caliby 0.1.3

English | 中文

_{⭐ If you find Caliby useful, please consider giving it a star!}

---

Caliby is an embeddable vector database designed for AI applications that need to scale beyond available memory without the complexity of distributed systems. Unlike client-server vector databases that require separate infrastructure, Caliby runs directly inside your application for embedded use cases with rich features including document storage, vector search, and filtered search.

🎯 Why Caliby?

The Problem with Existing Solutions

Solution	Limitation
HNSWLib / Faiss / Usearch	Memory-only — crash or slow down when data exceeds RAM
Pinecone / Weaviate / Qdrant	Requires separate server infrastructure, network latency, operational overhead
ChromaDB / LanceDB	Limited indexing options, no true buffer pool for efficient larger-than-memory

Caliby's Approach: Embeddable + Fast + Larger-Than-Memory

Caliby combines the simplicity of an embedded library with the scalability of disk-based storage while maintaining memory-fast vector search when data fits in memory:

• 🔌 Zero Infrastructure: pip install caliby — no Docker, no servers, no configuration
• 📦 Ship with Your App: Bundle Caliby directly into desktop apps, edge devices, or microservices
• 💾 1B+ Vectors on a Laptop: Handle datasets far larger than RAM with intelligent buffer management
• ⚡ In-Memory Performance: When data fits in RAM, matches or exceeds HNSWLib/Faiss speed
• 📉 Graceful Degradation: As data grows beyond RAM, performance degrades smoothly — not catastrophically

Perfect For

• 🤖 AI Agents — Persistent memory that survives restarts, scales with conversation history
• 📱 Desktop/Mobile Apps — Local-first semantic search without cloud dependencies
• 🔧 Developer Tools — Embed code search, documentation retrieval in IDEs and CLIs
• 🏭 Edge Computing — Run on resource-constrained devices without network access
• 🧪 Rapid Prototyping — Go from idea to working RAG pipeline in minutes, not hours

✨ Key Features

• 🔌 Embeddable: Single-process library, runs in your application's memory space
• 💾 Larger-Than-Memory: Innovative buffer pool handles datasets 10-100x larger than RAM
• 📚 Document Storage: Store vectors, text, and metadata with flexible schemas
• 🔍 Filtered Search: Efficient vector search with metadata filtering
• 🔗 Hybrid Search: Combine vector similarity and BM25 full-text search
• 🔥 In-Memory Speed: Matches HNSWLib/Faiss when data fits in RAM
• 🎯 Multiple Index Types: HNSW, DiskANN, IVF+PQ, B+tree, and Inverted Index

📱 Use Cases

Caliby excels where other vector databases struggle — embedded scenarios with large datasets:

Use Case	Why Caliby?	Example
🤖 Agentic Memory Store	Persistent agent memory that grows unbounded, survives restarts, no external DB needed	agentic_memory_store.py
📚 RAG Pipeline	Index millions of document chunks locally, hybrid search without API latency	rag_pipeline.py
🛒 Recommendation System	Ship recommendations with your app, works offline on edge devices	recommendation_system.py
🔍 Semantic Search	Local-first search for desktop apps, developer tools, and offline-capable systems	semantic_search.py
🖼️ Image Similarity	Visual search embedded in photo apps, no cloud upload required	image_similarity_search.py

🚀 Quick Start

Installation

From PyPI (Recommended):

pip install caliby

From Source:

git clone --recursive https://github.com/zxjcarrot/caliby.git
cd caliby
pip install -e .

