SB-SIMD Semantic ENN Search Engine

A stateless semantic Exact NN engine for dynamic subsets.

Perfect for workloads where every query might apply heavy filters, producing unique subsets, or even stream-like data that changes every second. ANN indexes struggle here, but SB-SIMD stateless ENN can simply rebuild on the fly. Eg. ideal for streams, RAG environments, complex website filters.

I have a pet project where I’ve been struggling a lot with one particular challenge: semantic dynamic search. The essence of similarity search is that it doesn’t look for exact matches, but for the closest results.

For example, if a user types “I love white sneakers”, the engine doesn’t just look for the exact keyword “white sneaker”, but understands semantic similarity. It can also return products described as “beloved white sneaker” by others, or where the description only says “casual sneaker white”.

Now, think about how many options this opens up in an e-commerce webshop. In my case, perhaps even more — since the data can also come dynamically through graphs.

Result is 0.5 - 5 ms overhead per query vs no index time.

👉 What you see here is the solution I’ve built for this pet project.

Perhaps you’ve already encountered the **HNSW filtering problem**. HNSW is extremely fast for large static datasets, but as soon as you apply **dynamic filters** (e.g. `WHERE user_id=...` or `category=...`), performance degrades dramatically:

The index is built globally, not per-filter.
When you restrict to a subset, the graph connectivity is no longer optimal.
The engine often ends up scanning far more nodes than expected, so query latency can jump from <1 ms → 20–100 ms.
Workarounds (pre-building per-filter indexes, or post-filtering results) are either memory-intensive or increase latency further.

Why this matters

This is why HNSW works great for global, static similarity search,
but it’s not ideal for highly dynamic, filter-heavy workloads.

By contrast, a stateless ENN (Exact Nearest Neighbor) engine, implemented here in Clojure, can rebuild on the fly for each filtered subset. Overload max 0,5ms - 5ms in normal cases.

That makes it practical for use cases with:

rapidly changing datasets,
stream-like data,
or queries that always narrow the search space dynamically.

✅ Best fit use cases

Small / medium subsets (< 100k, max. 700ms ~ 1Mx768D)

When user queries always filter down the space to 1k–10k vectors, the rebuild + search fits easily within low latency budgets.

Extreme filtering / rapidly changing custom subsets

Perfect for workloads where every query might apply heavy filters, producing unique subsets, or even stream-like data that changes every second.
(ANN indexes struggle here, but a stateless ENN can simply rebuild on the fly.)

Ephemeral / ad-hoc searches

Ideal for real-time logs, sensor data, or session-based datasets where the relevant subset changes constantly and indexing on-the-fly is more efficient.

Edge / mobile / microservices

Lightweight in-memory search when Postgres, Milvus, or Qdrant are too heavy.

Rapid prototyping / experimentation

When you need quick iteration without maintaining a persistent index.

⚠️ Less suitable

Very large, stable datasets

Millions of vectors (> 1M), where persistent ANN engines dominate with <10 ms query times.

Database-first environments

When transactional guarantees, SQL integration, or auditability are required.

Not a competitor to pgvector or Milvus, but rather a niche complement:

“A stateless ENN engine for dynamic subsets.”

Performance

Scaling Performance

Rebuild Cost Comparison

Benchmark Results

With SimSIMD Native Library on Apple M-series (10 threads):

Dataset	Vectors	Dimensions	Search Time	QPS
SIFT	10K	128	0.99ms	829
SIFT	100K	128	3.90ms	210
SIFT	1M	128	38.83ms	23
BERT	10K	768	0.95ms	564
BERT	100K	768	9.88ms	98
OpenAI	10K	1536	1.33ms	729
OpenAI	100K	1536	13.62ms	72
Bible	31K	768	2.83ms	345

Installation

Prerequisites

Java 11+
Clojure CLI tools

Setup

# Clone the repository
git clone https://github.com/yourusername/sb-simd.git
cd sb-simd

Usage

Quick Start

# Run benchmarks
java -cp "classes:simsimd.jar:$(clojure -Spath)" clojure.main -m sb-simd.benchmark

# Run specific test
java -cp "classes:simsimd.jar:$(clojure -Spath)" clojure.main -m sb-simd.benchmark SIFT-100K

As a Library

(require '[sb-simd.core :as simd])

