Store, index, and search high-dimensional vector embeddings — built for RAG systems, semantic search, and LLM applications.

ToucanDB is a lightweight, ML-native vector database written in Python. It transforms unstructured data (text, images, audio) into searchable vector embeddings and retrieves them with sub-millisecond precision — without the overhead of a full server deployment.

• Semantic Search — find by meaning, not keywords, using cosine / dot-product / Euclidean distance
• HNSW & IVF Indexing — fast approximate nearest-neighbour search, auto-tuned
• AES-256-GCM Encryption — all vectors and metadata encrypted at rest
• Rich Metadata Filtering — attach and query arbitrary JSON metadata alongside vectors
• Async Python API — fully async/await, type-safe with Pydantic schemas
• Embedding-model agnostic — works with OpenAI, Sentence Transformers, Cohere, Hugging Face, or any custom model
• Batch operations — bulk insert / search for high-throughput pipelines

Available on PyPI:

pip install toucandb       # pip
poetry add toucandb        # Poetry
uv add toucandb            # uv
pdm add toucandb           # PDM

Optional extras:

pip install toucandb[ml]   # + sentence-transformers, OpenAI, LangChain…
pip install toucandb[gpu]  # + GPU-accelerated FAISS
pip install toucandb[dev]  # development / testing

from sentence_transformers import SentenceTransformer
from toucandb import ToucanDB, VectorSchema, DistanceMetric, IndexType
import asyncio

async def semantic_search_demo():
    # Load a pre-trained sentence transformer model
    model = SentenceTransformer('all-MiniLM-L6-v2')
    
    # Sample documents to embed
    documents = [
        "Python is a versatile programming language used in AI and data science.",
        "Machine learning algorithms can predict patterns from historical data.",
        "Vector databases enable semantic search and similarity matching.",
        "Natural language processing helps computers understand human language.",
        "Deep learning models require large datasets for training.",
    ]
    
    # Generate embeddings
    embeddings = model.encode(documents)
    
    # Initialize ToucanDB
    db = await ToucanDB.create('./semantic_search.tdb', encryption_key='demo-key')
    schema = VectorSchema(
        name='semantic_docs', 
        dimensions=embeddings.shape[1],  # Auto-detect dimensions
        metric=DistanceMetric.COSINE, 
        index_type=IndexType.HNSW
    )
    collection = await db.create_collection(schema)
    
    # Store documents with embeddings
    vectors = []
    for i, (doc, embedding) in enumerate(zip(documents, embeddings)):
        vectors.append({
            'id': f'doc_{i}',
            'vector': embedding.tolist(),
            'metadata': {'text': doc, 'doc_id': i}
        })
    
    await collection.insert_many(vectors)
    
    # Semantic search
    query = "How does AI process language?"
    query_embedding = model.encode([query])[0]
    
    results = await collection.search(query_embedding.tolist(), k=3)
    
    print(f"🔍 Query: '{query}'")
    print("\n📋 Most similar documents:")
    for i, result in enumerate(results, 1):
        print(f"{i}. {result.metadata['text']}")
        print(f"   📊 Similarity: {result.score:.3f}")
        print()

asyncio.run(semantic_search_demo())

More examples in the examples/ directory (RAG pipeline, document search, semantic search).

ToucanDB adapts to any embedding dimension — just set dimensions in your schema.

AWS m5.4xlarge (16 vCPU, 64 GB RAM), 384-dim vectors, HNSW index

Just like the vibrant toucan bird, ToucanDB embodies the perfect combination of precision, adaptability, and intelligence that makes it exceptional for ML applications. Birds are my favourite animals, and toucans are one of my favourites! I've always been inspired by toucans. My grandfather was an ethnologist and explorer in the Amazon rainforest. He also discovered the Jora tribe. My grandparents even spent their honeymoon in the Amazon rainforest and lived in various Latin American countries for quite some time with my grandmother and my mum. His life has deeply inspired me since I was little, and my love for toucans is part of this beautiful legacy. Nature has always played a hugely positive role in my success.

🎯 Precision: Toucans have incredibly precise beaks that can reach exactly where they need to go - just like ToucanDB's vector search that finds exactly the right data points with sub-millisecond accuracy.

🔄 Adaptability: These remarkable birds adapt to diverse environments and data sources - mirroring how ToucanDB seamlessly handles any type of unstructured data (text, images, audio, code).

🧠 Intelligence: Toucans are highly intelligent creatures with excellent memory - reflecting ToucanDB's smart caching, adaptive indexing, and ML-first design that learns and optimizes performance.

🌈 Vibrancy: The toucan's colorful nature represents ToucanDB's rich feature set and the diverse, multimodal data it can process and understand.

Just as toucans navigate complex forest ecosystems with ease, ToucanDB navigates the complex landscape of high-dimensional vector spaces, making ML applications soar! 🚀

Pierre-Henry Soria — a super passionate engineer who loves building cutting-edge AI infrastructure and automating intelligent systems efficiently!

Enthusiast of Machine Learning, Vector Databases, AI, and writing performant code!

Find me at pH7.me

Enjoying this project? Buy me a coffee (spoiler: I love almond extra-hot flat white coffees while coding ML algorithms).

We welcome contributions! Please see our Contributing Guide for details.

ToucanDB is released under the MIT License. See license for further details.

• Distributed clustering support
• Real-time streaming updates
• Multi-modal search (text + image)
• Integration with popular ML frameworks
• Cloud-native deployment options
• GraphQL API support

Built with ❤️ for the AI community