Pith. sign in

REVIEW 5 major objections 6 minor 48 references

A Graph-Retrieval-Augmented Generation Framework Enhances Decision-Making in the Circular Economy

T0 review · 5 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read CircuGraphRAG claims that routing LLM queries through a structured circular-economy knowledge graph—rather than through free text—lifts ROUGE-L F1 from below 0.08 to 1.0 and makes answers traceable.

desk verdict A sensible GraphRAG application for circular-economy queries, but the evaluation is self-referential and the paper's own tables contradict its headline claims. read the letter →

arxiv 2506.04252 v1 pith:YDID6UYB submitted 2025-06-01 cs.AI cs.CLcs.LG

Yang Zhao , Chengxiao Dai , Dusit Niyato , Chuan Fu Tan , Keyi Xiang , Yueyang Wang , Zhiquan Yeo , Daren Tan Zong Loong
show 2 more authors
Jonathan Low Zhaozhi Eugene H.Z. HO
This is my paper
classification cs.AIcs.CLcs.LG
keywords retrieval-augmentedgenerationknowledgegraphSPARQLcirculareconomyindustrialsymbiosishallucinationreductionGWP100emissionfactorswaste-to-resourcedecisionsupport
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

CircuGraphRAG claims that the way to stop LLMs from inventing industrial codes and emission factors is to remove free-text retrieval from the loop and route every question through a structured knowledge graph built for the circular economy. The system parses a natural-language query, selects and merges SPARQL templates from an 18-template library, retrieves a verified subgraph of 117,380 entities and 753,145 edges, ranks results by 100-year global warming potential, and only then asks the LLM to phrase an answer. On six benchmark queries, this raises ROUGE-L F1 from below 0.08 for standalone LLMs and naive RAG to 1.0 in the best configurations, and cuts response time and token use in representative tasks. A reader should care because a wrong waste-classification code or invented GWP100 figure can mean regulatory non-compliance or a misstated climate impact, so the paper's contribution is a reproducible architecture for grounding circular-economy decisions in auditable data rather than in the model's memory.

What carries the argument

The carrying object is a domain-specific RDF knowledge graph built from a waste-treatment and recycling dataset, harmonizing eight code systems (ISIC, NACE, WZ, CPA, EWC, HS, SSIC, CPC) into 117,380 nodes and 753,145 edges, with predicates such as hasEwcCode, hasHSCode, hasNaceCode, hasGwp100, hasProvider, and hasReceiver. The load-bearing mechanism is the combination of an 18-template SPARQL query library with LLM-driven template matching and query merging: the LLM is never asked to recall a code or number, only to choose and parameterize a query path, and the retrieved subgraph is ranked by GWP100 before the final answer is composed. This is what converts open-ended generation into a constrained lookup-and-explain process.

What would settle it

Give the system a query whose correct answer is deliberately missing from the knowledge graph—for example, a newly regulated waste code added after graph construction—and compare its answer to a standalone LLM and a domain expert; if the gap disappears or the system confidently returns an ungrounded fallback answer, the claim that grounding eliminates hallucination is falsified.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that knowledge-graph-grounded retrieval effectively eliminates the hallucination problem for structured industrial-symbiosis questions. CircuGraphRAG translates user questions into SPARQL queries through an LLM-driven template matcher, merges multiple templates into a single executable path, retrieves the exact subgraph rather than text chunks, ranks candidate waste-to-resource matches by GWP100, and then composes a concise answer from those retrieved facts. In the reported experiments this yields ROUGE-L F1 of 1.0 on several single-hop and multi-hop cases across two of the three LLM backends, while the same LLMs without graph grounding or with text-chunk retrieval score below 0.08. The design also makes each answer traceable: the template match, merged query, and retrieved subgraph are all shown, so a regulator or plant operator can audit how a code or emission value was obtained.

Load-bearing premise

The whole result rests on the benchmark being a true test of correctness, but every reference answer was checked against the same knowledge graph the system queries, so the gain is measured on questions the system's own database was built to answer.

