Pithos - Model-Isomorphic Vector Database (MIDB)

Pithos is a high-performance Model-Isomorphic Database (MIDB) and Ahead-of-Time (AOT) compiled vector search engine designed for Matryoshka-structured binary embeddings at planetary scale, compiled into a standalone native shared library (.dylib / .so) via GraalVM Native Image.

What is a Model-Isomorphic Database (MIDB)?

Traditional vector databases treat embeddings as generic high-dimensional arrays. A Model-Isomorphic Database (MIDB) structurally mirrors the latent geometry and spectral energy distribution of the neural embedding model:

• SVD-Driven Spectral Truncation: Columnar tiers are allocated according to cumulative singular value energy (Φ(k)) derived from model projection/LoRA weights.
• Isometric Preconditioning & Rotation: Eliminates spatial burstiness via Rademacher sign flipping and spreads embedding energy uniformly with block-diagonal Fast Walsh-Hadamard Transforms (H_BD).
• 3-Gate Read-Path Cascade: 1-cycle metadata filtering → Matryoshka early-exit Hamming scanning → exact FP16 in-engine reranking.
• Zero-GC Off-Heap Memory: Bypasses garbage collection entirely using the Java Foreign Function & Memory (FFM) API (Project Panama) and POSIX memory-mapped I/O (mmap).
• Hardware SIMD & CUDA: Vectorized with Java Vector API (AVX-512 / ARM NEON) and native NVIDIA CUDA kernels for batch distance computation and multi-family resonant voting.
• Pythonic Zero-Copy FFI: Seamless integration with NumPy arrays via pithosdb.

Python Quickstart

Install the official Python package:

pip install pithosdb
# or with uv:
uv pip install pithosdb

import pithosdb
import numpy as np

# 1. Open database off-heap (Zero JVM overhead)
with pithosdb.VectorDb() as db:
    # 2. Compile an index from float embeddings
    records = np.random.randn(10_000, 384).astype(np.float32)
    pithosdb.VectorDb.compile_index(
        base_path="temp/sample_index",
        records=records,
        tiers=[64, 128, 256, 384]
    )
    
    # 3. Memory-map index & run zero-copy batch k-NN search
    index = db.load_index("sample", "temp/sample_index")
    queries = np.random.randn(10, 384).astype(np.float32)
    results = index.search(queries, k=5)
    
    for q_idx, matches in enumerate(results):
        print(f"Query {q_idx} Top Matches: {matches}")

    # 4. Real-time Ingestion via LSM DeltaBuffer
    delta = db.create_delta_buffer("sample", flush_threshold=1000)
    delta.insert(record_id=42, vector=np.random.randn(384).astype(np.float32))

Precompiled Native Binaries

Precompiled native libraries are automatically published on GitHub Releases:

Download Latest Release Assets

System Architecture & Features

┌─────────────────────────────────────────────────────────────┐
│                 Client Layer (Python / C / C++)             │
│            pithosdb (ctypes Zero-Copy NumPy / FFI)          │
└──────────────────────────────┬──────────────────────────────┘
                               │ C-ABI (vdb_*)
┌──────────────────────────────▼──────────────────────────────┐
│                    Pithos Core Engine                       │
│  ┌───────────────────────────────────────────────────────┐  │
│  │   LMAX Disruptor Lock-Free Multi-Threaded Workers     │  │
│  └───────────────────────────┬───────────────────────────┘  │
│                              │                              │
│  ┌───────────────────────────▼───────────────────────────┐  │
│  │               3-Gate Read-Path Cascade                │  │
│  │  Gate 1: Metadata & Tombstone Filter (1 cycle)        │  │
│  │  Gate 2: Matryoshka Early-Exit Hamming Scan (SIMD)    │  │
│  │  Gate 3: In-Engine FP16 / Asymmetric Reranking        │  │
│  └───────────────────────────────────────────────────────┘  │
│                              │                              │
│  ┌───────────────────────────▼───────────────────────────┐  │
│  │       Project Panama Off-Heap Storage (POSIX mmap)    │  │
│  │       - <name> (64B Header)                           │  │
│  │       - <name>_tier_*.bin (Packed Columnar Bits)      │  │
│  │       - <name>_metadata.bin (Attributes & Flags)      │  │
│  │       - <name>_fp16.bin (Half-Precision Sidecar)      │  │
│  └───────────────────────────────────────────────────────┘  │
└─────────────────────────────────────────────────────────────┘

Documentation

Detailed architectural guides, mathematical specifications, and C-API references:

• Architectural Principles & Core Innovations: Mathematical foundations, block-diagonal Walsh-Hadamard rotations, SVD-driven spectral truncation, and the 3-gate read-path cascade.
• C-API Reference & Runtime Configuration: Complete declarations of entry points (libpithos), FFI mappings, CUDA wrappers, and hardware co-design guidelines (FPGA/DMA offloading).
• CUDA GPU Acceleration Guide: Shared memory popcount kernels, asynchronous stream pipelines, and multi-family voting.

Building from Source

Prerequisites

• GraalVM JDK 25 (with native-image)
• Apache Maven 3.9+
• (Optional) NVIDIA CUDA Toolkit 12+ for GPU kernels

1. Compile Native Library (macOS & Linux)

export JAVA_HOME=/path/to/graalvm-jdk-25
export PATH=$JAVA_HOME/bin:$PATH

mvn clean package -DskipTests

The compiled shared library is generated in target/pithos.dylib (macOS) or target/pithos.so (Linux).

2. Run Test Suite

mvn test

3. Build with CUDA Support (Linux)

mvn clean package -Pcuda -Dcuda.enabled=true

License

Licensed under the Apache License, Version 2.0.