases and data warehouses. If you later want a language model to answer from your material, the same search has to fetch it, so it is worth getting right first. More under AI chatbots and RAG.