Editorial extensions

If this is right

  • If the claim holds, waste-to-resource planning can be automated without sacrificing auditability: each recommendation carries the SPARQL path and graph facts behind it.
  • The reported gains are model-agnostic across the three evaluated LLM backends, so deployers can swap backends without rebuilding the retrieval layer, as long as the backend is capable enough to follow the template logic.
  • The approach reduces the token and latency cost of typical queries, because the LLM writes a short grounded answer instead of a long speculative one.
  • The same architecture should transfer to other regulated domains with standardized codes and numeric factors, by replacing the ontology and template library.
  • The multi-hop results suggest that industrial-symbiosis matching—finding who can take another facility's by-product—can be discovered systematically rather than by manual cross-referencing.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The benchmark is self-referential: every gold answer was tested against the same knowledge graph that CircuGraphRAG queries, so the headline 1.0 versus 0.08 gap measures retrieval fidelity inside the graph, not general question-answering ability; on questions whose answers are absent or mis-mapped in the graph, the advantage should shrink.
  • Because the LLM's only role is template selection, parameter filling, and answer phrasing, the real ceiling of the system is the 18-template library; a genuinely novel or vague query will fall back to ungrounded generation, as the fuzzy-template ablation already hints.
  • The traceability could be packaged as a compliance artifact: regulators and ESG auditors could require that every reported code or emission factor ship with its template and subgraph, turning the framework into an assurance tool rather than a search engine.
  • A natural stress test is to run the pipeline on a held-out set of real facility queries created after graph construction, or on deliberately corrupted code mappings, to see whether grounding still beats baselines when the graph is imperfect.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 6 minor

Summary. This paper presents CircuGraphRAG, a retrieval-augmented generation framework that grounds LLM answers in a purpose-built knowledge graph for circular-economy waste-to-resource queries. The knowledge graph integrates Ecoinvent-derived resources with EWC, HS, NACE, CPA, ISIC, SSIC, and WZ codes and GWP100 emission factors (117,380 nodes, 753,145 edges). Question answering proceeds by LLM-driven entity extraction, selection among 18 SPARQL templates, LLM-based merging of matched templates, execution against GraphDB, GWP100-ascending ranking, and final answer composition. The paper evaluates the system on three single-hop and three multi-hop queries with Llama-3-70B, Qwen-QwQ-32B, and DeepSeek-R1-Distill-Llama-70B, reporting ROUGE-L F1 up to 1.0 with baselines below 0.08, decreased hallucination, reduced token usage, and halved response times, together with a template ablation study and a multi-round consistency study.

Significance. If the empirical claims were validated, CircuGraphRAG would be a useful domain-specific GraphRAG contribution: the framework design is coherent, the three-stage operation is described in enough detail to be reproduced, Table 7 provides a transparent per-stage token-and-time breakdown, and the paper honestly reports many failure cases (e.g., the F1 = 0.0 configurations in Tables 3 and 4) and states its limitations in Sections 2.6 and 4.5.4. However, the central evaluation is self-referential: the gold answers are co-refined against the same knowledge graph the system queries (Section 4.3), so perfect ROUGE-L scores are not independent evidence of correctness; unvalidated embedding-inferred code mappings (Section 4.2) weaken the underlying data; and the efficiency claims are contradicted by the paper's own tables. The significance of the contribution as stated is therefore not established. With an independent, externally validated gold standard and a larger, released query set, the approach could be a solid applied-AI contribution; as submitted, the evidence is insufficient.

