asg017/sqlite-vec

What it does

sqlite-vec is an extension you load into SQLite, written as a single portable C file, designed to run "anywhere SQLite runs" — Linux, macOS, Windows, inside a browser, and on small single-board computers. You create a vec0 table, put number lists in it, and query for nearest neighbours; you can keep ordinary columns alongside them for filtering and partitioning. It handles three storage sizes: full precision, eight-bit, and a one-bit form whose speed difference the author published himself — on one million long number lists, full precision took 8.52 seconds and the one-bit form returned in 124 milliseconds. It ships as a ready-made package for Python, Node, Ruby, Go and two data tools, so there is normally nothing to compile. It does not build an approximate index, which is the central thing to understand: the author states the project "is currently focused on really fast brute-force vector search", so search time grows with the number of rows rather than staying flat. Static read-only blobs do not support inserting, updating or deleting. And it does not let you add a column to a table that already exists — the numbers live in their own table, which makes queries more involved. Its own README says plainly: "sqlite-vec is a pre-v1, so expect breaking changes!"

Why it matters

Who it suits. Anyone building search-by-meaning into a program that already uses SQLite, where the collection is tens or hundreds of thousands of items rather than tens of millions, and where running a separate database server would be absurd. It suits a desktop application, a phone application, or a small service particularly well, because there is no server at all. Skip it if the collection is large and growing: with no approximate index, the cost of a query rises with the row count, and the author's own figures show a million long number lists needing about 6 GB of memory to hold.

What people say. This is the entry with genuinely strong outside sources, and the most important one is bad news. Dan Kennedy, one of SQLite's own core developers, posted on the SQLite user forum on 30 March 2026 that SQLite upstream is building its own vector extension: "There is still no vec1 release, but we are getting closer", with work remaining on processor-specific optimisation and testing he calls "woefully inadequate". The documentation for that extension says it "provides approximate nearest-neighbor (ANN) vector search", is "implemented in portable C and has no external dependencies", uses a named indexing method, and that "The current release is version 0.7" with "No further features are required before a 1.0 release." Neither page mentions sqlite-vec — and what SQLite is building is exactly the approximate indexing sqlite-vec does not have. The other substantial source is Marco Bambini's piece of 1 September 2025, which calls sqlite-vec "A clean re-implementation in C, easier to maintain, exposing a brute-force search strategy" and notes its table design "means vectors must live in separate tables and queries become more complex" — and he discloses in the article, as we do here, that he founded a company selling a competing extension and that his team first sponsored sqlite-vec. Simon Willison wrote two useful how-to pieces in August and October 2024, but he authors the two data tools sqlite-vec ships plugins for, so he is an interested neighbour, and both pieces are two years old.

Verdict. Use it today if the collection is small and the simplicity is the point — one file, no server, and it works in a browser. But go in knowing three things, and the third is the one nobody tells you. It is still before version 1.0 two and a half years in, with 52 unmerged pull requests and the last code landing on 18 May 2026 — about five months ago, well inside the line this report draws, and we are explicitly not calling it dormant. A third party has forked it, stating in the fork's README that it exists "to merge pending upstream PRs and provide community support while the original author is unavailable" — that is the fork's claim and not a verified fact, but it points the same way. And SQLite's own authors are now shipping a competing extension at version 0.7 that does the approximate indexing this one lacks, which will reshape the whole category when it lands. Its own predecessor, asg017/sqlite-vss, is the cautionary note: 2,004 stars, not archived, and no code since 5 May 2024, with a README that tells readers to come here instead.