ChatNetZero: Climate Commitment Analysis

• ChatNetZero is a specialized chatbot designed to assess and demystify corporate climate commitments using structured net-zero pledge data.
• It employs a streamlined retrieval-augmented generation workflow with a cosine-similarity filter to ensure strong data provenance and minimal hallucination.
• The system demonstrates superior energy efficiency with the lowest query energy among comparable models while delivering concise, verifiable responses.

Searching arXiv for the specified paper to ground the article and citation. ChatNetZero (CNZ) is a climate-specific chatbot, accessible at chatnetzero.ai, designed to demystify and assess net-zero and climate pledges by companies, countries and other actors. It supports climate and environmental analysis around corporate climate commitments, net-zero targets, emissions targets and timelines, and quantitative tasks such as counting entities with particular target types. In the available characterization of the system, CNZ is presented as a domain-specific retrieval-augmented generation workflow whose central design goal is to reduce hallucination and enforce strong provenance by constraining the LLM to verified, curated climate data, especially from the Net Zero Tracker (Bao et al., 31 Mar 2026).

1. Scope and domain focus

CNZ is oriented toward actor-level and pledge-level climate analysis rather than general conversational assistance. Its stated purpose is to assess net-zero and climate pledges and to answer questions such as whether a given company has made climate pledges, whether a net-zero target is credible, or how many entities satisfy a structured target condition. The system is therefore specialized for queries about corporate climate commitments, net-zero targets, emissions targets and timelines, and related structured variables (Bao et al., 31 Mar 2026).

A defining characteristic of CNZ is that it is not described as a general-purpose assistant with unconstrained synthesis. Instead, it is framed as a grounded climate-domain product that attempts to answer from curated sources and to acknowledge uncertainty if support is absent. This positions it within a class of domain-specific RAG systems in which the product objective is not only answer generation, but also traceability and manual verifiability.

The comparison class used in the same study clarifies this specialization. ChatNDC is also a climate-domain RAG chatbot, but it is focused on national climate policy and nationally determined contributions lodged under the Paris Agreement, whereas CNZ is focused on corporate and actor net-zero targets and is heavily centered on Net Zero Tracker and related climate-commitment data. This distinction is functional as well as corpus-level: the tasks CNZ emphasizes are narrower, more actor-specific, and more tightly aligned to structured target variables.

2. Knowledge base and retrieval regime

CNZ is grounded in a specialized corpus that includes the Net Zero Tracker database and reports as the primary source for actor-specific queries, together with other heterogeneous climate-related documents such as PDFs and spreadsheets. These materials are pre-processed by conversion to plain text, segmentation into paragraph-preserving chunks, and embedding of each chunk as a vector representation. Each chunk is tagged with document ID, page number, and location metadata, enabling semantic retrieval with provenance (Bao et al., 31 Mar 2026).

This retrieval design supports two query classes. For actor-specific questions, CNZ prioritizes entities in the Net Zero Tracker and searches the corresponding corpus partitions. For generic questions, such as definitional questions about emissions scopes, it retrieves top-ranked report chunks across the general corpus. In both cases, retrieval is vector-based semantic search over the pre-embedded chunks.

The workflow is explicitly described as a “simple RAG.” For actor-specific queries, the system retrieves a bounded set of top-ranked chunks per actor. For non-actor queries, it retrieves top-ranked report chunks across the general corpus. The bounded retrieval strategy is significant because it constrains the context window presented to the model and thereby limits downstream generation scope.

CNZ also includes what the study terms structured-data narration. Structured Net Zero Tracker variables, including target type, base year, and coverage, can be transformed into natural language, allowing quantitative and counting queries to be answered using the same retrieval-plus-grounding pipeline. This unifies narrative and structured evidence under a common response mechanism.

3. Response generation, verification, and citation

The retrieved chunks and the user question are passed to a self-fine-tuned GPT-4o-mini model. The prompting regime is tightly constrained: temperature is set to $0$, the model is explicitly instructed to answer only from the provided text, it must acknowledge uncertainty if the answer is not supported, and the target answer length is around $\leq 150$ words. CNZ therefore uses one single LLM call per query and does not employ multi-step agentic reasoning (Bao et al., 31 Mar 2026).

The anti-hallucination module is not implemented as an additional LLM call. Instead, the draft answer is split into sentences; each sentence is embedded and compared to the embeddings of the retrieved chunks using pairwise cosine similarity. If the similarity score is below a threshold, the sentence is deleted from the final output as potentially hallucinated or unsupported. This is implemented as a CPU-level vector operation.

