PaperlayerView exact revision on arXiv

Source-linked AI summary

Seed-Anchored Budget-Bounded Graph Rendering for Question Answering on Industry-Standard Power-Grid Information and Exchange Models

Jayakumar Manoharan, Yamini Sehgal

arXiv:2609.02011v1eess.SYcs.AIcs.IR

TL;DR

LLM assistants answering over power-grid models must preserve answer-relevant graph structure within a fixed context budget, but conventional rendering can truncate that evidence. The paper introduces deterministic seed-anchored rendering, which prioritizes query-local units and guarantees their preservation when their render mass fits the budget. On evaluated CIM/CGMES models, it improves bounded-retrieval outcomes and matches or exceeds tested LLM-extracted graphs without graph-construction tokens.

  • Problem

    Power-grid LLM question answering requires selecting a context-bounded subgraph without losing answer-relevant information, while prior native-graph deficits may reflect rendering truncation rather than missing structure.

  • Method

    Seed-anchored rendering deterministically places the query object and incident edges first, expands by hop distance, and uses a checkable D(S) ≤B preservation condition without additional tuned or learned parameters.

  • Results

    0.450 to 0.970 accuracy on a fresh SmallGrid-family 100-item bank accompanied evidence preservation from 0.12 and 0.00 under naive rendering to all multi-hop evidence under seed anchoring.

  • Takeaways & Limitations

    At zero graph-construction tokens, the standards-native seed-anchored graph matches or exceeds LightRAG, Microsoft GraphRAG, and HippoRAG under the common tested pipeline.

  • Takeaways & Limitations

    Evidence is limited to evaluated CIM/CGMES models, reader, context budget, and retrieval setting rather than general question answering or broad network-level generalization.

Abstract

from arXiv · show

Large language model question answering over power-grid models must respect a fixed context budget. We introduce seed-anchored graph rendering, a deterministic method that prioritizes query-local graph evidence without adding method-specific tuned or learned parameters beyond the shared hop bound and context budget. The method provides a checkable condition under which predefined seed-local answer-bearing render units are preserved in a greedy bounded-context prefix. We evaluate the approach on Common Information Model (CIM) network models exchanged through the Common Grid Model Exchange Standard (CGMES). On two budget-binding CGMES encodings, naive descriptions-first rendering retains local evidence for every single-hop item but only 0.12 and 0.00 of multi-hop items, whereas seed-anchored rendering retains all such evidence. On a preregistered fresh 100-item bank from the SmallGrid topology family, accuracy rises from 0.450 to 0.970 under a fixed 8,000-character context budget. Under a common retrieval and rendering pipeline, the standards-native seed-anchored graph matches or exceeds extracted graph representations produced by LightRAG, Microsoft GraphRAG, and HippoRAG, while avoiding LLM graph-construction tokens. The results are specific to the evaluated CIM/CGMES models, reader, and context budget; they concern budget-bounded retrieval rather than general question answering.

1. Introduction

The paper frames power-grid LLM assistance as budget-bounded graph rendering: retrieval must fit a finite context while preserving answer-relevant structure. It introduces seed-anchored rendering and evaluates its evidence preservation, accuracy, and comparison with LLM-extracted graphs on CGMES models.

  • Motivation: CGMES exports provide curated, schema-typed relational models that can be queried directly instead of reconstructed by an LLM.The paper contrasts standards-native representations with LLM extraction, which is expensive and lossy.
  • Problem: High-degree hubs can fill the context before answer-bearing edges are rendered, making retrieval order—not representation alone—the failure mechanism.The failure can flip between raw CIM and derived CGMES encodings.
  • Method: Seed-anchored rendering places the query object and incident edges first, then expands by hop distance under the shared budget and hop bound.It adds no tuned or learned parameters and provides a checkable seed-local render-mass condition.
  • Results: 0.12 and 0.00 of multi-hop evidence survive naive rendering on SmallGrid and SmallGrid-3.0, versus 1.00 on both networks with seed anchoring.Naive rendering retains all single-hop evidence, while seed anchoring retains all evidence in both strata.
  • Results: At zero graph-construction tokens, the standards-native arm matches or exceeds LightRAG, Microsoft GraphRAG, and HippoRAG under the common pipeline.The comparison is conditional on the tested extractor, reader, budget, and framework set.
  • Scope: The guarantee is domain-agnostic in proof but empirically validated only on CIM/CGMES topologies and the evaluated reader and budget.The paper does not claim general question-answering performance or empirical validation in other high-degree-hub settings.

