What's New in v0.3

AionDB v0.3 is the vector update. It turns vector search from a feature checkbox into a first-class product surface: pgvector-style SQL, HNSW, IVF-flat, Qdrant-style filtered retrieval, PostgreSQL ecosystem compatibility, and reproducible vector benchmarks all land in the same release line.

The message is direct: AionDB now gives application builders one engine for relational records, graph relationships, and semantic retrieval. The vector stack lives beside the SQL catalog and graph model instead of forcing the application to split data between a database, a graph service, and a separate vector store.

New in v0.3: pgvector-facing SQL, two ANN index families, Qdrant-style filtered vector helpers, and benchmarked recall/latency numbers from the repository.

Start with the current benchmark snapshot in v0.3 Vector Performance.

Product Highlights

v0.3 delivers five major upgrades:

broader pgvector-style SQL for vector, halfvec, sparsevec, bit helpers, casts, functions, and ORM-generated catalog lookups;
HNSW search tuned for high recall on raw vectors and product-quantized paths;
IVF-flat indexing with fast builds, configurable probes, parallel search work, and pgvector-style DDL;
Qdrant-style JSON filter options for metadata-aware vector retrieval;
a standalone vector comparison harness that publishes build time, recall@k, and query latency.

Vector SQL

The vector surface now speaks the SQL shape expected by PostgreSQL ecosystem tools:

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 operators and distance helpers;
binary quantization, Hamming distance, and Jaccard distance helpers;
pg_catalog.-qualified distance functions for generated SQL.

This is the practical bridge for teams that already prototype RAG, recommendation, similarity, or hybrid search on the PostgreSQL stack and want those vectors to sit next to graph and relational context.

Useful references:

ANN Indexes

v0.3 gives AionDB two vector ANN families.

HNSW brings graph-based approximate search with:

raw f32 search;
product quantization search with exact rescoring;
scalar and binary quantization paths;
search statistics and oversampling controls;
stronger search floors for large graphs.

IVF-flat brings clustered approximate search with:

USING ivfflat DDL compatibility;
configurable list and probe counts;
contiguous centroid storage;
parallel build assignment;
parallel large-list scanning;
per-list candidate trimming;
L2 ordering without square roots on the IVF hot path.

The result is a real choice: HNSW for very high recall, IVF-flat for fast builds and low-latency approximate scans, and exact brute force as the ground-truth reference.

Qdrant-Style Filters

AionDB v0.3 also moves vector search closer to real application retrieval. The vector helper functions accept Qdrant-style JSON options for filtered search, including:

must, should, and must_not clauses;
match.value, match.any, match.except, and match.text;
numeric range filters;
values_count;
nested JSONB paths such as payload.tags[] and payload.cities[].name;
has_id;
is_null and is_empty;
with_payload include/exclude controls;
with_vector and with_vectors.

That gives RAG and recommendation workloads the normal retrieval controls they need: tenant filters, permissions, metadata constraints, freshness rules, and graph-derived subsets before ranking.

Benchmark Snapshot

The v0.3 vector benchmark is built into the repository:

cd benchmarks/vector-compare
cargo run --release

Latest local run:

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 standout v0.3 story:

HNSW raw reaches 0.996 recall@10.
HNSW PQ keeps 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 time.
Exact brute force stays in the harness as the recall reference.

See:

Ecosystem Momentum

v0.3 also advances the PostgreSQL ecosystem route. The release improves the compatibility path for:

ORM-generated SQL;
migration-tool introspection;
pgvector extension metadata;
PostgreSQL casts and function lookup;
vector operators and helper functions;
typed JSONB payload behavior in vector helpers.

For builders, that is the whole point of AionDB: keep SQL tooling, add graph context, add vector retrieval, and run the application against one local engine.