Vector Plus provides a single Postgres index access method, ivfplus. It's built for the same job as pgvector's ivfflat, approximate nearest-neighbor search over vector columns, but combines three techniques to stay fast and small at large row counts: hierarchical Inverted File (IVF) clustering to narrow a search down to a handful of candidate groups, RaBitQ residual quantization to shrink each vector to a compact code, and a fast-scan packing layout to scan those codes efficiently.
Hierarchical IVF clustering
Searching every row against a query vector would be accurate but slow at scale, so ivfplus clusters rows into groups, called lists, when the index is built, then only scans the lists closest to the query at search time.
The index is either flat (one level of lists) or two-level (lists are themselves grouped into a smaller set of top-level clusters, searched first to find which lists to scan). Which one you get is automatic: a table crosses into the two-level hierarchy once lists reaches 1,800. See Tuning and sizing for what the two shapes mean for build time and parallelism.
Building the clustering works top-down: the index clusters the whole table first, then partitions and subdivides those clusters, rather than building up from individual rows. It reuses pgvector's Elkan k-means implementation for each clustering pass, which keeps the cost of clustering proportional to the number of rows and clusters rather than growing faster than that. When scanning a two-level index, how many top-level clusters get checked (before deciding which lists to scan) scales with how many probes you've set: the exact relationship is outer_probes = K_probes * sqrt(probes), where K_probes is a fixed constant (see Limits and fixed constants).
Residual quantization
Quantization shrinks each vector to a compact code so more of the index fits in memory and comparisons run faster, at the cost of some precision recovered later during reranking. RaBitQ encodes a 1 bit/dimension residual against the assigned centroid, plus {norm, bias, dot_cv, err_fac}, estimated asymmetrically at scan time with an optimal-stopping rule.
The rerank step is exact, against an inline f16 copy by default, or against the heap directly on a server that supports the executor rerank policy. See Index-rerank-policy builds.
A randomized Hadamard rotation (deterministic seed, no stored table) runs before encoding, which is what makes the 1-bit code work on nonisotropic embeddings. It isn't configurable — every index applies it.
Fast-scan packing
Fast-scan packing arranges quantized codes in memory so a scan can compare many of them at once instead of one at a time. Leaf lists use two packing formats, chosen by how the data arrived:
- Frozen packed — 32-wide
[u8;16]blocks with a 4-bit pshufb-style lookup table, used for bulk-built data. - Append unpacked — Bit-sliced, used for rows inserted after the index was built.
Rows inserted after the initial build land in the append region and are read by a slower scan until the index is rebuilt. See Writes, VACUUM, and REINDEX for how REINDEX reclaims this.