ChronosVector (CVX)
A high-performance temporal vector database that treats time as a geometric dimension of embedding space. ChronosVector answers not just "what is similar?" but "what was similar, what changed, and what will be similar?"
Documentation
| Resource | Link |
|---|---|
| Full Documentation | manucouto1.github.io/chronos-vector |
| White Paper | Research Overview |
| Temporal Analytics API (19 functions) | Toolkit Reference |
| Rust API (cargo doc) | API Reference |
| RFC-006: Anchor Projection | RFC |
Tutorials (with interactive Plotly visualizations)
| Tutorial | Domain | Key Result |
|---|---|---|
| Mental Health Explorer | Clinical NLP | 13 CVX features → F1=0.600 |
| Clinical Anchoring | Clinical NLP | DSM-5 anchors → F1=0.744, AUC=0.886 |
| Political Rhetoric | Political NLP | Trump tweets + S&P 500 alignment |
| Market Regimes | Finance | 11 changepoints, Hurst=0.74 |
| Anomaly Detection | Time Series | NAB benchmark, 4 detection strategies |
| MAP-Elites | Quality-Diversity | HNSW as adaptive niche discovery |
| MLOps Drift | Production ML | 5 independent drift signals |
Key Features
- Spatiotemporal kNN — composite distance
d_ST = α·d_semantic + (1-α)·d_temporal - 19 analytical functions — velocity, drift, Hurst, changepoints, path signatures, topology, anchor projection
- Anchor projection — project trajectories from ℝᴰ to interpretable ℝᴷ coordinates (RFC-006)
- SIMD-accelerated — auto-dispatched AVX2/NEON via
pulp(cosine, L2, dot product) - Index persistence — save/load HNSW graph via postcard binary serialization
- Tiered storage — hot (in-memory/RocksDB) → warm (file-based) with automatic compaction
- Crash safety — WAL with CRC32 validation and segment rotation
- Python bindings —
pip install chronos-vector(PyO3/maturin)
Python Quick Start
# Create and populate index
=
# persist for fast reload
# Trajectory analysis
=
=
=
=
# Anchor projection — measure relative to reference points
=
= # {mean, min, trend, last}
# All analytics work on projected trajectory too
=
=
Performance
| Metric | Value |
|---|---|
| HNSW recall@10 (1K, D=32) | 1.000 |
| HNSW recall@10 (10K, D=128) | 0.956 |
| Graph reachability (10K) | 100% |
| PELT F1 (3 planted CPs) | 1.000 |
| Bitmap memory (100K vectors) | 0.16 bytes/vector |
| Index save/load (225K, D=768) | < 1s (vs 500s rebuild) |
Architecture
┌──────────┐
│ cvx-api │ REST + gRPC endpoints
└────┬─────┘
│
┌──────────┼──────────┐
│ │ │
┌─────┴──┐ ┌─────┴──┐ ┌────┴──────┐
│ ingest │ │ query │ │ analytics │ 19 functions
└────┬───┘ └────┬───┘ └────┬──────┘ + anchor projection
│ │ │
┌────┴──────────┴──────────┴────┐
│ cvx-index │
│ ST-HNSW + SIMD + Roaring BM │ save/load persistence
└──────────────┬────────────────┘
│
┌──────────────┴────────────────┐
│ cvx-storage │
│ Hot → Warm + WAL + RocksDB │
└──────────────┬────────────────┘
│
┌──────────────┴────────────────┐
│ cvx-core │
│ Types, Traits, Config, Error │
└───────────────────────────────┘
8 crates + Python bindings, 280+ tests.
| Crate | Description |
|---|---|
cvx-core |
Types, traits, config, error handling |
cvx-index |
ST-HNSW with temporal filtering, scalar quantization, persistence |
cvx-analytics |
19 analytical functions: calculus, signatures, topology, anchors |
cvx-storage |
Hot/warm tiers, WAL, RocksDB, tiered routing |
cvx-ingest |
Delta encoding, input validation |
cvx-query |
Query engine |
cvx-api |
REST + gRPC handlers |
cvx-server |
Server binary with graceful shutdown |
cvx-python |
Python bindings via PyO3/maturin |
Rust Quick Start
# Build
# Run tests
# Generate API docs
# Start server
Python Installation
# From source (requires Rust toolchain)
# Or with pip (when published)
Cross-Domain Research
CVX has been validated across 7 investigations in 6 domains. See the White Paper for details.
| Domain | Dataset | CVX Contribution |
|---|---|---|
| Clinical NLP | eRisk 2017-2022 | DSM-5 anchor projection: F1=0.744, AUC=0.886 |
| Political NLP | Trump Twitter Archive | Rhetorical regime detection via changepoints + signatures |
| Finance | S&P 500 Sector ETFs | 11 regime changepoints, Hurst=0.74, path signatures |
| Anomaly Detection | Numenta NAB | Trajectory-geometric detection (velocity, topology) |
| Fraud Detection | IEEE-CIS | Transaction trajectory fingerprinting |
| Cybersecurity | CERT CMU | Behavioral regime shift detection |
| Quality-Diversity | MAP-Elites | HNSW as adaptive niche discovery |
License
Licensed under the Elastic License 2.0.
Free to use for research, education, and internal purposes. Commercial use as a managed service requires a separate license. See LICENSE for details.