major comments (5)
  1. [§4.3 (Ground Truth Dataset Construction), Table 2] The evaluation is circular and cannot support the abstract's claim of fact-checked, regulatory-ready answers. Step 3 of Section 4.3 states that each annotated answer was 'tested against our knowledge graph to confirm that the declared relationships were unambiguously retrievable' and that 'any inconsistencies resulted in further refinements to the knowledge graph or the ground truth.' The gold answers in Table 2 are therefore projections of the same graph that CircuGraphRAG queries, so a faithful SPARQL lookup reproduces the gold answer by construction, and ROUGE-L = 1.0 measures template and retrieval fidelity rather than factual correctness; the Standalone LLM and Naive RAG baselines are scored against facts they cannot access. The gold standard should be built from independent sources (e.g., official EWC/HS descriptions and Eurostat NACE-CPA correspondence tables) without consulting the graph, and all six cases should be re-scored against it.
  2. [§2.1.3, §4.2, Table 2] The knowledge graph's correctness is not established, and the paper contains a concrete numeric inconsistency. Section 4.2 reports that incomplete crosswalks are completed by sentence-embedding inference and that the inferred codes are only 'flagged for domain experts to validate', but no validation results are reported anywhere, so errors in these mappings could propagate into both the graph and the gold answers. A concrete numeric inconsistency makes the problem visible: Section 2.1.3 says CircuGraphRAG with Qwen and Llama retrieves the minimal GWP100 value 0.008930631 for Case 3, while Table 2 gives the Case 3 ground truth as 0.008826959. A system cannot receive ROUGE-L = 1.0 against a reference that contains a different number, so either the narrative or the Table 2 answer is wrong; the discrepancy sits precisely in the numeric-accuracy regime the framework claims to guarantee. The plausibility of several mappings is itself doubtful: Table 2 Case 2 pairs a query about a provider 'producing waste cement' with the gold answer 'Municipal incineration of waste polyethylene', a mismatch that suggests the graph's inferred crosswalks need external audit.
  3. [Tables 3 and 4, §3.1.2, Appendix 6.8] The claimed 'superior performance' is not uniform even on the six in-sample queries. In Tables 3 and 4, CircuGraphRAG with Llama scores ROUGE-L F1 = 0.0 in Cases 1, 3, 4, and 6, and with DeepSeek in Cases 1, 3, 4, and 6 as well; only Qwen succeeds consistently, so the abstract's blanket comparison ('superior performance... while baseline scores below 0.08') overstates the results. The narrative in Section 3.1.2 that CircuGraphRAG 'achieves exact matches in such scenarios' is also contradicted by the paper's own showcase run in Appendix 6.8, where the Case 4 Llama run ends with 'The SPARQL query returned no results' and the model falls back to guessing 'CPA code 38', and the DeepSeek run guesses 'CPA code 37.20' — exactly the hallucinated code assignments the framework is claimed to eliminate. The tables are honest, but Sections 2.1.1, 3.1.1, and 3.1.2 should be rewritten to match them; note also that the qualitative showcase in Appendix 6.7 (waste polyurethane) is not among the six benchmark cases, so the qualitative and quantitative evaluations do not cover the same queries.
  4. [§2.3, Tables 5 and 6, Abstract] The efficiency claims are contradicted by the paper's own tables when total token counts (input plus output) are considered. The abstract promises 'halving the response time and reducing token usage by 16%', and Section 2.3 says CircuGraphRAG 'consistently reduces both total token usage and inference latency', but Table 5 Case 2 with Qwen shows CircuGraphRAG at 1,206 + 1,787 = 2,993 tokens and 4.45 s versus Naive RAG at 420 + 646 = 1,066 tokens and 1.48 s; Table 6 Case 4 with Qwen shows 6,041 tokens and 11.50 s versus 1,696 tokens and 2.89 s for Naive RAG; and the same pattern appears in Case 3 with DeepSeek (3,069 tokens and 6.24 s versus 1,827 tokens and 5.54 s). The 16% token reduction and halved latency hold only for isolated configurations, so the blanket efficiency claim in the abstract and Section 2.7 is not supported by the reported data.
  5. [§4.3, §2.4, §2.5] The empirical base is too thin to support the scope of the claims. The entire benchmark consists of six hand-curated queries in Table 2 with no significance testing, no variance reporting, and no error analysis over repeated runs; the template ablation in Table 1 and the multi-round consistency study in Section 2.5 are each conducted on a single representative case. Section 2.5 additionally reports 60% exact-match rates for Llama and Qwen in template matching and query merging, which directly undermines the reliability of the pipeline for regulatory use, a point the authors themselves make in the same section. Finally, the 'reproducible graph reasoning benchmark' announced in Section 1 cannot currently be reproduced: the Data/Codes availability statements defer release until after publication, and the full gold-answer strings are not specified.
minor comments (6)
  1. [Figure 1] The caption for panel (b) repeats 'single-hop QA tasks'; it should read 'multi-hop QA tasks'.
  2. [§4.7.3] The paper should state explicitly whether the <think> reasoning traces that appear in Appendix Tables 8 and 9 are stripped from the outputs before ROUGE-L scoring, since this materially affects every reported score.
  3. [§2.3, Tables 5-6] 'Token usage' should be defined as input, output, or total; the abstract's 16% figure cannot be located in the tables as they now stand.
  4. [§4.1] The relation between the 3,896 Ecoinvent entries and the 117,380 nodes and 753,145 edges of the knowledge graph should be explained, as the node count is not obviously derivable from the stated source.
  5. [§4.6.2] The exact model versions, the Groq API snapshot date, and the decoding parameters should be pinned so that the experiments can be reproduced; the current description (e.g., 'Llama3 (70B)') is under-specified.
  6. [§6.9] The abbreviation table lists RDFlib, which is not used in the reported pipeline, while the pipeline's embedding models (all-MiniLM-L6-v2 and multi-qa-mpnet-base-cos-v1) are omitted from the table.