A subsequent reference module links each remaining sentence to citations comprising document title, page number, and matched passage location. The resulting output is intended to preserve traceability and to enable users to manually verify claims. In effect, CNZ combines semantic retrieval, constrained generation, lightweight post hoc filtering, and explicit citation attachment in a single path.

The contrast with ChatNDC is architecturally instructive. Both systems are grounded climate-domain RAG chatbots using GPT-4o-mini as the main generation model and both include anti-hallucination and reference mechanisms. However, ChatNDC performs LLM-based query classification using GPT-4o, then retrieval, then GPT-4o-mini answer generation, and then an additional GPT-4o-mini call for sentence-level hallucination checking and citation verification. CNZ, by contrast, uses one retrieval step, one GPT-4o-mini generation call, and a non-LLM cosine-similarity hallucination filter. The study therefore treats CNZ as a simple, non-agentic pipeline and ChatNDC as a more agentic multi-call pipeline.

4. Inference-time energy model

The energy analysis decomposes total workflow cost as

$$E_{\text{total}} = E_{\text{retrieval}} + E_{\text{inference}} + E_{\text{hallucination}}.$$

For CNZ, retrieval includes semantic search and non-LLM post-processing, including the cosine-similarity hallucination check. The system therefore has one GPT-4o-mini call contributing to $E_{\text{inference}}$, while $E_{\text{hallucination}}$ is effectively folded into the CPU-bound retrieval and post-processing component (Bao et al., 31 Mar 2026).

The study adapts an end-to-end API-based LLM energy model in which query energy is estimated from decoding time, latency, power draw, utilization fractions, and power usage effectiveness. The general form is

$$E_{\text{query}} \, (\text{kWh}) = \left( \frac{\text{Output Length}}{\text{TPS}} + \frac{\text{Latency}}{3600} \right) \left( P_{\text{GPU}} \times U_{\text{GPU}} + P_{\text{non-GPU}} \times U_{\text{non-GPU}} \right) \cdot \text{PUE}.$$

For GPT-4o-mini, the study uses pre-computed assumptions yielding an effective factor of $0.3843\ \text{kW}$, so the inference energy for CNZ’s single model call is expressed as

$$E_{\mathrm{GPT\text{-}4o\text{-}mini}\,(\mathrm{kWh})} = T_{\mathrm{LLM}\,(\mathrm{h})} \cdot 0.3843\,(\mathrm{kW}).$$

Retrieval and cosine-similarity hallucination checking run in a Google Colab CPU-only environment on a single core of Intel Xeon @ 2.20 GHz, with power per core estimated at $8.5\ \text{W}$ and Google data-center PUE set to $1.09$. Retrieval energy is therefore modeled as

$\leq 150$0

The experimental setup fixes temperature at $\leq 150$1 for all OpenAI API calls and logs total execution time for each step, together with token usage measured by tiktoken configured for GPT-4o-mini. The focus is inference-time energy; training and embodied energy are explicitly out of scope.

5. Comparative performance, energy, and quality

The core empirical comparison places CNZ alongside ChatNDC, unconstrained GPT-4o-mini, and GPT-4o-mini with a 200-word output cap. The reported median energy per successful query and the reported embellishment statistics are as follows (Bao et al., 31 Mar 2026).

System	Median energy per query	Embellishment
ChatNetZero	$\leq 150$2 kWh	mean $\leq 150$3, SD $\leq 150$4
GPT-4o-mini	$\leq 150$5 kWh	mean $\leq 150$6, SD $\leq 150$7
GPT-4o-mini (200-word cap)	$\leq 150$8 kWh	mean $\leq 150$9, SD $$E_{\text{total}} = E_{\text{retrieval}} + E_{\text{inference}} + E_{\text{hallucination}}.$$0
ChatNDC	$$E_{\text{total}} = E_{\text{retrieval}} + E_{\text{inference}} + E_{\text{hallucination}}.$$1 kWh	mean $$E_{\text{total}} = E_{\text{retrieval}} + E_{\text{inference}} + E_{\text{hallucination}}.$$2, SD $$E_{\text{total}} = E_{\text{retrieval}} + E_{\text{inference}} + E_{\text{hallucination}}.$$3

CNZ has the lowest mean and median energy per query among the four pipelines. ChatNDC is about $$E_{\text{total}} = E_{\text{retrieval}} + E_{\text{inference}} + E_{\text{hallucination}}.$$4 more energy-intensive per query than CNZ, and about $$E_{\text{total}} = E_{\text{retrieval}} + E_{\text{inference}} + E_{\text{hallucination}}.$$5 more than the unconstrained GPT-4o-mini baseline. GPT-4o-mini with a 200-word cap is only slightly above CNZ, at $$E_{\text{total}} = E_{\text{retrieval}} + E_{\text{inference}} + E_{\text{hallucination}}.$$6 versus $$E_{\text{total}} = E_{\text{retrieval}} + E_{\text{inference}} + E_{\text{hallucination}}.$$7 kWh, indicating that strict length control can approach CNZ’s energy efficiency even without RAG.

