ChatNDC: Climate Policy RAG System

• ChatNDC is a domain-specific retrieval-augmented generation system that provides climate-policy analysis using verified documents and layered LLM calls.
• It processes queries through classification, vector-based retrieval with page-level provenance, and GPT-4o-Mini generation to balance answer quality and energy use.
• Comparative studies indicate that its multiple LLM calls for verification enhance factual reliability but also result in higher energy consumption than streamlined alternatives.

Searching arXiv for the cited paper and any directly relevant ChatNDC references. ChatNDC is a climate-policy retrieval-augmented generation system examined as a domain-specific LLM product in research on inference-time energy use and output quality. In the reported architecture, a user’s natural-language question is processed through query classification, document retrieval over verified climate documents, response generation, and hallucination checking, with final output assembled only from verified sentences carrying page-level references (Bao et al., 31 Mar 2026). The system is studied comparatively against ChatNetZero and direct GPT-4o-Mini usage, with emphasis on how architectural choices such as multiple LLM calls, verification design, and output length shape both energy consumption and response characteristics (Bao et al., 31 Mar 2026).

1. System role and architectural definition

ChatNDC is presented as one of two LLM-based climate analysis chatbots used to study the energy footprint of domain-specific RAG workflows relative to generic model usage (Bao et al., 31 Mar 2026). Its function is climate-domain analysis grounded in verified climate documents, and its workflow is explicitly decomposed into retrieval, generation, and hallucination-checking components (Bao et al., 31 Mar 2026).

The pipeline begins with query classification (GPT-4o). The user’s natural-language question is first passed to a GPT-4o model that determines which sub-corpora are relevant and tags the query with thematic labels. The listed examples of sub-corpora are UNFCCC submissions, IPCC reports, national policy assessments (Bao et al., 31 Mar 2026). This stage therefore performs both routing and semantic categorization within a climate-policy evidence base.

The second stage is document retrieval (CPU-bound). Based on the classification output, the system issues a vector-search query over a precomputed embedding index of verified climate documents. Retrieved chunks carry provenance metadata, specifically document and page number (Bao et al., 31 Mar 2026). This provenance design is central to later citation attachment and verification.

The third stage is response generation (GPT-4o-Mini). Retrieved evidence is concatenated with a database-aware prompt template, and a GPT-4o-Mini call generates a draft answer under a constrained prompt with temperature = 0 and max length 400–600 words (Bao et al., 31 Mar 2026). The use of a constrained prompt indicates an attempt to standardize output behavior within a bounded answer length.

The fourth stage is hallucination checking (GPT-4o-Mini). Each sentence of the draft is re-embedded and sent, along with its source citations, to a second GPT-4o-Mini invocation. This “anti-hallucination” agent confirms or rejects each claim and attaches fine-grained, auditable citations (Bao et al., 31 Mar 2026). The final assembly stage then includes only sentences verified by the hallucination-checking step, and per-sentence references are linked to the original pages (Bao et al., 31 Mar 2026).

This design differs from the comparison systems in specific ways. ChatNetZero uses a single GPT-4o-Mini generation call and a non-LLM cosine-similarity check, whereas GPT-4o-Mini baselines have no retrieval or explicit verification calls (Bao et al., 31 Mar 2026). A plausible implication is that ChatNDC embodies a more agentic RAG design, with stronger procedural grounding but higher orchestration overhead.

2. Retrieval, generation, and verification workflow

ChatNDC’s workflow is organized around an explicit separation of evidence selection, answer drafting, and post hoc claim validation. This is not merely an implementation detail; in the reported study, it is the basis for decomposing per-query energy use and for analyzing design trade-offs (Bao et al., 31 Mar 2026).

In the classification stage, GPT-4o identifies relevant sub-corpora and thematic labels. This means that retrieval is not performed over a monolithic corpus in an undifferentiated fashion; instead, the retrieval scope is narrowed using an upstream LLM routing decision (Bao et al., 31 Mar 2026). Because the system targets climate-policy analysis, the sub-corpus selection mechanism functions as a domain-aware control layer over the evidence store.

Retrieval itself is vector search over a precomputed embedding index of verified climate documents. The retrieved chunks preserve provenance metadata including document and page number (Bao et al., 31 Mar 2026). This metadata is carried forward into answer construction and later into per-sentence references. The emphasis on verified documents and page-level provenance indicates that grounding is implemented not only through semantic retrieval but also through auditable source linkage.

During generation, retrieved evidence is combined with a database-aware prompt template and processed by GPT-4o-Mini under a constrained decoding configuration. The prompt temperature is fixed at zero, and the target output length is capped within 400–600 words (Bao et al., 31 Mar 2026). These settings suggest an operational preference for determinism and bounded verbosity.