Circularity Check

2 steps flagged · score 7.0 of 10

The benchmark is self-referential: Section 4.3 co-refines gold answers against the same knowledge graph that CircuGraphRAG queries, and Section 2.1.3 reports a perfect F1 for a value that does not match the Table 2 ground truth, showing the gold standard is effectively the KG value itself.

  1. self definitional [Section 4.3 (Ground Truth Dataset Construction), Cross-Checking]
    "Each annotated answer was tested against our knowledge graph to confirm that the declared relationships were unambiguously retrievable. Any inconsistencies resulted in further refinements to the knowledge graph or the ground truth."

    The evaluation gold answers are defined, at least in part, by what is unambiguously retrievable from the authors' knowledge graph. CircuGraphRAG answers by executing SPARQL queries against that same graph (Section 4.5.3: 'we execute the FinalQuery against the knowledge graph'). Therefore a correct template match reproduces the gold answer by construction; ROUGE-L measures retrieval/template fidelity against a KG-derived reference, not correctness against an independent external standard. The co-refinement loop, which can revise either the KG or the ground truth, removes any independent check on factual content. Section 4.2 also shows that many code crosswalks are only inferred by embeddings and 'flagged for domain experts to validate', with no validation results reported.

  2. other [Section 2.1.3 Numeric Retrieval, cf. Table 2 Case 3]
    "In Case 3, CircuGraphRAG with both Qwen and Llama correctly retrieves the minimal GWP100 value (0.008930631), achieving an ROUGE-L F1 score of 1.0."

    Table 2 lists the Case 3 ground-truth answer as 0.008826959, not 0.008930631. Since ROUGE-L is lexical and these strings differ, the reported F1=1.0 cannot be computed against the published ground truth. The only way to obtain perfect F1 is to score against the value stored in the knowledge graph itself, confirming that the gold standard is KG-relative rather than an independently established datum. This is the same self-referential evaluation identified in Section 4.3 and undermines the claim that the method provides 'fact-checked, regulatory-ready' numeric answers.

full rationale

The paper's central performance claim is that CircuGraphRAG grounds LLM outputs in a circular-economy knowledge graph and achieves ROUGE-L scores up to 1.0 while baselines fall below 0.08. That comparison is not circular in the sense of a fitted parameter being renamed a prediction, and there is no load-bearing self-citation chain: the GraphRAG and survey references are external works. The circularity lies in benchmark construction. Section 4.3 states that each annotated answer was tested against the same knowledge graph that CircuGraphRAG queries, and that inconsistencies led to refining either the knowledge graph or the ground truth. Consequently, the gold answers are not independent of the system's retrieval source; any system that correctly issues the corresponding SPARQL query will trivially reproduce the reference. Section 2.1.3 exacerbates this: the paper claims exact retrieval of GWP100 value 0.008930631 with perfect F1, while Table 2 gives the gold answer as 0.008826959, so the perfect score cannot have been computed against the published ground truth. The evaluation therefore supports a claim about template/SPARQL fidelity, not about factual correctness of the code mappings or emission factors. The absence of external validation for the embedding-inferred crosswalks (Section 4.2) and the limited six-query benchmark further weaken the transferability of the claimed gains. Because the main evidence for 'fact-checked, regulatory-ready support' reduces to agreement with a KG-derived reference, a partial-circularity score of 7 is appropriate.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The paper's contribution is an engineered retrieval stack. The central evaluation rests on a self-built knowledge graph and a self-constructed ground truth, so the key assumptions are the correctness of the Ecoinvent-derived mappings and the representativeness of the hand-curated queries. No new physical or mathematical entities are introduced.

