NEURONDB/AI Database Extension

NeuronDB

PostgreSQL AI Extension for Vector Search, ML Inference and RAG Pipeline. AI applications in PostgreSQL with GPU acceleration, 52 ML algorithms, and hybrid search.

StartRead docsGitHub
psql
Quickstart
CREATE EXTENSION neurondb;
PostgreSQL 16 to 186 Vector Types52 ML Algorithms665+ SQL Functions
NeuronDB
AI Database Extension
NeuronDB: PostgreSQL AI Extension

Architecture

Architecture with vector search, ML inference, and RAG pipeline

NeuronDB Architecture

PostgreSQL 16-18

ACID | MVCC | WAL | Replication

Vector Engine

  • • HNSW & IVF Indexing
  • • 10+ Distance Metrics
  • • Quantization
  • • SIMD Optimized

ML Engine

  • • 52 ML Algorithms
  • • ONNX Runtime
  • • Batch Processing
  • • Pure C Implementation

Embedding Engine

  • • Text Embeddings
  • • Multimodal Support
  • • Hugging Face Integration
  • • Caching & Batching

GPU Accelerator

  • • CUDA (NVIDIA)
  • • ROCm (AMD)
  • • Metal (Apple)
  • • Auto Detection

Advanced Features

Hybrid Search
Vector + FTS
Reranking
Cross-encoder, LLM
RAG Pipeline
Complete In-DB
Background Workers
4 Workers

SQL API Layer

665+ SQL Functions | Operators | Types | Views

NeuronDB Capabilities

PostgreSQL AI Extension for Vector Search, ML Inference and RAG Pipeline. AI applications in PostgreSQL with GPU acceleration, 52 ML algorithms, and hybrid search.

NeuronDB
console • demo
v1.0ready
HNSW + IVF indexing
High-performance vector search with multiple distance metrics
SQL
Vector Searchdemo
-- Create vector table CREATE TABLE embeddings ( id SERIAL PRIMARY KEY, embedding vector(384), metadata JSONB ); -- Create HNSW index CREATE INDEX ON embeddings USING hnsw (embedding vector_cosine_ops); -- Vector search SELECT id, 1 - (embedding <=> query_vec) as similarity FROM embeddings ORDER BY embedding <=> query_vec LIMIT 10;
Results5 rows
idsimilaritytext
420.9523Vector search enables semantic similarity matching in high-dimensional spaces using embeddings to find related content…
380.9234HNSW indexes provide fast approximate nearest neighbor search for vector databases with logarithmic query time…
350.8945RAG combines retrieval with generation for accurate LLM responses by finding relevant context first…
310.8656Embeddings convert text into numerical vectors for machine learning models to process semantic meaning…
280.8367PostgreSQL extensions enable vector operations directly in the database without external services…
Performance
Query Time
8.42ms
Latency (P95)
12.5ms
QPS
8.2k
Status
ready
Query Statistics
Execution
Rows Returned5
Cache Hit96%
PlanOptimized
Connection
Databaseneurondb
Versionv1.0
Statusactive
Summary
Total Queries1,247
Success Rate99.8%
Avg Latency8.2ms
Vector Engine Docs
AI Database Features

Why NeuronDB

Vector Search & Indexing

HNSW Index10+ Metrics
  • •5 vector types: vector (float32, up to 16K dims), vectorp (Product Quantization), vecmap (sparse vectors), vgraph (graph-based), rtext (retrieval text)
  • •HNSW and IVF indexing with automatic tuning
  • •10+ distance metrics: L2, Cosine, Inner Product, Manhattan, Hamming, Jaccard, and more
  • •Product Quantization (PQ) and Optimized PQ (OPQ) for 2x-32x compression
  • •DiskANN support for billion-scale vectors on SSD

ML & Embeddings

RFXGBSVMEmbeddings52 Algorithms
  • •52 ML algorithms in pure C: Random Forest, XGBoost, LightGBM, CatBoost, SVM, KNN, Neural Networks, and more
  • •Built-in embedding generation with intelligent caching
  • •ONNX runtime integration for model inference
  • •Batch processing with GPU acceleration
  • •Model catalog with versioning and A/B testing

Hybrid Search & Retrieval

VectorTextHybrid70% + 30%
  • •Combine vector similarity with full-text search (BM25)
  • •Configurable weighted scoring (e.g., 70% vector + 30% text)
  • •Multi-vector document support
  • •Faceted search with category filters
  • •Temporal decay for time-sensitive relevance ranking

