NeuronDB Documentation
GPU-accelerated vector search, model inference, hybrid retrieval, and RAG orchestration built into PostgreSQL. NeurondB is an AI PostgreSQL extension. Use this documentation to deploy NeurondB, operate background workers, and embed ML pipelines in SQL.
Key Capabilities
Vector Search
HNSW, IVF, product quantization, and custom distance metrics for billion-scale similarity search.
ML Inference
ONNX runtime integration, GPU offload, and batch execution for deep learning workloads in SQL.
Hybrid Retrieval
Blend keyword, metadata, and vector signals to deliver highly relevant multimodal results.
RAG Pipelines
In-database retrieval augmented generation with prompt templates, metadata policies, and observability.
Documentation Library
Getting Started
Install NeurondB on PostgreSQL 16–18, verify GPU support, and apply baseline configuration.
- Installation
Build from source or install packages.
- Quick Start
Load sample data and run first vector searches.
- Configuration
GUC parameters for CPU/GPU execution paths.
Core Features
Learn how NeurondB models vectors, maintains indexes, and tunes recall versus latency.
- Vector Types
Supported dimensionality and storage formats.
- Distance Metrics
Cosine, L2, IP, dot, and hybrid scoring.
- Indexing
HNSW, IVF, PQ, and adaptive index selection.
- Quantization
Reduce memory footprint with scalar and vector quantization.
ML & Embeddings
Generate, store, and serve embeddings alongside model lifecycle management.
- Embeddings
Transform text, audio, and images into dense vectors.
- Inference
Deploy ONNX models with GPU batching and caching.
- Model Management
Version control, approvals, and rollback workflows.
Hybrid Search & Reranking
Combine text search, BM25, and neural rerankers for production retrieval pipelines.
- Hybrid Overview
Architectures for multi-signal retrieval.
- Reranking
Cross-encoder and LLM reranking playbooks.
- RAG Workflows
Orchestrate retrieval augmented generation end to end.
Background Workers
Operational guidance for queue execution, auto-tuning, and index maintenance workers.
- Worker Overview
Understand worker architecture and lifecycles.
- Queue Worker (neuranq)
Batch ingestion and asynchronous scoring.
- Auto-Tuner (neuranmon)
Automated index health and GPU utilization tuning.
- Index Maintenance (neurandefrag)
Defragment and rebalance vector indexes online.
Components
NeuronDB ecosystem components: NeuronAgent for agent runtime and NeuronMCP for MCP protocol support.
- NeuronAgent
AI agent runtime with REST API, WebSocket, long-term memory, and tool execution.
- NeuronMCP
Model Context Protocol server for MCP-compatible clients like Claude Desktop.