free parameters (3)
  • top-k retrieval candidates = top-5 default; top-k unspecified
    The number of candidate entities retrieved from FAISS and the default top-5 result limit are set manually and are not varied or reported, so their effect on the scores is unknown.
  • LLM decoding parameters = temperature=0.7, top-p=0.9, max_tokens=4096
    Hyperparameters are shared across models and methods, but no sensitivity analysis is given.
  • Embedding models = multi-qa-mpnet-base-cos-v1, all-MiniLM-L6-v2
    Two off-the-shelf sentence encoders are chosen without comparison; they affect both crosswalk inference and entity retrieval.
assumptions (5)
  • domain assumption Ecoinvent 'Waste Treatment and Recycling' entries (3,896 records) provide accurate GWP100 values and process-resource relationships.
    Section 4.1 and 4.2: the knowledge graph is built from this dataset without independent verification of emission factors.
  • domain assumption Crosswalks between ISIC, CPC, NACE, WZ, CPA, EWC, HS, and SSIC codes used to harmonize the knowledge graph are correct and complete.
    Section 4.2: incomplete crosswalks are augmented by sentence-embedding inference, which can introduce errors; correctness is assumed.
  • ad hoc to paper The library of 18 SPARQL templates covers the space of user intentions for circular economy questions.
    Section 4.5.3: the template set is the core mechanism; no coverage analysis is provided, and vague prompts are acknowledged as a limitation in Section 2.6.
  • ad hoc to paper The six hand-curated queries and gold answers are representative and correct.
    Section 4.3: answers were cross-referenced against official guidelines and the paper's own knowledge graph, then refined until retrievable.
  • domain assumption LLM outputs via the Groq API are sufficiently deterministic for one-shot evaluation.
    Section 4.6: main results appear to be single-run; Section 2.5 shows template matching is only 60% exact-match across five runs for Llama and Qwen.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Graph-Retrieval-Augmented Generation Framework Enhances Decision-Making in the Circular Economy." pith.science (2026). https://pith.science/paper/YDID6UYB

@misc{pith2026250604252,
  author       = {Pith},
  title        = {Pith review of: A Graph-Retrieval-Augmented Generation Framework Enhances Decision-Making in the Circular Economy},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YDID6UYB}},
  note         = {Machine review of arXiv:2506.04252}
}
read the original abstract

Large language models (LLMs) hold promise for sustainable manufacturing, but often hallucinate industrial codes and emission factors, undermining regulatory and investment decisions. We introduce CircuGraphRAG, a retrieval-augmented generation (RAG) framework that grounds LLMs outputs in a domain-specific knowledge graph for the circular economy. This graph connects 117,380 industrial and waste entities with classification codes and GWP100 emission data, enabling structured multi-hop reasoning. Natural language queries are translated into SPARQL and verified subgraphs are retrieved to ensure accuracy and traceability. Compared with Standalone LLMs and Naive RAG, CircuGraphRAG achieves superior performance in single-hop and multi-hop question answering, with ROUGE-L F1 scores up to 1.0, while baseline scores below 0.08. It also improves efficiency, halving the response time and reducing token usage by 16% in representative tasks. CircuGraphRAG provides fact-checked, regulatory-ready support for circular economy planning, advancing reliable, low-carbon resource decision making.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

48 extracted references · 40 canonical work pages

  1. [1]

    Annual review of energy and the environment25(1), 313–337 (2000)

    Chertow, M.R.: Industrial symbiosis: literature and taxonomy. Annual review of energy and the environment25(1), 313–337 (2000)

    work page 2000

  2. [2]

    Journal of industrial ecology16(1), 28–37 (2012)

    Lombardi, D.R., Laybourn, P.: Redefining industrial symbiosis: Crossing academic–practitioner boundaries. Journal of industrial ecology16(1), 28–37 (2012)

    work page 2012

  3. [3]

    Business & Information Systems Engineering66(6), 831–848 (2024) https://doi.org/10.1007/s12599-024-00904-4

    Püchel, L., Wang, C., Buhmann, K., Brandt, T., Schweinitz, F., Edinger-Schons, L.M., Brocke, J., Legner, C., Teracino, E., Mardahl, T.D.: On the pivotal role of data in sustainability transformations. Business & Information Systems Engineering66(6), 831–848 (2024) https://doi.org/10.1007/s12599-024-00904-4

    work page doi:10.1007/s12599-024-00904-4 2024

  4. [4]

    Sustainability16(14) (2024) https://doi

    Solaimani, S.: From compliance to capability: On the role of data and technology in environment, social, and governance. Sustainability16(14) (2024) https://doi. org/10.3390/su16146061

    work page doi:10.3390/su16146061 2024

  5. [5]