Reranking

ResultsBeforeRerankTop KAfter<10ms latency
  • •Cross-encoder neural reranking for precision improvement
  • •LLM-powered scoring (GPT-4, Claude integration)
  • •ColBERT late interaction models
  • •MMR (Maximal Marginal Relevance) for diversity
  • •Ensemble strategies combining multiple rankers
  • •Sub-10ms latency for production workloads

RAG Pipeline

DocsRetrieveGenerateSQL APIIn PostgreSQL
  • •Complete Retrieval Augmented Generation in PostgreSQL
  • •Intelligent document chunking and processing
  • •Semantic retrieval with automatic reranking
  • •LLM integration for answer generation
  • •Context management and guardrails for content safety
  • •RAG operations available directly in SQL

Background Workers

QMDL4 Workers
  • •neuranq: Async job queue executor with SKIP LOCKED, retries, and batch processing
  • •neuranmon: Live query auto-tuner for search params and cache optimization
  • •neurandefrag: Automatic index maintenance, compaction, and rebuild scheduling
  • •neuranllm: LLM job processing with crash recovery
  • •All workers are tenant-aware with QPS and cost budgets

ML Analytics Suite

ClusterPCAQualityDriftTopicsAnalytics
  • •19 clustering algorithms: K-means, DBSCAN, GMM, Hierarchical (all GPU-accelerated)
  • •Dimensionality reduction: PCA, PCA Whitening, OPQ
  • •Outlier detection: Z-score, Modified Z-score, IQR
  • •Quality metrics: Recall@K, Precision@K, F1@K, MRR, Silhouette Score
  • •Drift detection: Centroid drift, Distribution divergence, Temporal monitoring
  • •Analytics: Topic discovery, Similarity histograms, KNN graph building

GPU Acceleration

CUDAROCmMetalGPU100x Speedup
  • •Multi-platform support: CUDA (NVIDIA), ROCm (AMD), Metal (Apple Silicon)
  • •GPU-accelerated ML algorithms: Random Forest, XGBoost, LightGBM, SVM, KNN, and more
  • •Batch distance computation with 100x speedup
  • •Automatic GPU detection with intelligent CPU fallback
  • •Multi-stream compute overlap for maximum throughput
  • •Efficient memory management and allocation

Performance & Optimization

SIMDCacheWAL10.1x Faster<1ms Search
  • •HNSW index building: 606ms for 50K vectors (128-dim), 10.1x faster than pgvector
  • •SIMD-optimized distance calculations (AVX2, AVX-512, NEON)
  • •In-memory graph building using maintenance_work_mem for optimal index construction
  • •Efficient neighbor finding during insert (not after flush) for faster builds
  • •Squared distance optimization avoiding sqrt() overhead in comparisons
  • •Intelligent query planning with accurate cost estimates
  • •ANN buffer cache for hot centroids and frequent queries
  • •WAL compression with delta encoding
  • •Parallel kNN execution across multiple cores
  • •Predictive prefetching for reduced latency
  • •Sub-millisecond searches on millions of vectors

Security

🔒EncryptionRLS
  • •Vector encryption using AES-GCM via OpenSSL
  • •Differential privacy for embedding protection
  • •Row-level security (RLS) integration
  • •Multi-tenant isolation with resource quotas
  • •HMAC-SHA256 signed results for integrity verification
  • •Audit logging with tamper detection
  • •GDPR-compliant data handling and governance

Monitoring & Observability

MetricsLoggingPrometheusObservability
  • •pg_stat_neurondb view with real-time performance metrics
  • •Worker heartbeats and watchdog monitoring
  • •Query latency histograms and percentile tracking
  • •Cache hit rate tracking and optimization insights
  • •Recall@K monitoring for search quality
  • •Model cost accounting and usage analytics
  • •Prometheus exporter ready for integration
  • •Structured JSON logging with neurondb: prefix

PostgreSQL Native Architecture

PostgreSQLNeuronDB665+ Functions
  • •Pure C implementation following 100% PostgreSQL coding standards
  • •665+ SQL functions, types, and operators
  • •7 new monitoring views for comprehensive observability
  • •Shared memory for efficient caching
  • •WAL integration for durability and crash recovery
  • •SPI for safe operations and transaction handling
  • •Background worker framework integration
  • •Standard extension with zero external dependencies
  • •SIMD-optimized (AVX2, AVX-512, NEON) with runtime CPU detection

NeuronDB vs. Alternatives

How NeuronDB compares to other PostgreSQL AI and vector extensions.

