A high-performance multi-model database with ACID guarantees + Native AI/LLM Integration

"ThemisDB keeps its own llamas." – Run LLaMA, Mistral, Phi-3 directly in your database, no API calls needed.

"ThemisDB keeps its own llamas." – Run AI/LLM workloads directly in your database - no external API costs!

• 🧠 Embedded LLM Engine - llama.cpp integrated, run LLaMA/Mistral/Phi-3 (1B-70B params) on GPU
• ⚡ Zero-Copy RAG - Direct memory access between vector DB and LLM (4x faster, 0ms transfer)
• 💰 100-1000x Cost Reduction - vs. AWS/Azure/GCP APIs (€0.02 vs. €30 per 1M tokens)
• 🎯 All GPU Tiers Supported - Entry (<16GB), Mid-Range (<24GB), High-End (>24GB)
• 🔄 Distributed Reasoning - Brain-inspired multi-shard collaboration (3.6x faster complex tasks)
• 📊 Continuous Batching - vLLM-style optimization (2.6x throughput)
• 🐳 Docker/VM Ready - Full testing possible without GPU (CPU fallback mode)

→ See GPU-Tier Analysis & Hyperscaler Comparison
→ See Native LLM Integration Concept
→ See Complete Documentation

ThemisDB is a production-ready multi-model database that combines relational, graph, vector, and document models in a single system with full ACID transaction support. Built on RocksDB with advanced security and compliance features.

Key Features:

• 🔒 ACID Transactions - Full snapshot isolation with MVCC
• 🔍 Multi-Model - Relational, Graph, Vector, Document in one database
• 🚀 High Performance - 45K writes/s, 120K reads/s, GPU-accelerated vector search
• 🛡️ Enterprise Security - TLS 1.3, RBAC, field-level encryption, audit logging
• 📊 Advanced Analytics - Complex Event Processing (CEP), OLAP, Time-series
• 🌐 Distributed - Horizontal sharding, replication, Kubernetes-ready
• 🧠 AI-Ready - Hybrid search (RAG), embedding cache, FAISS integration, native LLM engine (v1.5.0)

# Pull and run the latest version
docker pull themisdb/themisdb:latest
docker run -d -p 8080:8080 -p 18765:18765 -v themis_data:/data themisdb/themisdb:latest

# Or use Docker Compose
docker compose up -d

# Clone repository
git clone https://github.com/makr-code/ThemisDB.git
cd ThemisDB

# Setup and build (Linux/macOS)
./setup.sh
./build.sh

# Setup and build (Windows)
.\setup.ps1
.\build.ps1

# Start server
./build/themis_server --config config.yaml

Linux (Debian/Ubuntu):

wget https://github.com/makr-code/ThemisDB/releases/latest/download/themisdb_1.2.0-1_amd64.deb
sudo apt install ./themisdb_1.2.0-1_amd64.deb
sudo systemctl start themisdb

macOS (Homebrew):

brew install themisdb
brew services start themisdb

Windows (Chocolatey):

choco install themisdb

# 1. Check server health
curl http://localhost:8765/health

# 2. Create an entity
curl -X PUT http://localhost:8765/entities/users:alice \
  -H "Content-Type: application/json" \
  -d '{"blob":"{\"name\":\"Alice\",\"age\":30,\"city\":\"Berlin\"}"}'

# 3. Create an index
curl -X POST http://localhost:8765/index/create \
  -H "Content-Type: application/json" \
  -d '{"table":"users","column":"city"}'

# 4. Query by index
curl -X POST http://localhost:8765/query \
  -H "Content-Type: application/json" \
  -d '{"table":"users","predicates":[{"column":"city","value":"Berlin"}],"return":"entities"}'

# 5. View metrics
curl http://localhost:8765/metrics

ThemisDB uses a unified storage architecture with specialized projection layers:

┌─────────────────────────────────────────────────────────┐
│                   Query Layer (AQL)                     │
│  SQL-like • Graph Traversals • Vector Search • Analytics│
├─────────────────────────────────────────────────────────┤
│                 Projection Layers                        │
│  Secondary Indices • Graph Adjacency • HNSW Vector      │
├─────────────────────────────────────────────────────────┤
│              Canonical Storage (Base Entity)             │
│         RocksDB LSM-Tree • MVCC Transactions            │
└─────────────────────────────────────────────────────────┘

Core Components:

