Proximum

Version-controlled vector indexing

Proximum is a pure-JVM HNSW index with immutable snapshots, branches and Merkle verification. It has no native runtime dependency.

What Proximum provides

Snapshots: Immutable index snapshots in O(1) (structural sharing).
Independent reads: Share immutable snapshots across workers without coordinating through a connection pool.
Reproducibility: Query historical index states for audits, debugging, and compliance workflows.
Verification: Merkle-hashed indices for verifiable, content-addressed retrieval states.
Storage: Pluggable persistence via Konserve with backends tailored to your environment.

Published SIFT-1M result with one million 128-dimensional vectors on an Intel Core Ultra 7:

Implementation	Insert (vec/s)	Search QPS	p50 Latency	Recall@10
Proximum	13,392	3,844	264 µs	98.63%
jvector	9,771	3,609	277 µs	95.95%
datalevin/usearch	2,492	3,616	268 µs	96.96%
Lucene HNSW	2,395	3,036	340 µs	98.53%
hnswlib-java	4,260	1,007	1,033 µs	98.29%

Insertion: In this run, 37% higher throughput than jvector and 5.6× Lucene's throughput
Recall: The highest recall@10 among the pure-JVM implementations in this run
Pure JVM: SIMD via Java Vector API, no native dependencies
Immutable: Branches structurally share unchanged index data

Full benchmarks including DBpedia-OpenAI-100K and Glove-100 in the GitHub README →

First-class JVM integrations ship alongside the core API.

Spring AI: Drop-in integration for Spring-based RAG and retrieval pipelines.
LangChain4j: Adapter for LangChain4j-based applications and evaluators.
APIs: Full Java API and full Clojure API.

Run Proximum in the same JVM as your service, without operating a separate cluster.

Use the same API with local, cloud or replicated Konserve storage backends.

If you need help getting Proximum into production, we can help with integration, custom development, and support contracts.