;; Load embeddings
(def index (simd/create-index "data/embeddings.json"))

;; Search
(def results (simd/search {:embeddings (:embeddings index)
                           :query query-vector
                           :k 10
                           :dimension 768}))

;; Results include similarity scores and indices
(:results results) ; => [[0.95 123] [0.94 456] ...]
(:time-ms results) ; => 3.14
(:qps results)     ; => 318

Data Format

Embeddings should be in JSON format:

[
  {
    "id": "item1",
    "text": "Example text",
    "embedding": [0.1, 0.2, ...]
  },
  ...
]

Or with metadata:

{
  "metadata": {...},
  "verses": [
    {
      "id": "item1",
      "text": "Example text",
      "embedding": [0.1, 0.2, ...]
    }
  ]
}

Architecture

SB-SIMD uses a brute-force approach with several optimizations:

SIMD Operations - Hardware-accelerated cosine similarity via SimSIMD
Parallel Search - Dataset split across threads for concurrent processing
Heap-based Top-K - Efficient tracking of best matches
Zero-copy Arrays - Direct memory operations on float arrays

Comparison with Other Methods

Method	Build Time	Search@100K	Recall	Update
SB-SIMD	0ms	3.9ms	100%	Instant
FAISS-Flat	100ms	7.7ms	100%	Rebuild
HNSW	8.3s	0.15ms	99%	Incremental
Annoy	611ms	0.05ms	95%	Rebuild

Examples

See the examples/ directory for benchmark scripts:

cd examples
./run-all.sh    # Run all benchmarks
./run-sift.sh   # SIFT dimension tests
./run-bert.sh   # BERT dimension tests

Native Libraries

The project includes pre-compiled native libraries for macOS ARM64:

libsimsimd.dylib - SimSIMD core library
libsimsimdjni.dylib - JNI wrapper

For other platforms, the system falls back to pure Java implementation (slower but functional).

BERT-10K (768D)

Method	Build (ms)	Rebuild 1%	Rebuild 5%	Rebuild 10%	Search P50 (ms)	Recall	Total (10%+Search)
SB-SIMD	0	0.0	0.0	0.0	0.756	100%	0.76
FAISS-Flat	2.4	0.02	0.12	0.24	0.44	100%	0.68
HNSW	2100	21.0	105.0	210.0	0.35	~99%	210.35
Annoy	350	353.5	367.5	385.0	0.02	~95%	385.02
FAISS-IVF	115	1.15–5.75	5.75–28.75	11.50–57.50	0.08	~98%	11.58–57.58

BERT-100K (768D)

Method	Build (ms)	Rebuild 1%	Rebuild 5%	Rebuild 10%	Search P50 (ms)	Recall	Total (10%+Search)
SB-SIMD	0	0.0	0.0	0.0	7.088	100%	7.09
FAISS-Flat	39.4	0.39	1.97	3.94	5.05	100%	8.99
HNSW	48000	480.0	2400.0	4800.0	0.9	99%	4800.9
Annoy	3600	3636.0	3780.0	3960.0	0.05	95%	3960.05
FAISS-IVF	1500	15.00–75.00	75.00–375.00	150.00–750.00	0.55	98%	150.55–750.55

BERT-1M (768D)

Method	Build (ms)	Rebuild 1%	Rebuild 5%	Rebuild 10%	Search P50 (ms)	Recall	Total (10%+Search)
FAISS-Flat	372.9	3.73	18.64	37.29	51.33	100%	88.62

OpenAI-10K (1536D)

Method	Build (ms)	Rebuild 1%	Rebuild 5%	Rebuild 10%	Search P50 (ms)	Recall	Total (10%+Search)
SB-SIMD	0	0.0	0.0	0.0	1.167	100%	1.17
FAISS-Flat	4	0.04	0.2	0.4	0.89	100%	1.29
HNSW	4200	42.0	210.0	420.0	0.7	99%	420.7
Annoy	700	707.0	735.0	770.0	0.04	95%	770.04
FAISS-IVF	230	2.30–11.50	11.50–57.50	23.00–115.00	0.16	98%	23.16–115.16

OpenAI-100K (1536D)