CNZ also shows the smallest variation, with mean absolute deviation $$E_{\text{total}} = E_{\text{retrieval}} + E_{\text{inference}} + E_{\text{hallucination}}.$$8 kWh, which the study associates with its highly constrained and consistent output length and pipeline structure. Step-wise decomposition indicates that inference dominates total energy in CNZ, as in all systems, while retrieval plus cosine checking contributes only a small fraction. The hallucination step contributes only $$E_{\text{total}} = E_{\text{retrieval}} + E_{\text{inference}} + E_{\text{hallucination}}.$$9 of CNZ’s total energy because it is not an extra LLM call. By contrast, hallucination checking accounts for $E_{\text{inference}}$0 of ChatNDC’s total energy, and ChatNDC’s hallucination step uses about $E_{\text{inference}}$1 more energy than CNZ’s in absolute terms.

The quality analysis uses a statement-level Factual Index and Embellishment Index. Factual scores are described as broadly similar across CNZ, GPT-4o-mini, GPT-4o-mini (200), and ChatNDC, with no model clearly dominating in factual correctness. CNZ, however, has clearly the lowest embellishment. The study further reports no clear link between higher energy and better factuality, while higher energy tends to correlate with more embellishment, especially for ChatNDC and unconstrained GPT-4o-mini. Output length is central to this pattern: CNZ, GPT-4o-mini, and GPT-4o-mini (200) show a strong correlation between output token count and energy use, whereas ChatNDC shows a weaker correlation because a larger share of its energy is consumed by additional LLM calls rather than only by generating the main answer.

A common misconception addressed by these results is that RAG systems are inherently more efficient than generic GPT usage. The reported comparison does not support that proposition. The study instead states that efficiency depends strongly on how RAG is architected, particularly the number of LLM calls, answer length, and verification approach.

6. Temporal variation, limitations, and design implications

The experiments include four runs over the evaluation dataset: one at a random time and three in fixed windows of 08:00–10:30, 14:00–16:30, and 20:00–22:30. All runs originate from IPs in the Netherlands. These time-of-day measurements are used to capture load-dependent timing variation and to stabilize average latency and energy estimates. However, the energy model does not modulate energy or carbon by grid carbon intensity at different times or locations. Instead, it uses fixed PUE values of $E_{\text{inference}}$2 for Microsoft data centers and $E_{\text{inference}}$3 for Google data centers. Geographic variation is therefore not explored beyond Netherlands-origin traffic, and time variation is not converted into time-dependent carbon footprints (Bao et al., 31 Mar 2026).

Several limitations are explicit. The energy estimates are based on end-to-end latency rather than direct server-side measurements, so hardware allocations, batching behavior, and data-center details are inferred from secondary sources. The evaluation covers one-shot questions rather than multi-turn conversations. Both CNZ and ChatNDC are tied to the GPT model family, limiting cross-provider generalizability. The query set contains $E_{\text{inference}}$4 domain-specific questions, with $E_{\text{inference}}$5–$E_{\text{inference}}$6 per model manually evaluated. The study also does not integrate time- and region-dependent grid carbon intensity into $E_{\text{inference}}$7 estimates.

The design implications extracted from CNZ are correspondingly specific. Pipeline design is presented as being as important as model choice, since CNZ and ChatNDC differ by more than $E_{\text{inference}}$8 in energy per query despite sharing GPT-4o-mini as the main generation model. Keeping answer length proportional to user needs reduces energy consumption and embellishment without harming factual accuracy in this evaluation. Lightweight verification, such as CNZ’s cosine-similarity filter, can eliminate many unsupported statements with minimal energy overhead. More agentic or multi-step verification can substantially increase energy without proportional quality gains.

Future directions proposed for CNZ-like systems include session-level analysis of multi-turn conversations, dynamic workflows that decide when to invoke expensive hallucination checks, movement from fixed to adaptive compute budgets per query, benchmarking across models and providers, and integration of carbon-intensity-aware scheduling. A plausible implication is that CNZ functions as an existence proof for a particular design point in climate-domain LLM products: a smaller model, strict output-length control, RAG grounding on high-quality climate corpora, and static lightweight verification can jointly deliver low energy use, low embellishment, and comparable factuality within the scope of the reported evaluation.