turbovec-sqlite sqlite-v0.3.0

0.3.0 — 2026-10-11

• Commit inserts and deletes without loading the compressed index. A new
  <name>_ids shadow table answers rowid conflict checks, rowid lookups, full
  rowid scans, and count(*); a new <name>_info row holds the row count and
  delta counts a commit needs. Connecting reads that one row. The index loads on
  the first search, or when a commit compacts. On a 1,000,000-row,
  1,536-dimensional index (745 MiB base), a one-row delete-and-insert commit
  from a fresh process fell from 468 ms and 830 MiB peak RSS to 0.8 ms and
  24.5 MiB, with a warm page cache. The first search costs what it did.
• Upgrade 0.2.0 storage in place on its first write: the commit creates _info
  and _ids from the index it already loaded and empties <name>_meta. This
  took 1.1 s on the index above. Reads alone do not upgrade. The upgrade is
  one-way: 0.2.0 then refuses the index with "Query returned no rows", including
  a connection opened before the upgrade, so it cannot write without
  maintaining _ids. Chunk and delta formats are unchanged.
• Record a storage format number in _info and refuse an unknown one by
  number.
• Fix a savepoint rollback that lost a replaced row. On a base under 1 MiB, an
  insert after an INSERT OR REPLACE dropped its vector, so a later delete took
  its destructive checkpoint after the replacement, and rolling back past the
  checkpoint undid the replacement by removing the row. Inserts now keep their
  vectors after any delete or replacement. 0.2.0 has this bug. The test script now
  runs the transaction model check for 300 seeds of 30 steps, which reproduces
  it; the previous 20 seeds of 20 did not.
• Pin the extension on Windows so it stays loaded after its last connection
  closes. SQLite unloads it then, but every search starts TurboVec's rayon
  worker threads, which keep running code from the library and fault once
  Windows unmaps it. Linux and macOS already keep it loaded. 0.2.0 has this
  too; a test now checks that the library outlives its connections.
• PRAGMA integrity_check compares _ids and the _info counts with the
  index. A load checks the index's row count against _info.
• Each insert and delete writes one _ids row, which total_changes() counts.
  A 100,000-row bulk insert in one transaction took 11% longer (1.29 s to
  1.43 s).

turbovec-sqlite sqlite-v0.2.0

0.2.0 — 2026-09-26

• Bring a warm reader forward by replaying only the delta batches other
  connections committed, instead of rebuilding the whole index. A reader's first
  query after a concurrent one-row commit dropped from 7.6–12.3 ms to about
  1.4 ms on a 20,000-row by 1,536-dimension index; the saving grows with index
  size. A compaction, a gap in the delta chain, or any other surprise still
  falls back to the full read.
• Cap the delta budget at the base image. A high-dimensional index pays a fixed
  rotation and codebook cost — an empty 1,536-dimension base already serializes
  to 1.67 MiB — so the 16 MiB floor alone let small tables accumulate raw
  float32 deltas several times larger than the 4-bit base they described. A
  fresh 2,000-row by 1,536-dimension table now stores 3.21 MiB instead of
  13.97 MiB, and its cold open plus first query fell from 27 ms to 1 ms. Small
  commits on a large index are unchanged.
• Stream base chunks straight into TurboVec's loader instead of materializing
  the image first. IdMapIndex::from_bytes copies what it is handed, so opening
  an index held two full copies; peak growth on a 48.5 MiB image fell from
  104 MiB to 61 MiB.
• Plan joins whose MATCH is unusable in one loop order. xBestIndex now
  declines those orders instead of failing the statement, so batch KNN over a
  table of queries works: queries q join v on v.embedding match q.embedding and v.k=n. A query no order can serve reports SQLite's "no query solution"
  rather than a tailored message.
• Report the planned top-k as the KNN plan's estimated row count instead of a
  fixed 10, and keep xBestIndex off the blocking path when another statement
  holds the warm index.
• Free the previous virtual-table error message before replacing it. SQLite's
  savepoint callbacks never reclaim it, so a failing xSavepoint, xRelease,
  or xRollbackTo leaked its message.
• Fail the build if a future libsqlite3-sys grows sqlite3_module, which
  would silently stop PRAGMA integrity_check from reaching turbovec0.
• Assert in the transaction model check that each rowid holds the vector version
  the model expects, not only that the rowid exists. Exercise turbovec0
  creation, insertion, search, and damaged-storage reporting from the
  statically linked C smoke test.
• Say why turbovec0 is unavailable on a host below SQLite 3.38, naming the
  required and found versions, instead of reporting a missing internal callback.
  A statically linked build inherits whatever libsqlite3 it was linked
  against, so this is the message that path produces; the scalar turbovec_*
  functions keep working there.