2. Related work

Related work commonly applies locality at graph retrieval or scoring, while this paper places it at the render/truncation boundary. It situates the contribution within GraphRAG, power-domain retrieval, CIM/CGMES semantics, and bounded-context rendering.

  • GraphRAG: GraphRAG systems typically build a knowledge graph, retrieve a relevant subgraph, and render it for an LLM reader.LightRAG and HippoRAG pursue cheaper or memory-based alternatives to Microsoft GraphRAG.
  • Locality priors: Existing locality priors select candidate nodes or edges by proximity, but a later render ordering can still determine which units survive truncation.The paper identifies this second ordering as the site of the standards-native failure.
  • Locality priors: A relevance-ranked budget fill can let a relevant high-degree hub flood the window, whereas hop-locality rendering places seed-incident edges first.The methods can have similar reader-free answer presence, but seed anchoring adds a checkable guarantee.
  • Power-domain retrieval: Power-domain work includes LLM-extracted graphs, grid-code retrieval, and earlier power knowledge-graph QA, alongside reliability evaluations of LLM assistants.The paper contributes a mechanism-level account of budget-induced retrieval failure in this setting.
  • CIM/CGMES: CIM/CGMES is already a schema-defined semantic resource commonly accessed as a graph, so the contribution is linking direct standards reading to budget preservation.Established tooling generally assumes full in-memory access or small inputs rather than bounded LLM contexts.
  • Bounded rendering: Bounded-context rendering can degrade access to relevant content, and this paper studies the specific high-degree-hub truncation mechanism quantitatively.The paper distinguishes this mechanism from the broader known phenomenon of overstuffed contexts.

3. Background

The paper represents CGMES exports as deterministic standards-native or derived graphs and as a deterministic text corpus, then evaluates each through a frozen bounded-context QA pipeline. Its motivating question is whether the native graph’s deficit reflects representation or rendering truncation.

  • CGMES representation: CGMES exports describe typed grid objects joined by explicit schema associations in RDF/XML profiles.The profiles cover equipment, topology, steady-state hypothesis, and state variables.
  • CGMES representation: The deterministic parser produces GSTD, GHYB, and a corpus rendering of the same network facts.GSTD retains raw CIM associations, GHYB derives five relation types, and the corpus renders facts into text.
  • Evaluation setup: No language model produces the ground truth, corpus, or questions in this setup.This separates source representation and benchmark construction from LLM graph extraction.
  • Evaluation setup: The frozen pipeline name-matches seed entities, renders retrieved structure, caps context at B = 8,000 characters, and uses a fixed Qwen2.5-7B-Instruct reader.The renderer is tokenizer-independent because the cap is specified in characters.
  • Motivating question: The prior 0.48 versus 0.98 result motivates testing whether the native graph’s deficit is caused by bounded rendering rather than missing structure.Forty-five of fifty prior contexts reached the cap.

4. Budget-bounded retrieval on CGMES power-system graphs

The formal section models retrieval as ordering render units under a character budget and makes seed-local answer preservation and naive-rendering failure explicit. Its key guarantee is conditional: preservation holds when the seed-local render mass fits the budget.

  • Guarantee: Proposition 2 states that seed-anchored rendering is never worse than naive rendering in answer presence within the guaranteed regime.This supports replacing a deployed renderer without regression for covered queries.
  • Guarantee: The guarantee preserves predefined seed-local answer-bearing units whenever their total render mass satisfies D(S) ≤B.The formal results characterize answer preservation, naive failure, and the no-harm condition.
  • Formal setup: The formalization treats node descriptions and edge renderings as distinct units with deterministic lengths and identifiers.Each unit’s length includes its separator.
  • Formal setup: The retrieval graph uses a fixed character budget B, with the experiments using B=8,000.The graph is undirected and edges carry relation labels.
  • Candidate units: The candidate universe contains node descriptions and edge renderings whose endpoints lie within hop bound k of the seed set.The experiments use k = 2 and deterministic name matching defines the seeds.

Define the anchored tier

Seed-anchored rendering places the seed and incident edges before outward-hop units, so a budget-fitting seed-local evidence set enters the greedy context prefix before nonlocal content. The implemented renderer preserves this seed-local guarantee, while broader radius guarantees do not apply to its minimum-endpoint edge priority.

