Pith. sign in

REVIEW 3 major objections 7 minor 85 references

Introducing the Swiss Food Knowledge Graph: AI for Context-Aware Nutrition Recommendation

T0 review · 3 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper claims a unified Swiss Food Knowledge Graph can be populated by small off-the-shelf LLMs and support graph-grounded nutrition question answering at 80 percent accuracy.

desk verdict SwissFKG is a real integration effort and a first LLM-enrichment benchmark, but its quantitative claims rest on a small, unreleased, author-built ground truth, so treat them as proof-of-concept. read the letter →

arxiv 2507.10156 v1 pith:QJ7WJC5E submitted 2025-07-14 cs.AI

Lubnaa Abdur Rahman , Ioannis Papathanail , Stavroula Mougiakakou This is my paper
classification cs.AI
keywords SwissFoodKnowledgeGraphpersonalizednutritionlargelanguagemodelsGraph-RAGallergensdietaryrestrictionsingredientsubstitution
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

SwissFKG aims to solve a gap the authors identify in automatic dietary assessment: images alone cannot tell a system whether a meal fits a user's allergies, restrictions, or culture. The paper's claim is that this missing context can be captured in one knowledge graph—recipes, ingredients and substitutes, nutrient values, allergens, dietary restrictions, and Swiss food-pyramid categories—and that off-the-shelf open-weight LLMs under 70 billion parameters can populate such a graph with usable quality. It reports per-task benchmark scores for four LLMs and three embedding models, then shows a Graph-RAG layer over the graph answering 50 curated nutrition questions with up to 80 percent accuracy. If the claim holds, this is a foundation for context-aware dietary advice that goes beyond food recognition and a national stage toward a global food knowledge graph.

What carries the argument

The machinery is the SwissFKG ontology plus its LLM enrichment pipeline. The ontology defines node types—Recipe, Ingredient, Instruction, DietRestriction, Season, Cuisine, Utensil, AllergenCategory, SwissFoodPyramidCategory, and CompositeSubstitute—with typed edges such as CONTAINS, IS_SUITABLE_FOR, ALLERGEN_OF, CLASSIFIED_AS, SUBSTITUTED_BY, and HAS. The enrichment pipeline uses a designated LLM per task (translation, text splitting, allergen/diet/SFP labeling) together with embedding-based cosine similarity for ingredient matching, and the Graph-RAG stage embeds graph triplets, retrieves the ten most similar pieces of knowledge to a user query at a 0.5 similarity cutoff, and lets an LLM synthesize an answer. This combination carries the argument: the graph supplies verifiable structure, and retrieval decides how much of that structure reaches the answering model.

What would settle it

Have independent Swiss dietitians label allergens, dietary restrictions, and food-pyramid categories for a fresh sample of recipes and ingredients drawn from the same Swiss sources, then compare their labels to the graph's stored labels; if agreement is materially below the paper's F1 scores, the claim that off-the-shelf LLMs can effectively enrich the graph is refuted. For the QA claim, replace the 50 graph-generated questions with dietitian-authored questions from outside the graph and require exact or human-rated answers instead of containment.

Watch

Extended reading notes

Core claim

The central claim, on the authors' terms, is that a comprehensive national food knowledge graph is constructible and LLM-populatable: SwissFKG comprises 5,896 nodes and 62,499 relations across 1,000 recipes and 2,548 ingredients, connecting each recipe to ingredients, instructions, utensils, cuisines, seasons, and diet restrictions, and each ingredient to nutrients, allergen categories, Swiss pyramid categories, and substitutes. The authors present what they describe as the first benchmark of four off-the-shelf sub-70B LLMs for food knowledge augmentation; no single model dominates, with Gemma3 (27B) best at translation and text splitting, Mistral Small 3.2 (24B) best at allergen and diet-restriction mapping, Phi-4 (14B) best at Swiss-pyramid mapping, and Mxbai Embed Large best at ingredient matching. They then show that Graph-RAG over the graph can retrieve and answer user questions, reaching 80 percent accuracy with Gemma3 paired with Mxbai Embed Large, and that swapping embedding models changes accuracy by up to 16 percentage points. The paper's claim is that these results demonstrate LLMs can effectively enrich the graph with relevant nutritional information.

