Reproduce structure pricing and low cost graph retrieval from EffiRAG

Finding

EffiRAG, arXiv:2609.18099, submitted September 16, 2026, reports a materially different graph retrieval cost frontier. On UltraDomain, 120 open ended questions across four domains, EffiRAG is preferred over LightRAG hybrid on 93 questions versus 7, with 20 splits, while reducing total ingestion plus query cost by 57 percent from USD 0.952 to USD 0.408. At larger corpus sizes the originating team reports 4.2 times and 4.5 times lower cost using a lightweight non LLM salience filter to skip low value chunks.

Evidence class: originating team measured result. This is not independently reproduced by RuV. No production cost claim is made until model prices, corpus construction, judge protocol, and index build work are reproduced.

Primary source: https://arxiv.org/abs/2609.18099

Opportunity score

4.30 out of 5 under the RuV daily SOTA weighting model.

Scores: applicability 5, performance impact 4.5, implementation speed 4, cross stack leverage 5, commercial value 5, strategic differentiation 4, security improvement 2.5, open source leadership 3.5, evidence confidence 3.5, long horizon option value 4.5, experiment reversibility 5.

Hypothesis

RuVector should price graph structure explicitly. A graph edge, community, or relation should only be materialized when its expected retrieval value exceeds its ingestion, storage, and query cost. A cheap deterministic or statistical salience gate may reduce graph construction calls while preserving original source passages for answer generation.

This is complementary to issue #998. #998 separates graph path multiplicity from evidence origin multiplicity. This issue measures whether graph structure is worth constructing at all.

Frozen arms

1. Current RuVector hybrid retrieval baseline.
2. Vector plus lexical retrieval without graph construction.
3. Existing graph retrieval with current ingestion path.
4. Graph retrieval with a non LLM salience gate before structure construction.
5. Oracle using retrospective query relevance to estimate the upper bound on structure pricing.

All arms must use the same source corpus, embedding model, answer model, judge protocol, and query set.

Acceptance

Advance only if the candidate reduces total ingestion plus query cost by at least 30 percent versus the strongest current graph baseline while retaining answer quality within 1 absolute preference point or improving it. It must also reduce ingestion model calls by at least 40 percent or index construction wall time by at least 25 percent.

If the cheap gate disproportionately removes rare but decisive evidence, record the negative result and reject the optimization even if cost falls.

Benchmark contract

Report exact commits, model and embedding versions, provider pricing date, corpus, document count, token count, query count, seeds, ingestion calls, query calls, input and output tokens, build time, index bytes, query p50 and p95 latency, answer quality, source recall, rare evidence recall, total dollar cost, CPU and memory overhead, failures, and reproduction steps.

Cross stack mapping

RuVector owns structure pricing, salience gating, and retrieval metrics. RuVector WASM can run the cheap salience gate locally when feasible. WorldGraph can reuse the rule for deciding when to materialize derived relations. MetaHarness independently measures quality and cost. RuFlo can allocate retrieval budget but cannot change acceptance thresholds. RVF binds source passages and graph construction receipts. Autogenous and Dream Machine may tune salience parameters only after the benchmark is frozen. MidStream records ingestion and query cost events. RuView and RuField can use structure pricing for high volume sensing graphs. LatentMesh can avoid transmitting low value graph updates. Cognitum gets a direct cost control for enterprise retrieval. MCP resources preserve original source identity. Distributed infrastructure can route cheap gates to CPU or browser workers before expensive model calls.

Governance

Reversible benchmark only. Do not import an upstream runtime, change production retrieval, or claim the reported 57 percent saving without independent RuV measurement. No autonomous merge or deployment.

The issue names no files or tests. Start by locating the current hybrid and graph retrieval ingestion paths, the WASM salience-gate entry point, and the MetaHarness benchmark configuration. Run the five frozen arms on the same corpus, models, queries, and judge protocol; done means meeting the cost, quality, and ingestion or build-time thresholds with the full benchmark contract reported.