SQLite-Vector

Why it's ranked here

The case rests on honest numbers and portability. The project publishes a benchmark on one million 768-dimension vectors: exact search at 484 ms per query, the 8-bit signed mode at 37.6 ms with 99.5% recall at 20, or 114 ms when held to a 30 MB memory budget. The benchmark tooling refuses to print paste-ready rows for nonstandard runs and stamps the real SIMD backend itself, so the table is hard to fake. It ships Apache 2.0, with prebuilt binaries for Linux, macOS, Windows, Android and iOS, plus npm and Python packages and a WebAssembly build.

What's good

The documentation tells you when not to use its own features. It warns that the unsigned 8-bit mode collapses to 33.8% recall under cosine distance, that the default picks that mode for non-negative data, and that the 1-bit mode is a pre-filter at 10% recall, not an answer. It also concedes TurboQuant saves memory but is slower than plain 8-bit. Concurrency is handled with care: a preloaded index is reference-counted, so reloading or dropping it cannot free memory a running query is still reading. The build file documents that one test target silently skips the AVX kernels and adds a second target that actually tests them.

Tradeoffs

Search is linear in table size by design, so latency grows with every row; the only published hardware result is one Apple M5 Pro. Recall figures come from uniform random data, which the project calls a floor, so you must measure your own. The compressed copy goes stale when data changes and must be rebuilt by hand. Several options are trust-me assertions: declaring vectors normalized gives wrong distances if they are not, and binary input is not validated. Initialization must be repeated on every connection, at most 128 vector columns can be registered, and the Windows binary is x86 only. The README also funnels toward a paid managed cloud.

How to use it well

Fits apps that already live in SQLite, especially on phones, desktops or edge boxes, where a few hundred thousand to a million embeddings must be searchable offline without a separate server. Pick the signed 8-bit mode for cosine and set it explicitly, preload it if memory allows, and verify recall against the exact scan with the SQL recipe the docs supply. Use streaming mode when you need to filter by other columns. Rebuild the quantized copy after bulk writes. For tens of millions of vectors or sub-millisecond latency, a real approximate index elsewhere is the better fit.

Observed

License

Apache-2.0 (LICENSE.md, also declared in both package manifests)

Language

C, built as a loadable SQLite extension via a Makefile

Interfaces

SQL scalar functions plus two table-valued scan functions; no server or CLI of its own

Packaging

npm package with per-platform binaries, Python platform wheels, WebAssembly build, prebuilt binaries

Platform support

Linux x86 and ARM, macOS x86 and ARM, Windows x86, Android, iOS

SIMD backends

SSE2, AVX2, AVX-512, NEON, RISC-V vector, with a scalar CPU fallback

Search method

Brute-force scan over full or quantized vectors; no approximate graph index

Tests

C unit test suite in test/test_vector.c plus a SIMD-backend test target