Load-bearing premise

The results all depend on the author-curated ground truth being correct and representative: 10 percent of the data for enrichment scores, and 50 expected answers generated from the graph itself for question-answering, so any bias in those labels would inflate or distort the reported accuracy.

Editorial extensions

If this is right

  • A recipe's allergen and dietary-restriction profile can be derived from its ingredients, so a single graph can answer 'is this meal safe for me?' questions without separate manual tagging per dish.
  • Existing image-based dietary assessment systems can be extended with recipe, substitution, and cultural context, which the authors identify as necessary for safe guidance such as insulin-dosing decisions.
  • LLM-based enrichment is viable for food knowledge graphs, but the best model is task-dependent, so a pragmatic pipeline should route tasks to different models rather than rely on one model.
  • Retrieval quality in graph question-answering is strongly affected by the embedding model, so embedding selection deserves the same attention as LLM selection.
  • The SwissFKG is a national building block that can be combined with other national food graphs into a larger unified food knowledge graph.

Reading between the lines

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

  • A stricter evaluation than the paper reports would use dietitian-authored questions blind to the graph, since the current 50 expected answers come from the same graph being queried and the accuracy metric counts a response as correct whenever it contains the expected answer.
  • The reported F1 scores rest on a 10-percent author-curated subset whose selection and disagreement resolution are not described; an external expert audit on a fresh sample would test whether the scores generalize.
  • The same enrichment blueprint could be applied to other national food composition databases, and comparing accuracy across countries would reveal whether the observed LLM quality is specific to Swiss food data or general.
  • A human-in-the-loop validation stage for low-confidence LLM labels, which the authors mention as future work, may be necessary before the graph is used in clinical or insulin-dosing settings.
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

3 major / 7 minor

Summary. The paper introduces SwissFKG, a knowledge graph that integrates 1,000 Swiss recipes, ingredient-level nutrient data, substitutions, allergen categories, dietary restrictions, and Swiss Food Pyramid guidelines. It proposes a semi-automated LLM-based enrichment pipeline and evaluates four open-weight LLMs under 70B parameters on translation, ingredient text splitting, allergen/SFP/DR mapping, and ingredient matching. It then builds a Graph-RAG question-answering system over the graph and reports up to 80% QA accuracy over 50 questions. The central claims are that SwissFKG is the first resource uniting these Swiss-specific food dimensions in one graph and that off-the-shelf LLMs can effectively enrich it.

Significance. If the results hold, SwissFKG would be a useful integrated resource for context-aware nutrition recommendation, and the LLM enrichment pipeline would be a practical template for populating food knowledge graphs. The paper has strengths: it uses a clearly specified ontology, evaluates four recent open models under fixed generation settings, reports deterministic inference settings for reproducibility, and its Section 5 limitations candidly acknowledge that the chosen metrics do not capture hallucination, semantic drift, human alignment, or schema consistency. The novelty claim, however, is only as strong as the evaluation: all quantitative evidence rests on a small, unreleased, author-constructed ground truth, and the QA benchmark is generated from the same graph the system queries. Independent validation is needed before the effectiveness claims can be accepted.