  • Budgeted prefix: The greedy context is the maximal order prefix whose cumulative render-unit length is at most budget B.Render-unit lengths are measured in characters of the rendered UTF-8 text, including fixed newline separators.
  • Anchored render order: The anchored order emits seed descriptions and seed-incident edges first, followed by un-emitted descriptions and edges grouped by outward hop.Identifiers break ties within each tier, and seed-incident edges are not emitted again in their ordinary hop tier.
  • Implementation scope: The evaluated implementation uses minimum endpoint distance for edge priority, so radius-1 units need not form a prefix beyond the seed-local tier.An edge with endpoint distances (1, 2) can share a priority tier with a (1, 1) edge; therefore no r-local preservation guarantee is claimed beyond seed-local evidence.
  • Preservation condition: If the seed-local render mass D(S) is at most B, all seed-local answer-bearing units enter the anchored budget prefix.The proof relies on placing every seed-local unit before every non-seed-local unit and on the answer set being contained in those units.
  • Checkability: The precondition can be checked before retrieval by summing seed and seed-incident edge lengths, with cost O(|S| · deg).The paper characterizes this condition as mild for the low-degree line and transformer query entities studied.
  • Dominance and failure mode: The guarantee is never worse than naive rendering for seed-local answers satisfying D(S) ≤B, and it can be strictly better when naive descriptions consume the budget first.For SmallGrid, the hub-flooding condition corresponds to naive S2 accuracy 0.12 and SmallGrid-3.0 accuracy 0.00.

5. Method and experimental design

The study evaluates deterministic render-ordering effects in budget-bounded retrieval, using fixed graphs, seeding, reader, scorer, and an 8,000-character cap. It tests CGMES encodings, graph and corpus arms, preregistered question strata, and confirmatory and exploratory statistical analyses.

  • Design: The study varies only render-unit ordering within a fixed budget-bounded retrieval pipeline, rather than changing the graph, reader, or scoring stages.The primary objective is retrieval and ordering effects, not general-purpose question answering.
  • Method: The guarantee preserves predefined seed-local answer-bearing render units when their seed-local render mass satisfies the checkable budget precondition.The method changes which material survives truncation when the precondition fails, rather than preventing overflow generally.
  • Method: Algorithm 1 takes a graph, query, budget B, and hop bound k, then deterministically emits seed-prioritized render units under greedy budget filling.It introduces no additional tuned or learned parameters, degree threshold, or LLM graph-construction tokens.
  • Networks: The evaluation covers six ENTSO-E CGMES configurations spanning CGMES 2.4.15 bus-branch and CGMES 3.0 node-breaker representations.The CGMES 2.4.15 trio supports development, while the re-encoded CGMES 3.0 trio is held out from retrieval-method development.
  • Arms: The experiment compares standards-native and LLM-extracted graph arms, three corpus baselines, and both naive and seed-anchored render orders under a common 8,000-character pipeline.Extracted graphs include LightRAG, Microsoft GraphRAG, and HippoRAG; all arms share the seeding rule, cap, reader, and scorer.
  • Questions and analysis: Each network contributes 50 pre-registered S1–S2 items plus 17–24 extended items across containment, aggregation, and global-superlative strata.The benchmark yields 129 extended items overall, with typed machine-checkable gold specifications and unique-name filtering.

6. Results on ENTSO-E CGMES test configurations

Across ENTSO-E CGMES configurations, seed-anchored rendering preserves answer-bearing evidence under the context cap and improves the preregistered fresh-bank endpoint over naive rendering. Comparisons with extracted GraphRAG graphs further show that the standards-native anchored arm can match or exceed them without LLM graph-construction tokens, within the evaluated scope.