    Nookala, G., Gade, K.R., Dulam, N., Thumburu, S.K.R., et al.: Governance for data ecosystems: Managing compliance, privacy, and interoperability (2024)

    work page 2024

  6. [6]

    Energy, Sustainability and Society11(1), 13 (2021) https://doi.org/10.1186/s13705-021-00289-2

    Kullmann, F., Markewitz, P., Stolten, D., Robinius, M.: Combining the worlds of energy systems and material flow analysis: a review. Energy, Sustainability and Society11(1), 13 (2021) https://doi.org/10.1186/s13705-021-00289-2

    work page doi:10.1186/s13705-021-00289-2 2021

  7. [7]

    247–262 (2016)

    Moriguchi, Y., Hashimoto, S.: Material Flow Analysis and Waste Management, pp. 247–262 (2016). https://doi.org/10.1007/978-3-319-20571-7_12

    work page doi:10.1007/978-3-319-20571-7_12 2016

  8. [8]

    Moraga, G., Huysveld, S., Mathieux, F., Blengini, G.A., Alaerts, L., Van Acker, K., de Meester, S., Dewulf, J.: Circular economy indicators: What do they measure? Resources, Conservation and Recycling146, 452–461 (2019) https: //doi.org/10.1016/j.resconrec.2019.03.045

    work page doi:10.1016/j.resconrec.2019.03.045 2019

Show all 48 references
  1. [9]

    Sustainable Production and Consumption31, 113–126 (2022) https://doi.org/10.1016/j.spc.2022.02.004

    Roy, T., Garza-Reyes, J.A., Kumar, V., Kumar, A., Agrawal, R.: Redesign- ing traditional linear supply chains into circular supply chains–a study into its challenges. Sustainable Production and Consumption31, 113–126 (2022) https://doi.org/10.1016/j.spc.2022.02.004

    2022 doi

  2. [10]

    The British Accounting Review55(1), 101149 (2023)

    Tsang, A., Frost, T., Cao, H.: Environmental, social, and governance (esg) disclosure: A literature review. The British Accounting Review55(1), 101149 (2023)

    2023

  3. [11]

    Corporate social responsibility and environmental management10(3), 153–164 (2003)

    Hedberg, C.-J., Von Malmborg, F.: The global reporting initiative and corporate sustainability reporting in swedish companies. Corporate social responsibility and environmental management10(3), 153–164 (2003)

    2003

  4. [12]

    In: World Scientific Encyclopedia of Climate Change: Case Studies of Climate Risk, Action, and 22 Opportunity Volume 3, pp

    Hales, J.: Sustainability accounting standards board (sasb). In: World Scientific Encyclopedia of Climate Change: Case Studies of Climate Risk, Action, and 22 Opportunity Volume 3, pp. 37–41. World Scientific, Singapore (2021)

    2021

  5. [13]

    Final Report: Recommendations of the Task Force on Climate-Related Financial Disclosures (2017)

    Board, F.S.: Task force on climate-related financial disclosures. Final Report: Recommendations of the Task Force on Climate-Related Financial Disclosures (2017)

    2017

  6. [14]

    Resources, conservation and recycling136, 473–485 (2018)

    Fraccascia, L., Yazan, D.M.: The role of online information-sharing platforms on the performance of industrial symbiosis networks. Resources, conservation and recycling136, 473–485 (2018)

    2018

  7. [15]

    Resources, conservation and recycling157, 104799 (2020)

    Fraccascia, L., Giannoccaro, I.: What, where, and how measuring industrial sym- biosis: A reasoned taxonomy of relevant indicators. Resources, conservation and recycling157, 104799 (2020)

    2020

  8. [16]

    Journal of environmental management277, 111430 (2021)

    Capelleveen,G.,Amrit,C.,Zijm,H.,Yazan,D.M.,Abdi,A.:Towardbuildingrec- ommender systems for the circular economy: Exploring the perils of the european waste catalogue. Journal of environmental management277, 111430 (2021)

    2021

  9. [17]

    arXiv preprint arXiv:2409.11524 (2024)

    Vidali, A., Jean, N., Pera, G.L.: Unlocking nace classification embeddings with openai for enhanced analysis and processing. arXiv preprint arXiv:2409.11524 (2024)

    2024 arXiv

  10. [18]

    ACM Computing Surveys (Csur)54(4), 1–37 (2021)