turbovec-sqlite sqlite-v0.1.6

0.1.6 — 2026-09-22

• Make small commits proportional to changed vectors instead of total index
  size. Commits append checksummed operation batches; bounded lazy compaction
  occasionally folds them into the existing format v7 base image.
• Remove the full-index copy from ordinary deletes and replacements. Actual
  destructive savepoint rollback rebuilds from committed storage and replays
  the retained transaction prefix.
• Extend turbovec_info() with base, delta, and pending-operation byte counts.
• Add a 700,000-row by 1,536-dimension WAL benchmark. It times mutation and
  commit separately, compares released and candidate search results exactly,
  and covers single edits, scattered and neighboring deletes, replacements,
  and repeated churn.
• Existing index files open without rebuilding. A database with unmerged 0.1.6
  deltas intentionally refuses to open in 0.1.5 or older instead of silently
  returning stale results.

turbovec-sqlite sqlite-v0.1.5

0.1.5 — 2026-09-15

• Upgrade from 0.1.4 without rebuilding indexes. The SQL API and TurboVec
  format v7, revision 2 remain compatible.
• Reject invalid shadow-table lengths, chunk sequences, and chunk types before
  allocating an index read buffer; report allocation failure as a SQLite error.
• Preserve last_insert_rowid() across shadow-table writes, including failed
  commits that have already written some chunks.
• Let SQLite handle duplicate INSERT OR IGNORE constraints so affected-row
  counts and last-insert rowids reflect successful inserts only. Document the
  host SQLite RETURNING limitation and a query that avoids it.
• Stream committed indexes through a 4 MiB buffer and compare one old chunk at
  a time. Preserve the serialized bytes and changed-byte-span writes. Commit
  work still scales with index size; the full rollback checkpoint for destructive
  writes is unchanged.
• Add storage regressions for corruption, duplicate bookkeeping, chunk growth
  and shrinkage, byte compatibility, WAL snapshots, and partial commit failures.

turbovec-sqlite sqlite-v0.1.4

0.1.4 — 2026-08-24

• Prepared the repository and install path for public use.
• Added a private security-reporting policy and third-party license notices.
• Made the static-link smoke portable across Apple and GNU archive tools.
• Reduced routine pull-request CI to Linux quality and SQLite compatibility
  gates; tags and manual runs retain the complete five-platform release matrix.
• Normalized contributor identity display through .mailmap.

turbovec-sqlite sqlite-v0.1.3

0.1.3 — 2026-08-24

• Added turbovec_info(table) diagnostics and explicit format v7 revision 2
  checks alongside turbovec_version().
• Added static libraries, a C header, and a static-registration smoke test.
• Added a fixed real-embedding recall gate using a public-domain GloVe fixture.
• Updated the Go client to load every pooled connection with ConnectHook and
  demonstrate filtered KNN.
• Reworked the README around install, core SQL, operating costs, drift recovery,
  compatibility, and release consumption.
• Moved architecture and benchmark notes under docs/ and the CLI demo under
  examples/.

turbovec-sqlite sqlite-v0.1.2

0.1.2 — 2026-08-24

• Made xBegin and xSavepoint metadata-only for insert transactions.
• Added change-log rollback for inserts and lazy checkpoints for deletes and
  replacements.
• Added nested swap-and-pop rollback coverage and an FTS5 savepoint-cost gate.
• Fixed OR FAIL and other conflict policies by passing SQLite's required
  SQLITE_VTAB_CONSTRAINT_SUPPORT argument explicitly.

turbovec-sqlite sqlite-v0.1.1

Changelog

0.1.1 — 2026-08-24

• Added writable turbovec0 virtual tables backed by SQLite shadow tables.
• Added transaction, WAL, savepoint, integrity, rename, and reopen coverage.
• Added native rowid IN (...) pushdown for filtered nearest-neighbor search.
• Added repeatable correctness, performance, and sqlite-vec comparison tools.
• Added Python, JavaScript, and Go examples plus native release packaging.
• Fixed extension loading against the supported SQLite 3.44 runtime API.
• Improved planner errors and documented the supported KNN query shapes.