The hallucination-checking stage is distinctive. Each sentence of the draft is re-embedded and evaluated with a second GPT-4o-Mini call together with source citations. The “anti-hallucination” agent either confirms or rejects each claim and attaches fine-grained citations (Bao et al., 31 Mar 2026). Only verified sentences survive into the final output. This suggests a sentence-level filtration regime rather than whole-answer acceptance or rejection.

A plausible implication is that ChatNDC implements a two-pass LLM architecture in which the first pass optimizes for synthesis from retrieved evidence and the second pass optimizes for claim-level validation. The study, however, also shows that this additional structure has a measurable energy cost (Bao et al., 31 Mar 2026).

3. Energy-estimation methodology

The study estimates ChatNDC’s inference-time energy by decomposing total per-query energy into retrieval, generation, and verification components using a refined API-based framework from Jegham et al. (2025) (Bao et al., 31 Mar 2026). The governing equation is:

$$E_{\text{total}} = E_{\text{retrieval}} + E_{\text{inference}} + E_{\text{verification}}$$

Here, $E_{\text{retrieval}}$ estimates CPU-core energy for semantic search in Colab, $E_{\text{inference}}$ covers the GPT-4o and GPT-4o-Mini generation calls, and $E_{\text{verification}}$ covers the extra GPT-4o-Mini hallucination check (Bao et al., 31 Mar 2026).

For retrieval-stage energy, the reported formula is:

$$E_{\rm retrieval} \;=\; T_{\rm retrieval}\;(\mathrm{h}) \;\times\; P_{\rm CPU}\;(\mathrm{kW}) \;\times\; \mathrm{PUE}_{\rm Google} \;=\; T_{\rm retrieval}\;(\mathrm{h}) \;\times\; 0.0085\;\mathrm{kW} \;\times\; 1.09 \;=\; T_{\rm retrieval}\;(\mathrm{h})\;\times\;0.00927\;\mathrm{kW}$$

For LLM generation and verification, the study defines $T_{\rm LLM}$ as elapsed API-call time in hours and introduces a node power draw term:

$$W = U_{\rm GPU}\times P_{\rm GPU} + U_{\rm nonGPU}\times P_{\rm nonGPU}$$

after multiplying by PUE (Bao et al., 31 Mar 2026). For GPT-4o-Mini, the energy estimate is:

$E_{\rm GPT4o\mbox{-}Mini} \;=\; T_{\rm LLM}\;(\mathrm{h}) \;\times\;0.343125\;\mathrm{kW} \;\times\;1.12 \;=\; T_{\rm LLM}\;(\mathrm{h}) \;\times\;0.3843\;\mathrm{kW}$

and similarly for GPT-4o (large) (Bao et al., 31 Mar 2026).

Within ChatNDC, the study states that $E_{\text{inference}}$ = sum of all GPT-4o-Mini generation calls (here, the second stage), and $E_{\text{verification}}$ = one GPT-4o-Mini call (third stage), both at $E_{\text{retrieval}}$0 kW $E_{\text{retrieval}}$1 call duration (Bao et al., 31 Mar 2026). This decomposition enables component-level attribution rather than only end-to-end energy reporting.

The experimental protocol ran 102 climate-policy queries $E_{\text{retrieval}}$2 4 time windows (morning, afternoon, evening, random) from the Netherlands, converting execution times into energy via the stated formulas (Bao et al., 31 Mar 2026). Geographic variation beyond a single origin was not tested, and cross-location effects remain an open question (Bao et al., 31 Mar 2026).

4. Measured energy footprint

The reported quantitative results provide median and mean per-query energy for ChatNDC over 408 runs. The values are presented both in kilowatt-hours and in kilojoules, using $E_{\text{retrieval}}$3 (Bao et al., 31 Mar 2026).

Component	Median (kWh)	Mean (kWh)
Retrieval	0.00010	0.00011
Generation (GPT-4o-Mini)	0.00280	0.00295
Verification (GPT-4o-Mini)	0.00163	0.00168
Total	0.00453	0.00474

Expressed in kilojoules, the median values are 0.36 kJ for retrieval, 10.08 kJ for generation, 5.87 kJ for verification, and 16.31 kJ in total; the mean total is 17.06 kJ (Bao et al., 31 Mar 2026). The distribution indicates that retrieval contributes a small fraction of total energy relative to the LLM stages, while verification is itself a substantial component rather than a negligible post-processing step.