major comments (3)
  1. [§3.5, Tables 1–4 and 7] The quantitative support for the central claim that off-the-shelf LLMs can effectively enrich SwissFKG rests entirely on a 'curated expected output considered as ground truth' covering 10% of the data (100 recipes, 200 ingredients), but the paper does not state how this subset was sampled, how the expected labels were produced, or how annotator disagreements were resolved, and the ground-truth file is not released. Because the same labels are used as targets for the similarity, F1, and accuracy scores in Tables 1–4 and 7, a non-representative or biased gold standard would invalidate all reported performance numbers. The authors' own Section 5 limitation that the metrics overlook hallucination rates, semantic drift, human alignment, and structural inconsistencies makes this validation gap concrete; the paper needs either a released, independently annotated gold standard or an external benchmark before the enrichment claim can be accepted.
  2. [§3.4, Eq. (12)] The Graph-RAG QA evaluation uses 50 expected answers generated from the same KG that the system queries, and Eq. (12) scores a response as correct whenever the expected answer is a subset of the response. This means a verbose answer that includes the expected text but also hallucinated or incorrect extra content is counted as correct, so the reported 0.80 accuracy does not measure factual correctness; moreover, no error bars or statistical tests accompany any of the small-sample results. I recommend a stricter metric (e.g., precision-oriented or human-annotated correctness), an external QA set not derived from the graph, and reporting of retrieval-failure cases separately.
  3. [§4.1–§4.2] The best-performing models are selected per task on the same 10% evaluation set and then used to construct the final graph, whose statistics in Tables 5–6 are presented as the outcome. With differences between models as small as 0.005 in SFP F1 (Table 3) and 0.02 in matching accuracy (Table 4), the per-task selection is not robust without confidence intervals or a held-out validation split; the construction choices should be justified with uncertainty estimates or a separate validation set.
minor comments (7)
  1. [§3.1, paragraph 2] 'we also had French versions which we to evaluate translations' is missing a verb; it should read 'which we used to evaluate translations'.
  2. [§4.2] The sentence 'the top 5 most used ingredients were salt, pepper, olive oil, pepper, water, and sugar' lists 'pepper' twice and names six items; please correct the list and the count.
  3. [§3.5] The experiment settings state top-p=0, which is not a standard sampling parameter choice; unless this is intentional, please clarify whether this should be a conventional value or a different sampling configuration.
  4. [§3.2.3, Eq. (5)] The similarity metric is described as based on [53], but the paper does not explain how recursive longest-common-substring matching handles ties or empty strings; a brief clarification would aid reproducibility.
  5. [§3.4] The retrieval settings (cosine cutoff 0.5 and keeping the 10 most relevant items) are asserted without sensitivity analysis; a sentence explaining these choices or showing their robustness would strengthen the Graph-RAG results.
  6. [Tables 1–3] The tables report no standard deviations or sample sizes per task; adding these would help readers judge whether the 'best model' differences are meaningful.
  7. [Abstract and §1] The 'first resource ... to ever unite' claim is strong; a precise comparison against FoodKG [21], Recipe1M+ [38], and the Indian FKG [19] would help situate SwissFKG's novelty.

Circularity Check

0 steps flagged · score 0.0 of 10

No derivation-equivalence circularity identified; the main weakness is an in-house evaluation set, not a circular argument.

full rationale

The paper's central claim is that off-the-shelf LLMs can enrich SwissFKG, and this is supported by comparing LLM outputs against a manually curated reference set: Section 3.5 states 'a curated expected output considered as ground truth covering 10% of the data: 100 recipes and 200 ingredients'. The scoring equations (COMET in Eq. (4), string similarity in Eq. (5), F1 in Eqs. (6)-(9), cosine retrieval accuracy in Eq. (11)) are evaluation metrics, not fitted relationships: no model parameter is learned from the reference labels, and no reported number is algebraically equal to its input by construction. The Graph-RAG evaluation in Eq. (12) uses substring containment of an expected answer in the model response, and the 50 expected answers are 'generated from the KG' that the system queries; this makes the QA benchmark in-sample and limits external grounding, but the response is not defined as the expected answer and retrieval can fail (accuracies range from 0.64 to 0.80), so correctness does not reduce to the input by construction. The author self-citations (e.g., refs. [3,47,49,79]) concern prior dietary-assessment systems and are contextual background, not load-bearing premises. No circular step meeting the quoted-evidence bar was found; the legitimate concerns are evaluation-transparency and benchmark-independence issues, not circularity.

