ParadeDB 0.26.0 ships vector pushdown for RRF and a vector tie-break

ParadeDB published v0.26.0 on 3 October 2026 at 05:40 UTC. The tag is not a prerelease. The release notes are a long list of pull requests. Two of them are the hybrid-search changes the project highlights as features: vector pushdown for an RRF query, and tie-break ordering on a vector distance. Both pull requests merged in July. The October date is when they shipped in a tagged release, not when the patches were written.

RRF no longer has to read the whole table

Pull request 5750, merged 31 July 2026, says Postgres sets limit_tuples to zero when a window aggregate sits between LIMIT and the scan, because a window function usually needs every row. The usual reciprocal-rank-fusion shape, a per-branch RANK() next to ORDER BY ... LIMIT, therefore read and sorted the full corpus on both the text side and the vector side. The change overrides that when the window is only a ranking in the same order as the limit, so the top N rows are the first N in window order. It applies only when every window function is position-only (row_number, rank, or dense_rank), there is no PARTITION BY, and the window’s ORDER BY matches the query’s ORDER BY. The pull request says that case is result-preserving. It does not claim a speedup number.

A second sort key can ride with the distance

Pull request 5772, merged 30 July 2026, says ORDER BY embedding <=> '[..]', other_field can be pushed down. The body of the pull request is that one sentence plus the title “Support tiebreak ordering with vector.” There is no benchmark attached. If your hybrid query sorted by distance and then by a column such as an id or a timestamp, this is the change that is supposed to keep that plan in the vector index path. Confirm it against EXPLAIN on your version before you depend on it.

The rest of the tag

The same notes bump pgrx to 0.19.2, move the workspace to Rust edition 2024, and include other vector commits (a paradedb.vector_clustering_threshold GUC defaulting to 500, a cap of 4 workers on vector index builds, and a change that passes training vectors by value). One commit title mentions a bounds-gate change and “(0.25.1)”, so that change belongs to 0.25.1 rather than to this tag. The release is much larger than the two search features; the notes also contain CI, docs, and Docker chores.

Practical takeaway

Upgrade to 0.26.0 if you run ParadeDB hybrid queries that combine RANK() and LIMIT, or that order by a vector distance and a second column. The pushdown is conditional on the window shape above. The July merge dates mean a main checkout from August may already have had the code; the tag is what makes it a release.