DRAG: Domain-Routed Adaptive Decay for Temporally Heterogeneous Retrieval

Standard retrieval treats all documents as equally persistent. DRAG applies domain-aware decay so volatile facts expire while persistent knowledge survives.

Any uniform temporal decay forces a strict Pareto trade-off: decay aggressively enough to suppress stale stock prices and you also bury valid corporate profiles; decay gently enough to preserve reference material and outdated news contaminates results. DRAG breaks this frontier by applying distinct half-lives per document domain — a simple multiplicative weight on standard ANN cosine similarity scores.

Key Results

Method	Overall MRR	Reference MRR	News MRR
HNSW raw	0.709	0.966	0.452
Best uniform decay (h=5.0)	0.914	0.946	0.882
DRAG (domain-aware)	0.958	0.966	0.950

On a synthetic financial corpus with genuine temporal invalidity, DRAG simultaneously maximizes retrieval of persistent reference material and volatile news — something no single decay rate can achieve.

Full Results Summary

Experiment	Question	Result
1. Semantic Separability	Does the 64D projection preserve retrieval structure?	✅ 1.40 separation ratio vs 1.14 baseline; 98.3% triplet accuracy
2. Retrieval Quality	Does Lightcone geometry beat HNSW?	❌ NDCG@10 = 0.008 vs HNSW+TW = 0.642
2b. Compression Ablation	Is the failure caused by 64D compression?	❌ 384D Lightcone NDCG@10 = 0.011 — cone geometry itself fails
2c. Per-Query c	Does per-query calibration rescue the geometry?	❌ NDCG@10 = 0.019 — still 33× worse than HNSW+TW
2d. Minkowski Reranking	Does τ work as a reranker on HNSW candidates?	❌ Every reranker degrades HNSW quality
2e. Dense Timestamps	Does temporal density help the geometry?	❌ Converges to parity — reduces to cosine ranking
2f. DRAG (geometric)	Does domain routing with Minkowski scoring help?	❌ Best variant is the one with no temporal scoring
2g. Heterogeneous Corpus	Does domain-aware decay beat uniform decay?	✅ +0.118 NDCG over time filter
2h. Independent GT	Does decay help on timeless factual QA?	❌ Decay hurts — answers don't expire
2i. Time-Varying Corpus	Does DRAG break the Pareto frontier?	✅ MRR = 0.958 vs 0.914 best uniform
2j. Learned Profiles	Can decay profiles be learned without labels?	⚠️ Learns direction but not persistence scale
2k–2m. StackOverflow	Does the result replicate on real-world data?	✅ Persistent domain validated; volatile limited by frozen GT
3. Concentration of Measure	Does the cone filter discriminate at 64D?	✅ Tunable, but temporally — not semantically

How It Works

Ingestion:
  chunk → encoder → embedding → ANN index
        → domain classifier → decay profile (half_life, floor)
        → timestamp

Query:
  query → encoder → ANN top-K → reweight by cos_sim × decay_weight → top-k

score(q, d) = cos_sim(q, d) × max(exp(-0.693 × Δt / half_life), floor)

Two parameters per domain: a half-life (how fast relevance decays) and an importance floor (minimum weight regardless of age).

Repository Structure

├── paper/
│   ├── drag.md                   # Standalone DRAG paper
│   ├── iteration-7.md            # Full empirical paper (LRAG + DRAG)
│   ├── iteration-6.md            # Earlier empirical paper
│   ├── iteration-5.md            # Pre-experiment theoretical paper
│   ├── lightcone_paper-iter-5-1.md
│   └── citations_search.json     # arXiv citation search results
├── experiments/
│   ├── local/                    # Standalone Python scripts (run anywhere)
│   │   ├── experiment_1.py       # Semantic separability
│   │   ├── experiment_2.py       # Retrieval quality vs HNSW
│   │   ├── experiment_2b.py      # 384D compression ablation
│   │   ├── experiment_2c.py      # Per-query c calibration
│   │   ├── experiment_2d.py      # Minkowski reranking
│   │   ├── experiment_2e.py      # Dense timestamp configs
│   │   ├── experiment_2f.py      # Domain-routed adaptive geometry
│   │   ├── experiment_2g.py      # Heterogeneous corpus decay
│   │   ├── experiment_2h.py      # Independent ground truth
│   │   └── experiment_3.py       # Concentration of measure
│   └── modal/                    # Modal cloud versions (GPU)
│       ├── experiment_1.py
│       ├── experiment_2.py – experiment_2m.py
│       └── experiment_3.py
├── results/                      # Raw JSON results from all experiments
│   ├── experiment_1_results.json
│   ├── experiment_2_results.json – experiment_2m_results.json
│   └── experiment_3_results.json
└── findings/                     # Per-experiment analysis writeups
    ├── experiment_1.md
    ├── experiment_2.md – experiment_2j.md
    └── experiment_3.md

Running the Experiments

Local (CPU or GPU)

pip install torch numpy scikit-learn sentence-transformers datasets hnswlib

# Experiment 1: Semantic separability (~2 min on GPU, ~10 min on CPU)
python experiments/local/experiment_1.py --device cuda

# Experiment 2: Retrieval quality vs HNSW (~5 min on GPU)
python experiments/local/experiment_2.py --device cuda

# Experiment 2b–2h: Repair strategies and decay experiments
python experiments/local/experiment_2b.py --device cuda
python experiments/local/experiment_2c.py --device cuda
python experiments/local/experiment_2d.py --device cuda
python experiments/local/experiment_2e.py --device cuda
python experiments/local/experiment_2f.py --device cuda
python experiments/local/experiment_2g.py --device cuda
python experiments/local/experiment_2h.py --device cuda

# Experiment 3: Concentration of measure (~3 min on GPU)
python experiments/local/experiment_3.py --device cuda

Modal (cloud GPU)

pip install modal
modal setup

modal run experiments/modal/experiment_1.py
modal run experiments/modal/experiment_2.py
# ... through experiment_2m.py
modal run experiments/modal/experiment_3.py

Results are saved to results/ (local) or the Modal volume (cloud).

Background

This repository began as an investigation into Lightcone Retrieval — modeling documents as events in a Lorentzian manifold with the Minkowski interval as a retrieval metric. That geometric approach failed decisively (see experiments 2–2f). However, the differential decay model proposed alongside the geometry — the idea that different document types should persist at different rates — proved empirically sound. DRAG extracts and validates that positive result as a standalone contribution.

Citation

@software{drag_2026,
  author       = {VectorJet},
  title        = {DRAG: Domain-Routed Adaptive Decay for Temporally Heterogeneous Retrieval},
  year         = {2026},
  url          = {https://github.com/VectorJet/drag},
  note         = {Domain-aware decay breaks the uniform-decay Pareto frontier on temporally heterogeneous corpora}
}

License

MIT