Note: The --recursive flag is required to initialize the pybind11 submodule. If you already cloned without it, run:

git submodule update --init --recursive

Collection API (Recommended)

The Collection API provides a high-level interface for storing documents with vectors, text, and metadata:

import caliby
import numpy as np

# Initialize
caliby.set_buffer_config(size_gb=1.0)
caliby.open('/tmp/my_database', cleanup_if_exist=True)

# Define schema
schema = caliby.Schema()
schema.add_field("category", caliby.FieldType.STRING)
schema.add_field("price", caliby.FieldType.FLOAT)

# Create a collection with 128-dimensional vectors
collection = caliby.Collection("products", schema, vector_dim=128)

# Add documents (returns assigned IDs)
contents = ["Wireless headphones with noise cancellation",
            "Running shoes for trail running"]
metadatas = [{"category": "electronics", "price": 99.99},
             {"category": "sports", "price": 79.99}]
vectors = np.random.rand(2, 128).astype(np.float32).tolist()

doc_ids = collection.add(contents, metadatas, vectors)

# Create indices
collection.create_hnsw_index("vec_idx", M=16, ef_construction=200)
collection.create_text_index("text_idx")
collection.create_metadata_index("category_idx", ["category"])

# Vector search with metadata filter
query = np.random.rand(128).astype(np.float32)
results = collection.search_vector(query, "vec_idx", k=10,
                                   filter='{"category": {"$eq": "electronics"}}')

# Hybrid search (vector + text)
results = collection.search_hybrid(query, "vec_idx",
                                   "wireless headphones", "text_idx", k=10)

for r in results:
    print(f"Doc {r.doc_id}: score={r.score:.4f}")

caliby.close()

📖 See docs/COLLECTION_API.md for complete documentation including advanced filtering, best practices, and performance tuning.

Low-Level Index API

For direct control over indices:

import caliby
import numpy as np

# Initialize the system and configure buffer pool
caliby.set_buffer_config(size_gb=1.0)  # Set buffer pool size
caliby.open('/tmp/caliby_data')  # Initialize catalog

# Create an HNSW index
index = caliby.HnswIndex(
    max_elements=1_000_000,     # Maximum number of vectors
    dim=128,                    # Vector dimension
    M=16,                       # HNSW parameter (connections per node)
    ef_construction=200,        # Construction-time search depth
    enable_prefetch=True,       # Enable prefetching for performance
    skip_recovery=False,        # Whether to skip recovery from disk
    index_id=0,                 # Unique index identifier for multi-index
    name='user_embeddings',     # Optional human-readable name
)

# Add vectors (batch)
vectors = np.random.rand(10000, 128).astype(np.float32)
index.add_points(vectors, num_threads=4)  # Parallel insertion

# Get index info
print(f"Index name: {index.get_name()}")  # Output: 'user_embeddings'
print(f"Dimension: {index.get_dim()}")

# Search (single query)
query = np.random.rand(128).astype(np.float32)
labels, distances = index.search_knn(query, k=10, ef_search_param=50)

# Batch search (parallel)
queries = np.random.rand(100, 128).astype(np.float32)
results = index.search_knn_parallel(queries, k=10, ef_search_param=50, num_threads=4)

# Close when done
caliby.close()

🏗️ Index Types

HNSW (Hierarchical Navigable Small World)

Best for: High recall requirements, moderate to large dataset sizes

import caliby
import numpy as np

# Initialize system
caliby.set_buffer_config(size_gb=2.0)
caliby.open('/tmp/caliby_data')

index = caliby.HnswIndex(
    max_elements=1_000_000,
    dim=128,
    M=16,                    # Higher = better recall, more memory
    ef_construction=200,     # Higher = better graph quality, slower build
    enable_prefetch=True,    # Enable prefetching
    skip_recovery=False,
    index_id=0,              # Unique ID for multi-index support
    name='my_vectors',       # Optional human-readable name
)

# Add points
vectors = np.random.rand(100000, 128).astype(np.float32)
index.add_points(vectors, num_threads=4)

# Search with ef_search_param
query = np.random.rand(128).astype(np.float32)
labels, distances = index.search_knn(query, k=10, ef_search_param=100)

DiskANN (Vamana Graph)