Method	Build (ms)	Rebuild 1%	Rebuild 5%	Rebuild 10%	Search P50 (ms)	Recall	Total (10%+Search)
SB-SIMD	0	0.0	0.0	0.0	12.24	100%	12.24
FAISS-Flat	75.7	0.76	3.79	7.57	9.39	100%	16.96
HNSW	96000	960.0	4800.0	9600.0	1.8	99%	9601.8
Annoy	7200	7272.0	7560.0	7920.0	0.1	95%	7920.1
FAISS-IVF	3000	30.00–150.00	150.00–750.00	300.00–1500.00	1.1	98%	301.10–1501.10

SIFT-10K (128D)

Method	Build (ms)	Rebuild 1%	Rebuild 5%	Rebuild 10%	Search P50 (ms)	Recall	Total (10%+Search)
SB-SIMD	0	0.0	0.0	0.0	0.535	100%	0.54
FAISS-Flat	0.9	0.01	0.05	0.09	0.11	100%	0.2
HNSW	359	3.59	17.95	35.9	0.06	~99%	35.96
Annoy	58.5	59.09	61.43	64.35	0.01	~95%	64.36
FAISS-IVF	19.3	0.19–0.97	0.97–4.83	1.93–9.65	0.01	~98%	1.94–9.66

SIFT-100K (128D)

Method	Build (ms)	Rebuild 1%	Rebuild 5%	Rebuild 10%	Search P50 (ms)	Recall	Total (10%+Search)
SB-SIMD	0	0.0	0.0	0.0	3.52	100%	3.52
FAISS-Flat	4.3	0.04	0.21	0.43	0.77	100%	1.2
HNSW	8270.2	82.7	413.51	827.02	0.15	~99%	827.17
Annoy	611	617.11	641.55	672.1	0.01	~95%	672.11
FAISS-IVF	249.9	2.50–12.50	12.50–62.48	24.99–124.95	0.09	~98%	25.08–125.04

SIFT-1M (128D)

Method	Build (ms)	Rebuild 1%	Rebuild 5%	Rebuild 10%	Search P50 (ms)	Recall	Total (10%+Search)
SB-SIMD	0	0.0	0.0	0.0	28.539	100%	28.54
FAISS-Flat	99.9	1.0	5.0	9.99	7.72	100%	17.71
HNSW	20742	207.42	1037.1	2074.2	2.89	~99%	2077.09
Annoy	10371	10474.71	10889.55	11408.1	5.78	~95%	11413.88
FAISS-IVF	6223	62.23–311.15	311.15–1555.75	622.30–3111.50	3.61	~98%	625.91–3115.11

GloVe-100K (100D)

Method	Build (ms)	Rebuild 1%	Rebuild 5%	Rebuild 10%	Search P50 (ms)	Recall	Total (10%+Search)
SB-SIMD	0	0.0	0.0	0.0	2.997	100%	3.0
FAISS-Flat	2.7	0.03	0.14	0.27	0.64	100%	0.91
HNSW	7581.4	75.81	379.07	758.14	0.13	~99%	758.27
Annoy	541.9	547.32	569.0	596.09	0.01	~95%	596.1
FAISS-IVF	254.9	2.55–12.75	12.75–63.73	25.49–127.45	0.07	~98%	25.56–127.52

Fashion-MNIST (60K × 784D)

Method	Build (ms)	Rebuild 1%	Rebuild 5%	Rebuild 10%	Search P50 (ms)	Recall	Total (10%+Search)
SB-SIMD	0	0.0	0.0	0.0	3.052	100%	3.05
FAISS-Flat	24.2	0.24	1.21	2.42	2.85	100%	5.27
HNSW	31180	311.8	1559.0	3118.0	0.73	~99%	3118.73
Annoy	2085.5	2106.36	2189.78	2294.05	0.03	~95%	2294.08
FAISS-IVF	374.3	3.74–18.71	18.71–93.58	37.43–187.15	0.38	~98%	37.81–187.53

Bible-31K (768D)

