How search works

How the default search and the optional fuzzy search work, how results are ranked, and what each costs.

There is one search box. An identifier is looked up exactly; anything else is a name search over the active subjects. Names are compared without diacritics, without case and without the legal form, so Zlatá Rybka, s.r.o., zlata rybka and ZLATÁ RYBKA a.s. are the same words.

Default search

The default search answers as you type. It runs two steps:

Name prefix: names that start with the query. zlata ryb finds Zlatá Rybka, s.r.o..
Word prefixes: every word of the query must be the start of a word of the name, in any order. rybka zlata and ryb find it too.

The second step runs when the first one returned fewer than 10 results, or when the rarest query word is rare enough to be cheap. It is driven by the rarest word and capped, so a very common word (obec, slovenska) does not make the query expensive.

Ranking: an exact name first, then names that contain all query words as whole words, then names that start with the query, then shorter names. There is no popularity signal in the data, so the register does not invent one.

Fuzzy search (optional)

mode=fuzzy corrects typos. It is never automatic: call it when the default search found nothing and the user asks for it, for example with a “Search with typos?” button.

Keep known words

Query words that are the start of a known word stay as typed.

Correct unknown words

Each unknown word is compared with the dictionary of all words in the register: first by trigrams, then by edit distance with adjacent transpositions. sporitelan → sporitelna is one transposition. A word of up to four letters may have one typo, a longer word two; a short word is checked against every word one edit away, since it has too few trigrams for the first step.

Search with the corrections

The corrected words go through the default word search. The response lists the corrections.

Fuzzy search is optional and potentially CPU-intensive, and it never substitutes exact identifier matching. Operators can switch it off (REGISTER_SEARCH_FUZZY=false); /v1/capabilities tells clients whether it is on. It has its own rate limit, a limit on concurrent fuzzy searches and a timeout.

What it costs

Measured on the real data (1 017 893 active subjects) on a small database:

	Default search	Fuzzy search
Typical	1–5 ms	3–5 ms
Worst queries, full machine	under 10 ms	about 60 ms
Worst queries, 1/8 vCPU database	about 50 ms	hundreds of ms

When the database is saturated, a 1/8 vCPU tier answers about 30 default searches a second and a 1/4 vCPU tier 75–140. With a client debounce of about 120 ms, a person typing sends one to three searches a second.

What search does not do

Terminated subjects are not in name search. An exact identifier finds them, flagged active: false.
Natural persons without an IČO (taxpayers known only by DIČ) are never in search.
A subject kept out of search after an objection is still found by its exact identifier.
Results are capped at 20; there is no paging in v1.
Inflected forms are different words: lipa does not find “U Lipy”. Type the start of the word (lip), or use fuzzy search for one changed letter.