thaki-AI/daily-paper-2026-07-09-retriever-vs-decomposition-skill-routing

Retriever Bottleneck vs. Decomposition

TL;DR — In a real 1,898-skill bilingual agent harness, the ceiling-gap diagnostic shows that neither decomposition nor a better retriever raises routing coverage: the binding constraint is skill-corpus redundancy. Decomposition still earns its place on execution ordering.

ThakiCloud AI Research · 2026-07-09 (v2, revised) · 📝 Tech blog (KO)

Problem

Operators of large agent harnesses that route natural-language requests to hundreds of skills face a costly decision: when routing quality is poor, invest in a decomposition layer or in a better retriever and index? The two investments have very different costs, and choosing wrongly wastes engineering effort.

Approach

We introduce the ceiling-gap diagnostic, a decomposition-free procedure that compares single-pass retrieval (SINGLE) against a gate-free oracle-decomposition upper bound (ORACLE) computed with no new LLM pass. It reads two numbers: ORACLE's absolute level and its gap to SINGLE. A high ORACLE with a large gap argues for decomposition; a low ORACLE argues that retriever, index, and corpus quality bind.

Key findings (12-case compositional benchmark, bootstrap CIs)

• SINGLE 52.9% (CI [34.0, 72.2]); ORACLE 63.6% (CI [42.8, 83.5]); ceiling gap 10.7pp (CI [2.1, 20.8]).
• Decomposition does not help coverage: SAD 35.0%, ISAD 37.1%, SAD-AGENT 41.9% — all below SINGLE.
• A better retriever does not raise the ceiling: dense 45.4%, BM25/dense hybrid 64.0% ≈ live BM25 63.6%.
• The binding constraint is corpus redundancy: of 12 misses, 5 are family saturation (gold ranked out by near-substitutable siblings, e.g. trading-position-sizer rank 38) and 7 are gold/corpus name mismatch (capability present only under sibling names). Dense fails identically; automatic dedup/equivalence moves ORACLE ≤2.1pp.
• Retrieval is a synonym-dictionary artifact: removing the hand-curated KO→EN dict collapses ORACLE 63.6→20.8; a prior "+21pp" was that dict tuned on eval vocabulary (train-on-test), retracted here.
• Decomposition's real value is execution ordering: Kendall τ=1.0 vs SINGLE's −0.22, invisible to coverage.

Correction vs. an earlier draft

An earlier version stated SkillWeaver (arXiv:2606.18051) reports "oracle top-1 recall ~99.5%". That was a misreading: SkillWeaver's 99% figure is context-window reduction, and its retrieval quality is category recall@1 34%→41%. The framing is corrected in this version.

Limitations

n=12 case study (all numbers with bootstrap CIs); single private corpus; dense arm used a small multilingual model (large models untested); substitutability not cleanly machine-measurable; English–Korean only; no live A/B. A sanitized public replication is the next artifact.

Files

• main.pdf / main.tex — the paper
• figures/ — figures (English labels)
• abstract.txt — abstract