sqlite_plus_vec 0.1.7

Brute-force vec0 vector search for sqlite_plus, with exact filtered ranking.

Changelog

0.1.7

Depends on sqlite_plus ^0.1.14. The library behavior is the 0.1.6 release.

0.1.6

Depends on sqlite_plus ^0.1.13. The library behavior is the 0.1.5 release.

0.1.5

Depends on sqlite_plus ^0.1.12. The README opens points.db with openFileImage. The library behavior is the 0.1.4 release.

0.1.4

Depends on sqlite_plus ^0.1.11. The library behavior is the 0.1.3 release.

0.1.3

Depends on sqlite_plus ^0.1.9. The README opens on the distance. The library behavior is the 0.1.2 release.

0.1.2

Depends on sqlite_plus ^0.1.8.

rankVec0 scores a vec0 shadow table one page at a time. Pass the eligible row ids and every one of those rows gets a distance; an empty list reads nothing. embedding MATCH ? AND k = ? is unchanged and still applies k before a later SQL filter. A MATCH scan keeps the embedding bytes of the returned neighbors rather than a second copy of every stored blob.

0.1.1

Depends on sqlite_plus ^0.1.7. vec0 is unchanged. The README links the other packages on pub.dev and shows a short example of each.

0.1.0

Brute-force vec0 virtual table for one little-endian float32 column.

• SqlitePlusVec.register accepts a SqliteDatabase, a MutableSqliteDatabase, or the connection's SqlFunctionRegistry.
• CREATE VIRTUAL TABLE ... USING vec0(embedding float[N] distance_metric=l2|cosine) scans every row for embedding MATCH ? AND k = ?.
• vecDistance, encodeVec, and decodeVec are the same float32 math. Dimensions are 1 to 4096. k is 1 to 100000.
• An ORM table uses DbModules.vec0 and @DbVector. nearest returns the k closest rows.
• The scan is not an approximate index. The file is not a sqlite-vec database.