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

The central contribution is a data resource, not a derivation; there are no fitted mathematical parameters. The main load-bearing assumptions are the authority of external food databases, the validity of Foodsubs substitution data, and the correctness of self-constructed ground truths. The cosine cutoff and retrieval limit are hand-chosen and affect results.

free parameters (3)
  • Cosine similarity retrieval cutoff = 0.5
    Hand-chosen threshold for retrieving graph content in Graph-RAG (Section 3.4); the authors note some correct matches fall below it, affecting retrieval reliability.
  • Number of retrieved knowledge items = 10
    Fixed top-10 retrieval limit to stay within context window (Section 3.4); arbitrary and not tuned against performance.
  • Evaluation split = 10% (100 recipes, 200 ingredients)
    Arbitrary 10% sample used for ground truth evaluation (Section 3.5); no justification for representativeness.
assumptions (5)
  • domain assumption Swiss FCDB and USDA FoodData Central provide accurate nutrient values for ingredients.
    Used as authoritative nutrient sources without validating against independent measurements (Section 3.1).
  • domain assumption Foodsubs substitution options are valid and appropriate for Swiss recipes.
    Substitutions are imported from Foodsubs without evaluating whether they match Swiss culinary practice (Section 3.1).
  • domain assumption The WHO-based 18 dietary restrictions and Swiss 14 allergen categories are the right set for the KG.
    These taxonomies are fixed a priori and shape all LLM mapping tasks (Section 3.2.2).
  • domain assumption The manually constructed ground truth for enrichment and QA is correct.
    All evaluation metrics depend on this assumption; construction and validation process is not detailed (Section 3.5).
  • domain assumption LLM outputs used for the final graph are accurate enough after model selection.
    The final KG is built automatically from the best-scoring models without full manual verification (Section 4.2).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Introducing the Swiss Food Knowledge Graph: AI for Context-Aware Nutrition Recommendation." pith.science (2026). https://pith.science/paper/QJ7WJC5E

@misc{pith2026250710156,
  author       = {Pith},
  title        = {Pith review of: Introducing the Swiss Food Knowledge Graph: AI for Context-Aware Nutrition Recommendation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QJ7WJC5E}},
  note         = {Machine review of arXiv:2507.10156}
}
read the original abstract

AI has driven significant progress in the nutrition field, especially through multimedia-based automatic dietary assessment. However, existing automatic dietary assessment systems often overlook critical non-visual factors, such as recipe-specific ingredient substitutions that can significantly alter nutritional content, and rarely account for individual dietary needs, including allergies, restrictions, cultural practices, and personal preferences. In Switzerland, while food-related information is available, it remains fragmented, and no centralized repository currently integrates all relevant nutrition-related aspects within a Swiss context. To bridge this divide, we introduce the Swiss Food Knowledge Graph (SwissFKG), the first resource, to our best knowledge, to unite recipes, ingredients, and their substitutions with nutrient data, dietary restrictions, allergen information, and national nutrition guidelines under one graph. We establish a LLM-powered enrichment pipeline for populating the graph, whereby we further present the first benchmark of four off-the-shelf (<70 B parameter) LLMs for food knowledge augmentation. Our results demonstrate that LLMs can effectively enrich the graph with relevant nutritional information. Our SwissFKG goes beyond recipe recommendations by offering ingredient-level information such as allergen and dietary restriction information, and guidance aligned with nutritional guidelines. Moreover, we implement a Graph-RAG application to showcase how the SwissFKG's rich natural-language data structure can help LLM answer user-specific nutrition queries, and we evaluate LLM-embedding pairings by comparing user-query responses against predefined expected answers. As such, our work lays the foundation for the next generation of dietary assessment tools that blend visual, contextual, and cultural dimensions of eating.

Figures

Figures reproduced from arXiv: 2507.10156 by the authors.

