Expand description
Temporal Tensor Compression with Tiered Quantization
Implements ADR-017: groupwise symmetric quantization with temporal segment reuse and access-pattern-driven tier selection (8/7/5/3 bit).
§Architecture
f32 frame → tier_policy → quantizer → bitpack → segment
segment → bitpack → quantizer → f32 output§Compression Ratios
| Tier | Bits | Ratio vs f32 | Use Case |
|---|---|---|---|
| Hot | 8 | ~4.0x | Frequently accessed tensors |
| Warm | 7 | ~4.57x | Moderately accessed |
| Warm | 5 | ~6.4x | Aggressively compressed warm |
| Cold | 3 | ~10.67x | Rarely accessed |
§Zero Dependencies
This crate has no external dependencies, making it fully WASM-compatible.
§Quick Start
use ruvector_temporal_tensor::{TemporalTensorCompressor, TierPolicy};
// Create a compressor for 128-element tensors
let mut comp = TemporalTensorCompressor::new(TierPolicy::default(), 128, 0);
comp.set_access(100, 0); // hot tensor -> 8-bit quantization
let frame = vec![1.0f32; 128];
let mut segment = Vec::new();
// Push frames; segment is populated when a boundary is crossed
comp.push_frame(&frame, 1, &mut segment);
comp.flush(&mut segment); // force-emit the current segment
// Decode the segment back to f32
let mut decoded = Vec::new();
ruvector_temporal_tensor::segment::decode(&segment, &mut decoded);
assert_eq!(decoded.len(), 128);§Random-Access Decode
// Decode only frame 0 without decoding the entire segment
let single = ruvector_temporal_tensor::segment::decode_single_frame(&seg, 0);
assert!(single.is_some());§Compression Ratio Inspection
let ratio = ruvector_temporal_tensor::segment::compression_ratio(&seg);
assert!(ratio > 1.0);Re-exports§
pub use compressor::TemporalTensorCompressor;pub use tier_policy::TierPolicy;
Modules§
- agentdb
- AgentDB adapter for pattern-aware tiering.
- bitpack
- Bitstream packer/unpacker for arbitrary bit widths (1-8).
- coherence
- Coherence gate: read-after-write validation for the temporal tensor store.
- compressor
- TemporalTensorCompressor: the main entry point.
- core_
trait - Abstract trait interface for tensor block storage.
- delta
- Delta compression, delta chains, and reconstruction policies (ADR-021).
- f16
- Software IEEE 754 half-precision (f16) conversion.
- metrics
- Witness logging and decision audit for the temporal tensor store.
- quantizer
- Groupwise symmetric quantization with f16 scales.
- segment
- Segment binary format: encode and decode.
- store
- Block-based storage engine for temporal tensor compression (ADR-018).
- tier_
policy - Tier policy for access-pattern-driven bit-width selection.
- tiering
- Enhanced temporal scoring with EMA + popcount + recency, hysteresis, and budgeted maintenance (ADR-020).