FeatureNeuronDBpgvectorpgvectorscalepgaiPostgresML
Index typesHNSW, IVF, PQ, hybrid, multi-vectorHNSW + IVFStreamingDiskANNUses pgvectorpgvector-based
ML InferenceONNX (C++)NoneNoneAPI callsPython ML libs
Embedding GenerationIn-database (ONNX)ExternalExternalExternal APIIn-database (Transformers)
Hybrid SearchNative (Vector+FTS)ManualManualManualManual
RerankingCross-encoder, LLM, ColBERT, MMRNoneNoneNoneNone
ML Algorithms52 algorithms: RF, XGBoost, LightGBM, CatBoost, SVM, KNN, DT, NB, NN, K-means, DBSCAN, GMM, PCA, etc.NoneNoneNoneXGBoost, LightGBM, sklearn suite, Linear/Logistic
Background Workers4 workers: neuranq, neuranmon, neurandefrag, neuranllmNoneNoneNoneNone
RAG PipelineComplete In-DBNoneNonePartial (API)Partial (Python)
QuantizationFP16, INT8, Binary (2x-32x)Binary onlyBinary onlyNoneNone
ImplementationPure CPure CPure CRust + SQLPython + C
GPU SupportCUDA + ROCm + Metal (native C/C++)NoneNoneNoneCUDA (via Python)
PostgreSQL Versions16, 17, 1812-1815-1816-1814-16
Vector Types6 types: vector, vectorp, vecmap, vgraph, rtext, sparse_vector1 type: vector1 type: vectorUses pgvectorUses pgvector
Distance Metrics10+ metrics: L2, Cosine, Inner Product, Manhattan, Hamming, Jaccard, etc.3 metrics: L2, Cosine, Inner Product3 metrics: L2, Cosine, Inner ProductUses pgvectorUses pgvector
SQL Functions665+ functions~20 functions~30 functions~15 functions~50 functions
Benchmark coverageRAGAS, MTEB, BEIR integratedManual setup requiredManual setup requiredManual setup requiredManual setup required
HNSW Index Build Performance606ms (50K, 128-dim L2, 10.1x), 583ms (50K, 128-dim Cosine, 8.8x), 146ms (10K, 768-dim L2, 27.1x), 1,208ms (100K, 128-dim L2, 13.0x)6,108ms (50K, 128-dim L2, baseline), 5,113ms (50K, 128-dim Cosine), 3,960ms (10K, 768-dim L2), 15,696ms (100K, 128-dim L2)Varies by configurationUses pgvectorUses pgvector
Performance (QPS)100K+ (with GPU)10K-50K50K-100KLimited (API overhead)5K-20K (Python overhead)
DependenciesZero (pure C, optional ONNX)Zero (pure C)Zero (pure C)Rust runtimePython + ML libraries

FAQ

Frequently asked questions

Common questions about NeuronDB and deployment.

NeuronDB supports PostgreSQL 16, 17, and 18. We recommend the latest stable release for best performance and security.
NeuronDB adds 665+ SQL functions, 52 ML algorithms, in-database RAG pipeline, 6 vector types, 6+ index types (HNSW, IVF, PQ, OPQ, hybrid), cross-encoder and LLM reranking, 4 background workers, and optional GPU acceleration (CUDA, ROCm, Metal). pgvector provides core vector storage and HNSW/IVF indexes only.
Yes. NeuronDB supports NVIDIA CUDA, AMD ROCm, and Apple Metal for batch distance computation, ML inference, and index build. Enable via GUC and optional GPU libraries.
The core NeuronDB PostgreSQL extension is open source. NeuronDB Cloud and NeuronDB Hub are managed/platform offerings; contact us for access.
NeuronDB supports vector encryption (AES-GCM), RLS, audit logging, and GDPR-aware data handling. See the Security docs for RLS, field encryption, and audit logging.
Start

Add AI Capabilities to PostgreSQL

Install NeuronDB. Build semantic search, RAG applications, and ML features in your PostgreSQL infrastructure.

View DocumentationView on GitHub