Mohamed KEITA

CortexDB v4

Full vector layer for CPU-first semantic search.

1. Overall Vision of V4

V4 adds a complete vector layer to CortexDB, making it AI-ready. It is divided into two sub-versions:

  • V4.1: Vector Core (collections, HNSW, 8-bit quantization, persistence)
  • V4.2: Vector DX (advanced filters, hybrid search, batch search, IVF-Lite, 4-bit quantization, ingestion helpers)

Main objectives:

  • Enable semantic search capabilities in CortexDB
  • Optimize for CPU, not GPU
  • Support embeddings and vector search
  • Deliver acceptable performance on modest machines

2. V4 Structure

V4.1 — Vector Core

Features:

  • Vector collections
  • L2, Cosine, and Dot-product distances
  • CPU-first HNSW index
  • 8-bit quantization
  • Snapshot persistence of the index
  • Simple filters (post-filtering)

V4.2 — Vector DX

Features:

  • Advanced filters with indexing
  • Hybrid search (vector plus text)
  • Batch search
  • IVF-Lite index as an alternative to HNSW
  • 4-bit quantization
  • Embedding ingestion helpers on the client side

3. V4 Non-Goals

  • No GPU support
  • No replication (V5)
  • No CortexQL (V6)
  • No distribution (V7)
  • No fine-tuning of embedding models

4. V4 Architecture Overview

┌───────────┐
│   Client  │
└─────┬─────┘
      │  vector_search(query, k=10)

┌───────────────────┐
│   Vector Engine   │
│  - Collections    │
│  - HNSW or IVF    │
│  - Search         │
└─────┬─────────────┘


┌───────────────────┐
│  Storage Engine   │
│  (V3)             │
└───────────────────┘

5. V4 Objective

At the end of V4 (V4.1 + V4.2), CortexDB becomes:

  • AI-ready: full vector search capabilities
  • CPU-optimized: acceptable performance on CPU
  • Memory-efficient: quantization to reduce memory usage
  • Production-ready: suitable for AI applications

This is the version that makes CortexDB unique through its CPU-first vector search capability.