• Storage Engine: RocksDB TransactionDB with LSM-Tree
• Transaction Manager: MVCC with snapshot isolation
• Query Engine: Advanced Query Language (AQL) with graph/vector support
• Index Manager: Automatic maintenance of secondary, graph, and vector indexes
• Security: TLS 1.3, RBAC, field encryption, audit logging
• Observability: Prometheus metrics, OpenTelemetry tracing

→ Full Architecture Documentation

Enterprise Features Release (December 2025)

• ✅ Hypertables - TimescaleDB-compatible time-series with automatic partitioning
• ✅ Hybrid Search - RAG-optimized search combining BM25 + vector similarity (85% recall@10)
• ✅ FAISS Advanced - IVF+PQ vector search with 10-100x memory reduction
• ✅ Embedding Cache - 70-90% cost reduction for LLM applications
• ✅ Time-Series Aggregates - SIMD-accelerated analytics (5-10x faster)

→ Full Changelog | → Release Notes

• Relational: SQL-like queries with secondary indexes
• Graph: BFS, Dijkstra, A* traversals with path constraints
• Vector: HNSW and FAISS for similarity search (GPU-accelerated)
• Document: JSON storage with flexible schema
• Time-Series: Gorilla compression, continuous aggregates

• Full ACID guarantees with snapshot isolation
• Write-write conflict detection
• Atomic updates across all index types
• Session-based and direct API

• CEP Engine: Complex Event Processing with pattern matching
• OLAP: CUBE, ROLLUP, window functions
• Time-Series: Compression, retention policies, aggregates
• Hybrid Search: BM25 + vector for RAG workflows

• TLS 1.3 with mTLS support
• Role-Based Access Control (RBAC)
• Field-level encryption
• Audit logging with SIEM integration
• Certificate pinning for HSM/TSA
• Secrets management (HashiCorp Vault)

• Horizontal sharding with consistent hashing
• Leader-follower and multi-master replication
• RAID-like redundancy (MIRROR, STRIPE, PARITY)
• Kubernetes operator with CRDs
• Auto-rebalancing and cloud deployment

• 10 backend options: CUDA, Vulkan, HIP, OpenCL, DirectX, OneAPI, ZLUDA
• 10-50x speedup for vector search
• Automatic platform detection and fallback

Getting Started:

• Installation Guide
• Docker Deployment
• Quick Start Tutorial

Core Concepts:

• Architecture Overview
• Multi-Model Design
• Transaction Management
• AQL Query Language

Features:

• Vector Search
• Graph Operations
• Time-Series Engine
• Security & Compliance
• Feature Overview

Operations:

• Configuration Guide
• Monitoring & Metrics
• Backup & Recovery
• Performance Tuning

Development:

• Build Guide
• Contributing
• API Reference
• Client SDKs

Full Documentation: https://makr-code.github.io/ThemisDB/

Completed (v1.0 - v1.2):

• ✅ ACID transactions with MVCC
• ✅ Multi-model support (relational, graph, vector, document)
• ✅ Horizontal sharding and replication
• ✅ GPU acceleration (10 backends)
• ✅ Enterprise security features
• ✅ Client SDKs (7 languages)
• ✅ Kubernetes operator

In Progress (v1.3 - Q1 2026):

• 🚧 Query optimizer enhancements
• 🚧 Multi-datacenter deployment
• 🚧 Advanced ML/GNN features
• 🚧 Production hardening

Planned (v1.4+ - 2026):

• 📋 Real-time materialized views
• 📋 Cross-region replication
• 📋 Advanced security compliance (SOC 2, HIPAA)
• 📋 Cloud-native optimizations

→ Detailed Roadmap

Benchmark Results (Release build, i7-12700K):

→ Detailed Benchmarks

• Documentation: https://makr-code.github.io/ThemisDB/
• GitHub Issues: Report bugs or request features
• Discussions: Community discussions
• Contributing: Contributing guidelines
• Security: Security policy

ThemisDB is released under the MIT License.

ThemisDB is inspired by and builds upon the ideas from:

• ArangoDB - Multi-model architecture
• CozoDB - Hybrid relational-graph-vector design
• Azure Cosmos DB - Multi-model with unified API
• RocksDB - High-performance LSM-Tree storage
• FAISS - Efficient similarity search

For a complete list of third-party libraries and feature attributions, see ATTRIBUTIONS.md.

Built with ❤️ for the database community