NeuronDB

NeuronDB is a PostgreSQL extension that adds vector search, machine learning, and AI capabilities directly to PostgreSQL.

What it is

• A PostgreSQL extension that defines types (for example vector), operators, and index access methods
• ~650+ SQL functions for vector operations, ML, embeddings, and RAG
• 25+ machine learning algorithm families
• GPU acceleration for CUDA, ROCm, and Metal platforms
• Background workers for async operations, auto-tuning, and maintenance

Key Modules & Features

Vector Operations

• Vector Types: vector, vectorp, vecmap, vgraph, rtext, sparse_vector for dense and sparse vectors
• Distance Metrics: L2, cosine similarity, inner product, and more
• Indexing: HNSW and IVF indexes for fast similarity search
• Quantization: Product Quantization (PQ) and Optimized Product Quantization (OPQ)

Machine Learning (25+ algorithm families)

• Classification: Random Forest, SVM, Logistic Regression, Naive Bayes, Decision Trees
• Regression: Linear, Ridge, Lasso, Neural Networks, Deep Learning
• Clustering: K-Means, Mini-batch K-Means, DBSCAN, Gaussian Mixture Model, Hierarchical
• Gradient Boosting: XGBoost, LightGBM, CatBoost
• Dimensionality Reduction: PCA, PCA Whitening
• Outlier Detection: Z-score, Modified Z-score, IQR
• Time Series: Forecasting and analysis
• Recommendation Systems: Collaborative filtering
• Quality Metrics: Recall@K, Precision@K, F1@K, MRR, Davies-Bouldin Index

Embeddings & LLM Integration

• Embedding Generation: Text, image, and multimodal embeddings
• ONNX Runtime: Model inference and management
• Hugging Face Integration: Direct model loading and inference
• OpenAI Integration: API-based embeddings and completions
• Model Management: Model catalog, versioning, and deployment

Hybrid Search & Retrieval

• Hybrid Search: Combine vector and full-text search
• Temporal Search: Time-decay relevance scoring
• Sparse Search: Sparse vector operations
• Multi-Vector: Multiple embeddings per document

Reranking

• Cross-Encoder: Neural reranking models
• LLM Reranking: GPT/Claude-powered scoring
• Ensemble Reranking: Combine multiple strategies
• Learning to Rank (LTR): Trainable reranking models

RAG Pipeline

• Document Processing: Text extraction and chunking
• Context Retrieval: Semantic search for context
• LLM Integration: Complete RAG workflows

Background Workers

• neuranq: Async job queue executor
• neuranmon: Live query auto-tuner for index optimization
• neurandefrag: Index maintenance and defragmentation
• neuranllm: LLM job processor for embeddings and completions

GPU Acceleration

• CUDA: NVIDIA GPU acceleration
• ROCm: AMD GPU acceleration
• Metal: Apple Silicon GPU acceleration
• Auto-Detection: Automatic GPU backend selection

Multi-Tenancy & Security

• Tenant Management: Per-tenant resource quotas and isolation
• Row-Level Security: RLS policies for vector data
• Encryption: Post-quantum encryption support
• Access Control: Fine-grained permissions

Observability

• Monitoring Views: Vector stats, index health, query performance
• Prometheus Integration: Metrics export
• Performance Tracking: Query performance metrics

Documentation

• Main README: README.md (repository root)
• Installation: INSTALL.md (repository root)
• Complete Docs: docs/
• SQL API: Functions defined in sql/neurondb--*.sql (repository root)
• Official Docs: https://www.neurondb.ai/docs

Docker

• Compose file: docker/docker-compose.yml. Services: neurondb (CPU) and GPU profiles (cuda, rocm, metal).
• See docker/README.md.

Quick Start

After Postgres is running:

Copy
CREATE EXTENSION IF NOT EXISTS neurondb;
SELECT neurondb.version();

For detailed setup, see INSTALL.md or Installation.