    Hogan, A., Blomqvist, E., Cochez, M., d’Amato, C., Melo, G.D., Gutierrez, C., Kirrane, S., Gayo, J.E.L., Navigli, R., Neumaier, S.,et al.: Knowledge graphs. ACM Computing Surveys (Csur)54(4), 1–37 (2021)

    2021

  11. [19]

    O’Reilly Media, Inc

    Powers, S.: Practical RDF: Solving Problems with the Resource Description Framework. " O’Reilly Media, Inc.", Sebastopol, CA (2003)

    2003

  12. [20]

    Lassila, O.: Resource description framework (rdf) model and syntax (1997)

    1997

  13. [21]

    Radford, A., Narasimhan, K., Salimans, T., Sutskever, I., et al.: Improving language understanding by generative pre-training (2018)

    2018

  14. [22]

    Environmental Science & Technology57(9), 3464–3466 (2023)

    Rillig, M.C., Ågerstrand, M., Bi, M., Gould, K.A., Sauerland, U.: Risks and benefits of large language models for the environment. Environmental Science & Technology57(9), 3464–3466 (2023)

    2023

  15. [23]

    Environmental Science & Technology57(46), 17667–17670 (2023)

    Zhu, J.-J., Jiang, J., Yang, M., Ren, Z.J.: Chatgpt and environmental research. Environmental Science & Technology57(46), 17667–17670 (2023)

    2023

  16. [24]

    ACM computing surveys55(12), 1–38 (2023)

    Ji, Z., Lee, N., Frieske, R., Yu, T., Su, D., Xu, Y., Ishii, E., Bang, Y.J., Madotto, A., Fung, P.: Survey of hallucination in natural language generation. ACM computing surveys55(12), 1–38 (2023)

    2023

  17. [25]

    Advances in Neural Information Processing 23 Systems33, 9459–9474 (2020)

    Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal, N., Küttler, H., Lewis, M., Yih, W.-t., Rocktäschel, T.,et al.: Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in Neural Information Processing 23 Systems33, 9459–9474 (2020)

    2020

  18. [27]

    arXiv preprint arXiv:2405.16506 (2024)

    Hu, Y., Lei, Z., Zhang, Z., Pan, B., Ling, C., Zhao, L.: Grag: Graph retrieval- augmented generation. arXiv preprint arXiv:2405.16506 (2024)

    2024 arXiv

  19. [28]

    https://arxiv.org/abs/2312.10997

    Gao, Y., Xiong, Y., Gao, X., Jia, K., Pan, J., Bi, Y., Dai, Y., Sun, J., Wang, M., Wang, H.: Retrieval-Augmented Generation for Large Language Models: A Survey (2024). https://arxiv.org/abs/2312.10997

    2024 arXiv

  20. [29]

    https://arxiv.org/abs/ 2005.11401

    Lewis,P.,Perez,E.,Piktus,A.,Petroni,F.,Karpukhin,V.,Goyal,N.,Küttler,H., Lewis,M.,Yih,W.-t.,Rocktäschel,T.,Riedel,S.,Kiela,D.:Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks (2021). https://arxiv.org/abs/ 2005.11401

    2021 arXiv

  21. [30]

    https: //arxiv.org/abs/2305.18846

    Kang, M., Kwak, J.M., Baek, J., Hwang, S.J.: Knowledge Graph-Augmented Language Models for Knowledge-Grounded Dialogue Generation (2023). https: //arxiv.org/abs/2305.18846

    2023 arXiv

  22. [31]

    Climatic Change52(1), 93–127 (2002)

    Godal, O., Fuglestvedt, J.: Testing 100-year global warming potentials: impacts on compliance costs and abatement profile. Climatic Change52(1), 93–127 (2002)

    2002

  23. [32]

    Accessed: 2025-05-18 (2023)

    United States Environmental Protection Agency: Understanding Global Warm- ing Potentials | US EPA. Accessed: 2025-05-18 (2023). https://www.epa.gov/ ghgemissions/understanding-global-warming-potentials

    2023

  24. [33]

    https://arxiv.org/abs/2501.00309

    Han, H., Wang, Y., Shomer, H., Guo, K., Ding, J., Lei, Y., Halappanavar, M., Rossi, R.A., Mukherjee, S., Tang, X., He, Q., Hua, Z., Long, B., Zhao, T., Shah, N., Javari, A., Xia, Y., Tang, J.: Retrieval-Augmented Generation with Graphs (GraphRAG) (2025). https://arxiv.org/abs/...

