An AI-native embedded database
Zero config. Single file. Built for ML workloads.
pip install synadb
Built for AI workflows
Everything you need for ML development in a single embedded database.
Vector Search
Native HNSW index for similarity search. Optional FAISS backend for billion-scale deployments.
MmapVectorStore
Ultra-high-throughput vector ingestion at 490K vectors/sec. 7x faster than standard VectorStore.
Gravity Well Index
Novel O(N) build time index. 168x faster than HNSW for streaming and real-time data.
Cascade Index
Three-stage hybrid index combining LSH, bucket trees, and sparse graphs. Tunable recall up to 99%.
Experiment Tracking
Log parameters, metrics, and artifacts locally. No external server required.
Tensor Storage
Store and retrieve tensors directly to NumPy and PyTorch with zero serialization overhead.
Model Registry
Version models with SHA-256 checksums. Promote through development, staging, and production stages.
LLM Integrations
Native support for LangChain, LlamaIndex, and Haystack. Drop-in vector stores and chat history.
ML Integrations
PyTorch Dataset/DataLoader and TensorFlow tf.data integration. Distributed training support.
Studio Web UI
Visual database explorer with 3D embedding clusters, model registry browser, and statistics dashboard.
CLI Tool
Command-line interface for database inspection, management, and scripting workflows.
GPU Direct
CUDA tensor loading with pinned memory and async streams. Zero-copy GPU memory access.
Single File
Your entire database in one portable file. No server process, no configuration files.
High Performance
Written in Rust. 100K+ writes per second. Sub-millisecond reads with in-memory indexing.
Cross-Platform
Native binaries for Linux, macOS (Intel & Apple Silicon), and Windows. C-ABI for any FFI language.
Get started
pip install synadbcargo add synadbfrom synadb import SynaDB, VectorStore
# Key-value storage
db = SynaDB("project.db")
db.put_float("accuracy", 0.95)
db.put_text("model_name", "bert-base")
# Vector similarity search
vectors = VectorStore("embeddings.db", dimensions=768)
vectors.insert("doc_1", embedding)
results = vectors.search(query, k=10)Three paradigms, one database
SynaDB synthesizes the best ideas from three database traditions.
SQLite
Single file, zero config, embedded library. No server to manage.
DuckDB
Columnar history, efficient tensor extraction, analytical queries.
MongoDB
Schema-free storage, flexible data types, no migrations needed.