spellfix1

Package spellfix1 implements a fuzzy-text-matching virtual table inspired by SQLite’s spellfix1 extension. Build a vocabulary by inserting words, then query with WHERE word MATCH ? to retrieve the closest matches by edit distance, grouped phonetically.

CREATE VIRTUAL TABLE vocab USING spellfix1;
INSERT INTO vocab(word) VALUES ('apple'), ('banana'), ('cherry'), ('aple');
SELECT word, distance FROM vocab WHERE word MATCH 'aple' LIMIT 3;
-- aple    0
-- apple   1
-- ... etc.

For a typed Go handle over this vtab — Create / Add / AddMany / Size / Correct / CorrectSQL / Drop, mirroring vec.Table and fts.Index — see Vocab. The raw SQL above and the typed API are interchangeable.

word      TEXT     -- the candidate vocabulary entry
rank      INTEGER  -- caller-supplied frequency score (0 if not given)
distance  INTEGER  -- Damerau-Levenshtein distance from the query
score     INTEGER  -- distance*1024 - rank (lower is better)
matchlen  INTEGER  -- length of the query prefix consumed
phonetic  HIDDEN   -- the Soundex key (4-char A-Z)
top       HIDDEN   -- caller-supplied LIMIT override
scope     HIDDEN   -- max distance bound (default 4)
srchcnt   HIDDEN   -- how many candidates the planner examined
soundslike HIDDEN  -- alternative search-spelling (rarely used)

• Cyrillic / Greek / other non-Latin transliteration tables.
• Caller-tunable cost matrices (the editdist3 cost-matrix API).
• The Russian-language phonetic encoder.
• The editdist1 SQL function (use (SELECT distance FROM vocab WHERE word MATCH ? LIMIT 1) to get the distance directly).

We use Soundex for phonetic grouping (well-defined, simple, language-agnostic) and Damerau-Levenshtein for distance (covers transpositions in addition to insert/delete/substitute).

The vocabulary lives in a shadow table named <vtab>_storage on the same schema. Vocabulary survives db.Close() / reconnect.

Ported in spirit from SQLite spellfix1 — same SQL surface, simpler algorithm.

Full API: pkg.go.dev/gosqlite.org/ext/spellfix1