    2025 arXiv

  25. [34]

    https://arxiv.org/abs/2408

    Peng, B., Zhu, Y., Liu, Y., Bo, X., Shi, H., Hong, C., Zhang, Y., Tang, S.: Graph Retrieval-Augmented Generation: A Survey (2024). https://arxiv.org/abs/2408. 08921

    2024

  26. [35]

    Journal of Cleaner Production13(13), 1337–1343 (2005) https://doi.org/10.1016/j.jclepro.2005.05.002

    Frischknecht, R., Rebitzer, G.: The ecoinvent database system: a comprehensive web-based lca database. Journal of Cleaner Production13(13), 1337–1343 (2005) https://doi.org/10.1016/j.jclepro.2005.05.002 . Life Cycle Assessment

    2005 doi

  27. [36]

    https://ai.meta.com/blog/meta-llama-3/

    Meta AI: Introducing Meta Llama 3: The most capable openly available LLM. https://ai.meta.com/blog/meta-llama-3/. Accessed: 2025-05-18 (2024)

    2024

  28. [37]

    https: 24 //qwenlm.github.io/blog/qwq-32b/

    Qwen Team: QwQ-32B: Embracing the Power of Reinforcement Learning. https: 24 //qwenlm.github.io/blog/qwq-32b/. Accessed: 2025-05-18 (2024)

    2024

  29. [38]

    Pellets of munici- pal waste

    DeepSeek AI: DeepSeek-R1-Distill-Llama-70B. https://huggingface.co/ deepseek-ai/DeepSeek-R1-Distill-Llama-70B. Accessed: 2025-05-18 (2025) 25 6 Appendix 6.1 Query–Answer Pairs from the Ground Truth. Table 2: Query–Answer Pairs from the Ground Truth. Case Type Query Answer 1 Si...

    2025

  30. [39]

    query": 17,

    Querying with Retrieved Context: construction and demolition wastes, industrial✓ Qwen 1. Template Matching: <think>...</think> - Possible Solutions: [{"query": 17, "input": "waste polyurethane", "output": "category information"}]

  31. [40]

    Query Merging: <think>...</think> 17

  32. [41]

    query": 17,

    Querying with Retrieved Context: <think>...</think> construction and demolition wastes✓ (continued on next page) 36 Method LLM Response (continued) Deepseek 1. Template Matching: <think>...</think> [{"query": 17, "input": "waste polyurethane", "output": "category of solid waste"}]

  33. [42]

    waste polyurethane,

    Query Merging: <think>...</think> Since the question directly asks for the category of solid waste for "waste polyurethane," and query 17 is specif- ically designed to retrieve information based on the entity’s name, the solution is to execute query 17 with the input "waste po...

  34. [43]

    Renting of air transport equipment

    Querying with Retrieved Context: <think>...</think> The category of solid waste that waste polyurethane belongs to isconstruction and demolition wastes✓, and it can also be classified as industrial waste. Note: For brevity, the<think>sections are partially omitted in this tabl...

  35. [44]

    Query Merging: Merging method:Query 10⇒Query 5

  36. [45]

    Querying with Retrieved Context: The SPARQL query returned no results. The providers under NACE code 3821 are likely related to waste treatment, so the resources gen- erated might be categorized underCPA code 38 (Waste collection, treatment and disposal activities; materials r...

  37. [46]

    Query Merging: <think>...</think> Merging Strategy:Chaining Execution Flow:Query 1⇒Query 7⇒Query 6

  38. [47]

    query": 18,

    Querying with Retrieved Context: <think>...</think> The SPARQL query returned the code cpa: 382150, hence the CPA code that categorizes the resources generated by providers under NACE 3821 is382150✓. (continued on next page) 41 Method LLM Response (continued) Deepseek 1. Templ...

  39. [48]

    Merging method:Query 18⇒Query 5

    Query Merging: <think>...</think> The merging method is chaining because the output of query 18 (resource codes) is used as the input for query 5 to get the CPA codes. Merging method:Query 18⇒Query 5

  40. [49]

    Querying with Retrieved Context: <think>...</think> The SPARQL query returned no results. Based on general knowledge, resources under NACE code 3821 (waste collection, treatment, and disposal activities) are typically categorized under CPA code 37.20 (Waste collection, treatme...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.