Time-of-day variation was reported as small, specifically less than 5% (Bao et al., 31 Mar 2026). The median per-query values were 0.00445 kWh (16.0 kJ) in the morning, 0.00451 kWh (16.2 kJ) in the afternoon, 0.00460 kWh (16.6 kJ) in the evening, and 0.00453 kWh (16.31 kJ) for the random run (Bao et al., 31 Mar 2026). The study does not report multi-region deployment behavior; it explicitly notes that geographic variation was not tested beyond the Netherlands (Bao et al., 31 Mar 2026).

These measurements support the paper’s broader conclusion that the energy consumption of domain-specific RAG systems depends strongly on system design, particularly the number and type of LLM calls added for routing and verification (Bao et al., 31 Mar 2026).

5. Comparative positioning against ChatNetZero and GPT-4o-Mini

The comparative analysis reports median energy use for ChatNDC, unconstrained GPT-4o-Mini, GPT-4o-Mini with responses capped at $E_{\text{retrieval}}$4 words, and ChatNetZero (Bao et al., 31 Mar 2026).

System	Median (kWh/query)	Relative to ChatNDC
ChatNDC	0.00453	100 %
GPT-4o-Mini (unconstrained)	0.00113	25 %
GPT-4o-Mini ($E_{\text{retrieval}}$5 words)	0.00050	11 %
ChatNetZero (CNZ)	0.000408	9 %

The study states that ChatNDC uses approximately 4$E_{\text{retrieval}}$6 more energy than unconstrained GPT-4o-Mini, approximately 9$E_{\text{retrieval}}$7 more than GPT-4o-Mini(200), and approximately 11$E_{\text{retrieval}}$8 more than ChatNetZero (Bao et al., 31 Mar 2026). A one-way ANOVA test across the four conditions yields $E_{\text{retrieval}}$9, indicating statistically significant differences in median energy use (Bao et al., 31 Mar 2026).

The paper attributes part of this gap to architectural differences. ChatNDC includes extra GPT-4o-Mini calls for classification and hallucination check, which account for approximately 36% of total energy, with 30.9% from verification alone (Bao et al., 31 Mar 2026). By contrast, ChatNetZero places verification in a lightweight cosine-similarity post-processing stage, accounting for 8.4% of its total energy (Bao et al., 31 Mar 2026). This establishes a concrete design contrast between LLM-mediated verification and non-LLM verification.

A plausible implication is that ChatNDC occupies a point in the design space where stronger procedural checking and richer grounding flexibility are traded for materially higher inference-time energy cost. The study explicitly frames this as a question of design trade-offs rather than a simple ranking of systems (Bao et al., 31 Mar 2026).

6. Quality trade-offs, interpretation, and open questions

The quality analysis reports that factual accuracy (statement-level) is broadly similar across ChatNDC, CNZ, and GPT-4o-Mini, within less than $E_{\text{inference}}$0 on a 0–1 index (Bao et al., 31 Mar 2026). At the same time, embellishment (non-factual) is higher for ChatNDC (0.62) than for CNZ (0.15), and the paper states that this suggests extra length or style conventions drive energy cost without proportional factual gains (Bao et al., 31 Mar 2026).

The study therefore concludes that agentic pipelines with multiple LLM calls can improve grounding flexibility but substantially increase energy use for only marginal improvements in factual performance (Bao et al., 31 Mar 2026). This conclusion is tightly linked to the system’s architecture: classification, retrieval, constrained generation, sentence-level hallucination checking, and final filtering all contribute to a more elaborate workflow whose added cost does not map linearly onto measured factual gains.

Several design recommendations follow directly from the reported findings. Energy should be treated as a first-class metric alongside latency and accuracy when designing domain-specific RAG products (Bao et al., 31 Mar 2026). Output length matters: even with identical models, longer or more elaborate answers drive up energy use and often increase embellishment, so constraining answer length is described as a simple lever to reduce both energy and non-factual content (Bao et al., 31 Mar 2026). The paper also points to dynamic routing and cascades, naming FrugalGPT and RouteLLM, as mechanisms for directing simple factual queries to smaller models or non-LLM modules while reserving heavyweight pipelines for more complex analytic tasks (Bao et al., 31 Mar 2026). It further notes that sparsely-activated MoE architectures may decouple capacity from energy use (Bao et al., 31 Mar 2026).

The open questions are stated explicitly: multi-round dialogue footprint; cross-regional carbon intensity; end-to-end session-level energy; real-time “agentic” budget allocation (Bao et al., 31 Mar 2026). For ChatNDC specifically, these unresolved issues delimit the current scope of evidence. This suggests that the reported measurements characterize a single-query inference setting rather than the full operational footprint of sustained interactive use.

In that sense, ChatNDC is significant not only as a climate-policy RAG system, but also as a concrete case study in how domain-specific LLM product design affects measurable energy use, verification behavior, and output form (Bao et al., 31 Mar 2026).