Method	Build (ms)	Rebuild 1%	Rebuild 5%	Rebuild 10%	Search P50 (ms)	Recall	Total (10%+Search)
SB-SIMD	0	0.0	0.0	0.0	1.593	100%	1.59
FAISS-Flat	15	0.15	0.75	1.5	2.5	100%	4.0
HNSW	15000	150.0	750.0	1500.0	0.5	~99%	1500.5
Annoy	1000	1010.0	1050.0	1100.0	0.02	~95%	1100.02
FAISS-IVF	200	2.00–10.00	10.00–50.00	20.00–100.00	0.2	~98%	20.20–100.20

Global Summary

Dataset	Method	Build (ms)	Rebuild 1%	Rebuild 5%	Rebuild 10%	Search P50 (ms)	Recall	Total (10%+Search)
BERT-10K (768D)	SB-SIMD	0	0.0	0.0	0.0	0.756	100%	0.76
BERT-10K (768D)	FAISS-Flat	2.4	0.02	0.12	0.24	0.44	100%	0.68
BERT-10K (768D)	HNSW	2100	21.0	105.0	210.0	0.35	~99%	210.35
BERT-10K (768D)	Annoy	350	353.5	367.5	385.0	0.02	~95%	385.02
BERT-10K (768D)	FAISS-IVF	115	1.15–5.75	5.75–28.75	11.50–57.50	0.08	~98%	11.58–57.58
BERT-100K (768D)	SB-SIMD	0	0.0	0.0	0.0	7.088	100%	7.09
BERT-100K (768D)	FAISS-Flat	39.4	0.39	1.97	3.94	5.05	100%	8.99
BERT-100K (768D)	HNSW	48000	480.0	2400.0	4800.0	0.9	99%	4800.9
BERT-100K (768D)	Annoy	3600	3636.0	3780.0	3960.0	0.05	95%	3960.05
BERT-100K (768D)	FAISS-IVF	1500	15.00–75.00	75.00–375.00	150.00–750.00	0.55	98%	150.55–750.55
BERT-1M (768D)	FAISS-Flat	372.9	3.73	18.64	37.29	51.33	100%	88.62
OpenAI-10K (1536D)	SB-SIMD	0	0.0	0.0	0.0	1.167	100%	1.17
OpenAI-10K (1536D)	FAISS-Flat	4	0.04	0.2	0.4	0.89	100%	1.29
OpenAI-10K (1536D)	HNSW	4200	42.0	210.0	420.0	0.7	99%	420.7
OpenAI-10K (1536D)	Annoy	700	707.0	735.0	770.0	0.04	95%	770.04
OpenAI-10K (1536D)	FAISS-IVF	230	2.30–11.50	11.50–57.50	23.00–115.00	0.16	98%	23.16–115.16
OpenAI-100K (1536D)	SB-SIMD	0	0.0	0.0	0.0	12.24	100%	12.24
OpenAI-100K (1536D)	FAISS-Flat	75.7	0.76	3.79	7.57	9.39	100%	16.96
OpenAI-100K (1536D)	HNSW	96000	960.0	4800.0	9600.0	1.8	99%	9601.8
OpenAI-100K (1536D)	Annoy	7200	7272.0	7560.0	7920.0	0.1	95%	7920.1
OpenAI-100K (1536D)	FAISS-IVF	3000	30.00–150.00	150.00–750.00	300.00–1500.00	1.1	98%	301.10–1501.10
SIFT-10K (128D)	SB-SIMD	0	0.0	0.0	0.0	0.535	100%	0.54
SIFT-10K (128D)	FAISS-Flat	0.9	0.01	0.05	0.09	0.11	100%	0.2
SIFT-10K (128D)	HNSW	359	3.59	17.95	35.9	0.06	~99%	35.96
SIFT-10K (128D)	Annoy	58.5	59.09	61.43	64.35	0.01	~95%	64.36
SIFT-10K (128D)	FAISS-IVF	19.3	0.19–0.97	0.97–4.83	1.93–9.65	0.01	~98%	1.94–9.66
SIFT-100K (128D)	SB-SIMD	0	0.0	0.0	0.0	3.52	100%	3.52
SIFT-100K (128D)	FAISS-Flat	4.3	0.04	0.21	0.43	0.77	100%	1.2
SIFT-100K (128D)	HNSW	8270.2	82.7	413.51	827.02	0.15	~99%	827.17
SIFT-100K (128D)	Annoy	611	617.11	641.55	672.1	0.01	~95%	672.11
SIFT-100K (128D)	FAISS-IVF	249.9	2.50–12.50	12.50–62.48	24.99–124.95	0.09	~98%	25.08–125.04
SIFT-1M (128D)	SB-SIMD	0	0.0	0.0	0.0	28.539	100%	28.54
SIFT-1M (128D)	FAISS-Flat	99.9	1.0	5.0	9.99	7.72	100%	17.71
SIFT-1M (128D)	HNSW	20742	207.42	1037.1	2074.2	2.89	~99%	2077.09
SIFT-1M (128D)	Annoy	10371	10474.71	10889.55	11408.1	5.78	~95%	11413.88
SIFT-1M (128D)	FAISS-IVF	6223	62.23–311.15	311.15–1555.75	622.30–3111.50	3.61	~98%	625.91–3115.11
GloVe-100K (100D)	SB-SIMD	0	0.0	0.0	0.0	2.997	100%	3.0
GloVe-100K (100D)	FAISS-Flat	2.7	0.03	0.14	0.27	0.64	100%	0.91
GloVe-100K (100D)	HNSW	7581.4	75.81	379.07	758.14	0.13	~99%	758.27
GloVe-100K (100D)	Annoy	541.9	547.32	569.0	596.09	0.01	~95%	596.1
GloVe-100K (100D)	FAISS-IVF	254.9	2.55–12.75	12.75–63.73	25.49–127.45	0.07	~98%	25.56–127.52
Fashion-MNIST (60K × 784D)	SB-SIMD	0	0.0	0.0	0.0	3.052	100%	3.05
Fashion-MNIST (60K × 784D)	FAISS-Flat	24.2	0.24	1.21	2.42	2.85	100%	5.27
Fashion-MNIST (60K × 784D)	HNSW	31180	311.8	1559.0	3118.0	0.73	~99%	3118.73
Fashion-MNIST (60K × 784D)	Annoy	2085.5	2106.36	2189.78	2294.05	0.03	~95%	2294.08
Fashion-MNIST (60K × 784D)	FAISS-IVF	374.3	3.74–18.71	18.71–93.58	37.43–187.15	0.38	~98%	37.81–187.53
Bible-31K (768D)	SB-SIMD	0	0.0	0.0	0.0	1.593	100%	1.59
Bible-31K (768D)	FAISS-Flat	15	0.15	0.75	1.5	2.5	100%	4.0
Bible-31K (768D)	HNSW	15000	150.0	750.0	1500.0	0.5	~99%	1500.5
Bible-31K (768D)	Annoy	1000	1010.0	1050.0	1100.0	0.02	~95%	1100.02
Bible-31K (768D)	FAISS-IVF	200	2.00–10.00	10.00–50.00	20.00–100.00	0.2	~98%	20.20–100.20

