Eidet Retrieval Benchmark Scorecard

This measures the part of recall Eidet owns: ranking, hybrid fusion, and type-budget quality on a curated, adversarial in-repo dataset. Each gold case scripts per-arm raw scores (lex, vec); the runner feeds them through the real RecallScoring.Fuse + ApplyTypeBudgets pipeline and compares the v2 fusion ranker against a flat baseline (the pre-#33 scoring: lexical hits = 1.0, vector-only hits = 0.9, no normalization, no UCB, no recency) over the same dataset. The lift is fusion recovering gold that flat scoring loses at the truncation boundary.

It does not measure embedding quality (RavenDB owns embeddings, which can’t run deterministically in CI), and it is not the external SWE Context Bench leaderboard — that is a deferred epic. This is a deterministic, no-LLM, CI regression guard.

v2 Fusion vs Flat Baseline

Capability Cases Metric v2 Fusion Flat Baseline Delta
Recall 14 Recall@k 0.857 0.321 +0.536
    MRR 0.833 0.405 +0.429
    nDCG@k 0.804 0.246 +0.559
    Gold survival 0.750 0.357 +0.393
StateUpdating 1 Recall@k 1.000 1.000 +0.000
    MRR 1.000 1.000 +0.000
    nDCG@k 1.000 1.000 +0.000
    Gold survival 1.000 1.000 +0.000
Overall 15 Recall@k 0.867 0.367 +0.500
    MRR 0.844 0.444 +0.400
    nDCG@k 0.817 0.296 +0.522
    Gold survival 0.767 0.400 +0.367

Capabilities not evaluated

  • CausalInference — N/A — requires LLM-in-loop; deferred to the SWE Context Bench epic (issue #36 follow-up).
  • StateAbstraction — N/A — requires LLM-in-loop; deferred to the SWE Context Bench epic (issue #36 follow-up).

© 2026 Steve Hansen. Eidet is MIT licensed.

This site uses Just the Docs, a documentation theme for Jekyll.