Spectrum CodeRAG Benchmark HUD

Retrieval-aware storage, measured as a serving path.

Public snapshot from the current large-corpus benchmark. The live benchmark runner is available in the repository and is best run on a larger VM or a local workstation.

Spectrum Overview HUD Source

Accurate Serving Profile

Linux-scale corpus

Documents

72,601

Raw Corpus

1.03 GB

Compact Store

0.3585x

Serving Footprint

0.4882x

Hit@1

0.8500

MRR

0.8625

Recall@5

0.8750

P95 E2E

6.35 ms

Serving Path

Search, rank, hydrate

Spectrum stores compact .spec payloads with a retrieval-aware token layer, serves snippets without full hydration, and decodes selected payloads on demand for exact reconstruction.

Current frontier: selected-payload decode, sidecar footprint, and memory efficiency.

Engine View

What the local HUD compares

Engine	Storage	Latency	Rebuild
Spectrum CodeRAG	Compact searchable store	Sub-10 ms serving signal	Lossless
Raw BM25	Raw chunks plus sparse index	Fast lexical baseline	Raw source retained
TF-IDF	Raw chunks plus vector matrix	Fast small-corpus baseline	Raw source retained
FAISS Flat	Dense vector sidecar	Vector scan baseline	Index only
Hybrid	Sparse plus dense sidecars	Blended retrieval baseline	Raw source retained