  • 6.1. Mechanism: answer presence under the cap: 3/25 versus 25/25: on development SmallGrid multi-hop items, naive rendering preserved answer-bearing units three times, while seed-anchored rendering preserved them on all 25.Across the run set, naive contexts hit the cap on 32–50 of 50 items per network, compared with at most 2 of 50 for anchored rendering.
  • 6.1. Mechanism: answer presence under the cap: 0.12 and 0.00: naive rendering retained multi-hop evidence on SmallGrid and SmallGrid-3.0, respectively, whereas seed-anchored rendering retained all multi-hop evidence on both.Naive retained evidence for every single-hop item; the reader-side join indicates the multi-hop failure was truncation rather than reader error.
  • 6.2. Seed-anchored retrieval improves CGMES power-system QA under context budgets: 0.450 to 0.970: on the preregistered fresh 100-item bank across the two budget-binding encodings, seed-anchored rendering outperformed naive rendering.The paired difference was +0.520, with 95% CI [+0.415, +0.625] and one-sided McNemar exact p=2.2 × 10^-14.
  • 6.2. Seed-anchored retrieval improves CGMES power-system QA under context budgets: 0.12 to 0.96: SmallGrid multi-hop S2 accuracy increased with seed-anchored rendering, while SmallGrid-3.0 increased from 0.00 to 0.92.The exploratory per-network analysis found significant gains on MicroGrid, SmallGrid, and held-out SmallGrid-3.0, but other network contrasts were inconclusive.
  • 6.3. Against LLM-extracted GraphRAG: 0 tokens: the standards-native anchored arm never fell behind tested LLM-extracted graphs under either render order.At fixed anchored order, it reached 0.96 versus 0.80 on held-out SmallGrid-3.0 and 0.92 versus 0.82 on MicroGrid; the extracted-graph ordering also improved some extracted graphs under binding budgets.
  • 6.3. Against LLM-extracted GraphRAG: 0.40 and 0.60: Microsoft GraphRAG scored below LightRAG and the zero-token anchored method on SmallGrid and SmallGrid-3.0, respectively.Its build cost was 235,312 to 1,703,707 tokens, or 2.3× to 2.7× LightRAG’s cost.

6.4. Extractor-controlled comparison

The extractor-controlled comparison separates graph source from rendering order and shows that extractor choice materially affects the apparent advantage of LLM-built graphs. Under matched evaluation, the standards-native anchored graph remains competitive while using zero construction tokens.

  • Extractor control: 0.98 to 0.66: replacing the proprietary development extractor with the held-out open-weight extractor lowers GLLM accuracy while anchored remains 0.98.The comparison changes only the extractor, making the held-out advantage extractor-conditional rather than a network artifact.
  • Representation dependence: 0.48 versus 0.94: the raw CIM graph changes from collapse on bus-branch SmallGrid to survival on node-breaker SmallGrid-3.0, while the derived support graph collapses in the latter.The encoding changes how neighbourhoods are distributed, so the representation that overflows can flip between versions of the same network.
  • Held-out transfer: 0.62 overall on held-out SmallGrid-3.0: the proposed method leads a mostly non-binding human-authored probe, while the LLM-descriptor arm reaches 0.16.The comparison primarily tests fidelity transfer rather than the budget-dominance margin.

7. Practical guidance for utilities

For utilities using CIM/CGMES models, the paper recommends deterministic parsing and seed-anchored rendering rather than relying on LLM graph reconstruction. These recommendations are conditional on the evaluated reader, extractor, budget, and model scope.

  • Practical guidance: Consider parsing the standard directly: the standards-native anchored arm performed at least as well as all three tested extracted graph representations under the shared pipeline.The comparison concerns LightRAG, Microsoft GraphRAG, and HippoRAG with matched retrieval, reader, seeder, and scorer.
  • Practical guidance: Seed-anchored ordering is recommended for seed-local CGMES queries after verifying D(S) ≤B, seeding accuracy, and rendering completeness.The rule changes rendering order, adds no tuned parameters, and carries an answer-preservation guarantee within the stated regime.
  • Practical guidance: Keep the parsed CGMES model as the numeric source of truth because extracted graphs can migrate numeric attributes between similarly named components.The paper frames this as a fidelity safeguard for values, not merely a retrieval-order choice.
  • Practical guidance: Monitor the reader in production: retrieval-level guarantees are reader-independent, but absolute accuracy varied across the tested readers.Reader changes require re-validating accuracy levels even when the rendering ordering and no-harm guarantee remain applicable.
  • Practical guidance: Local deterministic parsing and retrieval avoid external API calls and provide reproducible, auditable retrieval within the operator’s trust boundary.The paper limits this operational interpretation to fidelity, deterministic retrieval, and data handling rather than full safety or regulatory compliance.

8. Limitations and threats to validity

The evidence is strongest for seed-local, budget-bounded retrieval on a narrow set of CIM/CGMES networks and readers. Several benchmark, extractor, reader, scale, and query-type limitations constrain broader QA or deployment claims.