License

MIT

Credits

SimSIMD for SIMD operations
Clojure community for the amazing language

Name		Name	Last commit message	Last commit date
Latest commit History 2 Commits
classes/lightning_index		classes/lightning_index
data		data
examples		examples
research		research
resources		resources
src		src
.gitignore		.gitignore
README.md		README.md
deps.edn		deps.edn
run.sh		run.sh
simsimd.jar		simsimd.jar

Folders and files

Latest commit

History

Repository files navigation

SB-SIMD Semantic ENN Search Engine

A stateless semantic Exact NN engine for dynamic subsets.

✅ Best fit use cases

⚠️ Less suitable

Performance

Scaling Performance

Rebuild Cost Comparison

Benchmark Results

Installation

Prerequisites

Setup

Usage

Quick Start

As a Library

Data Format

Architecture

Comparison with Other Methods

Examples

Native Libraries

BERT-10K (768D)

BERT-100K (768D)

BERT-1M (768D)

OpenAI-10K (1536D)

OpenAI-100K (1536D)

SIFT-10K (128D)

SIFT-100K (128D)

SIFT-1M (128D)

GloVe-100K (100D)

Fashion-MNIST (60K × 784D)

Bible-31K (768D)

Global Summary

License

Credits

About

Resources

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Uh oh!

Contributors

Uh oh!

Languages

Packages