Best for: Filtered search, dynamic updates, very large graphs with tags/labels

import caliby
import numpy as np

# Initialize system
caliby.set_buffer_config(size_gb=2.0)
caliby.open('/tmp/caliby_data')

# Create DiskANN index
index = caliby.DiskANN(
    dimensions=128,
    max_elements=1_000_000,
    R_max_degree=64,    # Max graph degree (R)
    is_dynamic=True     # Enable dynamic inserts/deletes
)

# Build index with tags for filtering
vectors = np.random.rand(100000, 128).astype(np.float32)
tags = [[i % 100] for i in range(100000)]  # Tags for filtering

params = caliby.BuildParams()
params.L_build = 100       # Build-time search depth
params.alpha = 1.2         # Alpha parameter for Vamana
params.num_threads = 4

index.build(vectors, tags, params)

# Search with params
search_params = caliby.SearchParams(L_search=50)
search_params.beam_width = 4

query = np.random.rand(128).astype(np.float32)
labels, distances = index.search(query, K=10, params=search_params)

# Filtered search (only return vectors with specific tag)
labels, distances = index.search_with_filter(query, filter_label=42, K=10, params=search_params)

# Dynamic operations (if is_dynamic=True)
new_point = np.random.rand(128).astype(np.float32)
index.insert_point(new_point, tags=[99], external_id=100000)
index.lazy_delete(external_id=100000)
index.consolidate_deletes(params)

IVF+PQ (Inverted File with Product Quantization)

Best for: Very large datasets (10M+ vectors), memory-constrained environments

import caliby
import numpy as np

# Initialize system with buffer pool
caliby.set_buffer_config(size_gb=0.5)  # Small buffer for large datasets
caliby.open('/tmp/caliby_data')

index = caliby.IVFPQIndex(
    max_elements=10_000_000,
    dim=128,
    num_clusters=256,           # Number of IVF clusters (K)
    num_subquantizers=8,        # Number of PQ subquantizers (M), dim must be divisible by this
    retrain_interval=10000,     # Retrain centroids every N insertions
    skip_recovery=False,
    index_id=0,
    name='large_dataset'
)

# Train the index first (required for IVF+PQ)
training_data = np.random.rand(50000, 128).astype(np.float32)
index.train(training_data)

# Add points (after training)
vectors = np.random.rand(1000000, 128).astype(np.float32)
index.add_points(vectors, num_threads=4)

# Search with nprobe parameter
query = np.random.rand(128).astype(np.float32)
labels, distances = index.search_knn(query, k=10, nprobe=8)

🔧 Advanced Configuration

Persistence & Recovery

import caliby

# Indexes are automatically persisted via the buffer pool
caliby.set_buffer_config(size_gb=1.0)
caliby.open('/path/to/caliby_data')  # Data directory for persistent storage

# Create index (will be persisted automatically)
index = caliby.HnswIndex(
    max_elements=1_000_000,
    dim=128,
    M=16,
    ef_construction=200,
    enable_prefetch=True,
    skip_recovery=False,  # Set to False to enable recovery
    index_id=1,
    name='my_index'
)

# Manual flush to ensure all data is written
index.flush()

# Recovery happens automatically when reopening with same directory
caliby.close()

# Later: reopen and recover
caliby.open('/path/to/caliby_data')
recovered_index = caliby.HnswIndex(
    max_elements=1_000_000,
    dim=128,
    M=16,
    ef_construction=200,
    enable_prefetch=True,
    skip_recovery=False,  # Will recover existing index
    index_id=1,  # Must match original
    name='my_index'
)

if recovered_index.was_recovered():
    print("Index successfully recovered from disk!")