  • Benchmark validity: 83% of S1/S2 gold answers appear verbatim in the corpus, giving lexical methods a surface-form advantage and making the benchmark closer to retrievability than general QA.Human-authored and derived-answer probes provide additional evidence, but broader field-authored evaluation remains future work.
  • Reader dependence: Absolute accuracy is reader-specific: anchored accuracy ranges from 0.98 to 0.64 on SmallGrid across three readers, although ordering is more stable.The confirmatory statistics use one frozen Qwen2.5-7B-Instruct reader.
  • Network diversity and scale: The six-network dominance claim covers only three base topologies, and RealGrid does not establish transfer to fully operational transmission models of 10^5 to 10^6 objects.The retrieval-level guarantee is network-size independent under its seed-local precondition, but reader accuracy at larger scales remains undemonstrated.
  • Baseline scope: The LLM-extraction comparison omits packaged native retrievers and other builders, so its conclusion is conditional on the tested framework set and shared pipeline.One native HippoRAG scheme was reimplemented, while the packaged native retrievers were not run.
  • Guarantee scope: The guarantee covers answer presence for seed-local answers only when D(S) ≤B; it does not establish reader extraction, end-to-end accuracy equivalence, or non-inferiority.Aggregation, global superlatives, and some broader query strata lie outside the formal guarantee.
  • Operational assumptions: A high-degree seed can violate the precondition when its incident render mass exceeds the budget, while an aliased or missed seed can prevent query-object seeding even if the precondition holds.These are concrete failure cases for deployment beyond the guaranteed regime.
  • Budget scope: The primary QA evaluation uses one fixed B = 8,000-character budget and does not evaluate accuracy-versus-budget curves, adaptive budgeting, or multi-turn budgeting.Formal results are budget-agnostic, but the reader-level curve and adaptive settings remain unevaluated.
  • Scoring validity: Scoring audits changed stored verdicts but left all six-network anchored accuracies and the preregistered fresh-bank endpoint unchanged.The frozen assertion-aware entity score is 0.701 over 3,503 rows.

9. Conclusion

The paper concludes that standards-derived CIM/CGMES graphs can be rendered directly for fixed-budget LLM retrieval instead of being reconstructed by an LLM. Seed-anchored ordering preserves budget-fitting seed-local evidence and improves the evaluated endpoint, while broader query planning and generalization remain open.

  • Conclusion: 0.12 and 0.00 versus all evidence retained: naive rendering loses multi-hop evidence on the two budget-binding networks, whereas seed-anchored rendering preserves it.The reader-free item-level analysis attributes the failure to truncation rather than reader error.
  • Conclusion: 0.450 to 0.970: the preregistered fresh 100-item SmallGrid bank confirms the retrieval mechanism at item level under the fixed budget.This is strong evidence within one topology family, not a network-level generalization claim.
  • Conclusion: At zero graph-construction tokens, the standards-native arm is never behind the three tested extracted representations under the common retrieval-and-rendering pipeline.The comparison is conditional on the tested extractors, reader, budget, and framework set.
  • Conclusion: The no-harm guarantee is an elementary greedy-prefix property: whenever seed-local render mass fits the budget, anchored rendering is never worse than naive for seed-local answer presence.The property concerns retrieval-level answer presence, not grid control or general end-to-end QA.
  • Open boundary: Aggregation and global superlative queries remain outside the evaluated system and would require query planning over the parsed CIM model.Suggested routes include generated SPARQL or network-analysis calls, but the hybrid end-to-end system was not evaluated.
  • Open boundary: RealGrid contributes only an attribute-lookup no-harm result because its multi-hop stratum and subgraph comparisons lacked unique gold referents.Therefore it does not establish topology-retrieval or extraction-parity conclusions on that independent topology.
  • Open boundary: Broader field-authored, multi-network, noisy-documentation, multi-turn, protection, contingency, and dynamic-study evaluations remain future work.Such tasks may require reasoning beyond the seed-local, k-local neighbourhood covered by the present guarantee.

Data availability

The complete research artifact is archived as a citable public deposit and companion code repository, with identifiers withheld during double-anonymized review.

  • The archived artifact includes parsers, retrievers, QA generators, audit trails, and result files.
  • The deposit contains a 100-question human-authored stress test and machine-checkable gold specifications.
  • The DOI and repository URL will be provided on acceptance.

Funding

The research did not receive any specific grant from funding agencies.

  • No specific grant supported the research.
  • The funding statement covers public funding agencies.
  • The funding statement covers commercial funding agencies.

CRediT authorship contribution statement

Jayakumar Manoharan contributed across the research lifecycle, while Yamini Sehgal contributed software and validation; both contributed to manuscript review and editing.

  • Jayakumar Manoharan contributed conceptualization, methodology, software, formal analysis, investigation, and data curation.
  • Jayakumar Manoharan contributed to original drafting, review, editing, and visualization.
  • Yamini Sehgal contributed software, validation, and manuscript review and editing.
Loading 2609.02011v1…