AionDB v0.3 turns vector search into a first-class engine feature

v0.3 is where AionDB's vector system becomes a product surface.

The release brings pgvector-style SQL, HNSW, IVF-flat, Qdrant-style filtered helper options, PostgreSQL ecosystem compatibility work, and a reproducible benchmark harness into the same engine. The goal is simple: one database for relational records, graph relationships, and semantic retrieval.

For RAG, recommendation, knowledge-base, support, and agent memory workloads, the interesting query is rarely just "nearest vectors." It is "nearest vectors inside this tenant, for this permission scope, related to this graph neighborhood, with these metadata constraints." AionDB v0.3 moves directly toward that shape.

The v0.3 Vector Stack

The pgvector-facing SQL surface now covers more of the syntax and helper behavior that application tooling emits:

CREATE EXTENSION IF NOT EXISTS vector;
VECTOR(n), unconstrained VECTOR, HALFVEC(n), and SPARSEVEC(n);
casts between arrays, vector, halfvec, sparsevec, and real arrays;
vector arithmetic and distance helpers;
binary quantization, Hamming distance, and Jaccard distance helpers;
pg_catalog.-qualified distance functions for ORM-generated SQL.

The index surface now has two ANN families:

HNSW for high-recall graph-based vector search;
IVF-flat for fast builds and low-latency approximate scans.

The helper layer now accepts Qdrant-style JSON options: must, should, must_not, match clauses, numeric ranges, nested JSONB paths, id filters, null checks, empty checks, payload controls, and vector return controls.

That is the product point of v0.3. AionDB goes beyond a distance function by bringing vector retrieval into the same place where the application already keeps tables, metadata, permissions, and relationships.

Benchmark Snapshot

The vector benchmark is now a repository-level workflow:

cd benchmarks/vector-compare
cargo run --release

Default run:

Setting	Value
Vectors	50000
Dimensions	96
Queries	200
k	10

Latest local output:

Backend	Build ms	Recall@10	Mean us	p50 us	p95 us	p99 us
AionDB HNSW raw	36571	0.996	9423	8790	15674	17233
AionDB HNSW PQ	74742	0.994	13072	12355	17659	18471
AionDB IVF-flat nprobe=8	418	0.466	809	827	1223	1766
AionDB IVF-flat nprobe=32	416	0.863	2572	2474	3603	3977
Brute-force exact	0	1.000	9930	9520	14293	18991

The headline numbers:

HNSW raw reaches 0.996 recall@10.
HNSW PQ reaches 0.994 recall@10.
IVF-flat builds in about 416-418 ms.
IVF-flat with nprobe=32 reaches 0.863 recall@10 at about 2.57 ms mean query latency.

Why It Matters

The normal vector database pattern is to split data: relational state in PostgreSQL, relationships somewhere else, embeddings in a vector service, and glue code between them. AionDB's model is different. The table stays the source of truth, graph labels add connected context, and vector indexes rank semantic matches beside the same metadata.

That is why v0.3 matters. It gives the product a stronger vector core without giving up the PostgreSQL tooling path that application teams already understand.