📁 Project Structure

caliby/
├── include/caliby/          # C++ headers
│   ├── calico.hpp           # Core buffer pool system
│   ├── hnsw.hpp             # HNSW index
│   ├── ivfpq.hpp            # IVF+PQ index
│   ├── diskann.hpp          # DiskANN index (experimental)
│   ├── catalog.hpp          # Index catalog management
│   └── distance.hpp         # Distance functions
├── src/                     # C++ implementation
│   ├── bindings.cpp         # Python bindings
│   ├── hnsw.cpp
│   ├── ivfpq.cpp
│   └── calico.cpp
├── examples/                # Usage examples
├── benchmark/               # Performance benchmarks
├── tests/                   # Python tests
└── third_party/             # Dependencies
    └── pybind11/            # Python binding library (submodule)

🛠️ Building from Source

Prerequisites

Caliby requires the following system dependencies:

• C++17 compatible compiler (GCC 9+ or Clang 10+)
• CMake 3.15+
• OpenMP
• Abseil C++ library
• Python 3.8+

Ubuntu/Debian:

sudo apt-get update
sudo apt-get install -y build-essential cmake libomp-dev libabsl-dev python3-dev

Fedora/RHEL:

sudo dnf install -y gcc-c++ cmake libomp-devel abseil-cpp-devel python3-devel

Build

git clone https://github.com/zxjcarrot/caliby.git
cd caliby
mkdir build && cd build
cmake .. -DCMAKE_BUILD_TYPE=Release
make -j$(nproc)

# Install Python package
cd ..
pip install -e .

Run Tests

# C++ tests
cd build && ctest --output-on-failure

# Python tests
pytest python/tests/

📚 Documentation

• Collection API Guide - High-level API for documents with vectors, text, and metadata
• Usage Guide - General usage patterns and examples
• Benchmarks - Performance comparisons and benchmarking tools

🔬 How Caliby Handles Larger-Than-Memory

Unlike in-memory libraries that crash or grind to a halt when data exceeds RAM, Caliby uses a database-style buffer pool:

┌─────────────────────────────────────────────────────────────────┐
│                        Your Application                         │
├─────────────────────────────────────────────────────────────────┤
│                     Caliby (Embedded Library)                   │
│  ┌─────────────────────────────────────────────────────────┐   │
│  │                    Buffer Pool (RAM)                     │   │
│  │   ┌─────┐ ┌─────┐ ┌─────┐ ┌─────┐ ┌─────┐ ┌─────┐      │   │
│  │   │Hot  │ │Hot  │ │Warm │ │Warm │ │Cold │ │Cold │ ...  │   │
│  │   │Page │ │Page │ │Page │ │Page │ │Page │ │Page │      │   │
│  │   └─────┘ └─────┘ └─────┘ └─────┘ └─────┘ └─────┘      │   │
│  └─────────────────────────────────────────────────────────┘   │
│                            ▲  │                                 │
│                   Evict    │  │ Parallel Fetch on               │
│                   Cold     │  │ Access                          │
│                            │  ▼                                 │
│  ┌─────────────────────────────────────────────────────────┐   │
│  │                     Disk Storage                         │   │
│  │        (SSD/NVMe)                                          │   │
│  └─────────────────────────────────────────────────────────┘   │
└─────────────────────────────────────────────────────────────────┘

Key Insight: Most vector search workloads have locality — recently accessed vectors are likely to be accessed again. Caliby exploits this by keeping hot data in RAM and seamlessly paging cold data to disk.

Data Size vs RAM	Caliby Behavior
Data < RAM	🚀 Full in-memory speed (matches HNSWLib)
Data ≈ RAM	⚡ Mostly in-memory, occasional disk reads
Data >> RAM	💾 Working set in memory, graceful disk access

🤝 Contributing

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

1. Fork the repository
2. Create a feature branch (git checkout -b feature/amazing-feature)
3. Commit your changes (git commit -m 'Add amazing feature')
4. Push to the branch (git push origin feature/amazing-feature)
5. Open a Pull Request

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.

📬 Contact

• Issues: GitHub Issues
• Discussions: GitHub Discussions
• Email: xinjing@mit.edu

---