Figure 1
Figure 1. Knowledge Graph Nodes, Node Properties, and Relationships [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Performance of LLMs on enrichment tasks. COMET [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

85 extracted references · 63 canonical work pages

  1. [1]

    FoodSubs: Ingredient Substitutions and Ingredient Synonyms

    2025. FoodSubs: Ingredient Substitutions and Ingredient Synonyms. https: //www.foodsubs.com/. Accessed: 2025-07-09

    2025

  2. [2]

    Mahyar Abbasian, Iman Azimi, Amir M Rahmani, and Ramesh Jain. 2023. Con- versational health agents: A personalized llm-powered agent framework. arXiv preprint arXiv:2310.02374 (2023)

    arXiv 2023

  3. [3]

    Lubnaa Abdur Rahman, Ioannis Papathanail, Lorenzo Brigato, Elias K Spanakis, and Stavroula Mougiakakou. 2024. Food Recognition and Nutritional Apps. In Diabetes Digital Health, Telehealth, and Artificial Intelligence . Elsevier, 73–83

    2024

  4. [4]

    aha! Swiss Allergy Centre. 2024. Food allergy. https://www.aha.ch/swiss-allergy- centre/allergies-intolerances/food-allergies/food-allergy. Last update: 13 August 2024

    2024

  5. [5]

    Rohan Ajwani, Shashidhar Reddy Javaji, Frank Rudzicz, and Zining Zhu. 2024. LLM-generated black-box explanations can be adversarially helpful. arXiv preprint arXiv:2405.06800 (2024)

    arXiv 2024

  6. [6]

    Grace Ataguba and Rita Orji. 2025. Exploring Large Language Models for Per- sonalized Recipe Generation and Weight-Loss Management. ACM Transactions on Computing for Healthcare (2025)

    2025

  7. [7]

    Yejin Bang, Ziwei Ji, Alan Schelten, Anthony Hartshorn, Tara Fowler, Cheng Zhang, Nicola Cancedda, and Pascale Fung. 2025. Hallulens: Llm hallucination benchmark. arXiv preprint arXiv:2504.17550 (2025)

    arXiv 2025

  8. [8]

    Charalampos Chelmis and Bedirhan Gergin. 2021. A Knowledge Graph for Semantic-Driven Healthiness Evaluation of Recipes

    2021

Show all 85 references
  1. [9]

    Amélie Cordier, Valmi Dufour-Lussier, Jean Lieber, Emmanuel Nauer, Fadi Badra, Julien Cojan, Emmanuelle Gaillard, Laura Infante-Blanco, Pascal Molli, Amedeo Napoli, et al . 2014. Taaable: a case-based system for personalized cooking. Successful Case-based Reasoning Application...

    2014

  2. [10]

    Murillo Edson de Carvalho Souza, Murillo Edson De Carvalho Souza, and Li Weigang. 2025. Unveiling the Black Box: The Significance of XAI in Making LLMs Transparent. Authorea Preprints (2025)

    2025

  3. [11]

    Damion M Dooley, Emma J Griffiths, Gurinder S Gosal, Pier L Buttigieg, Robert Hoehndorf, Matthew C Lange, Lynn M Schriml, Fiona SL Brinkman, and William WL Hsiao. 2018. FoodOn: a harmonized food ontology to increase global food traceability, quality control and data integratio...

    2018

  4. [12]

    Takumi Ege and Keiji Yanai. 2017. Estimating food calories for multiple-dish food photos. In 2017 4th IAPR Asian Conference on Pattern Recognition (ACPR) . IEEE, 646–651

    2017

  5. [13]

    Federal Department of Home Affairs (FDHA), Switzerland. 2016. Verordnung des EDI vom 16. Dezember 2016 betreffend Information über Lebensmittel (OIDAl) – Annex 6. https://www.fedlex.admin.ch/eli/cc/2017/158/fr#annex_6. SR817.022.16; accessed June 26 2025

    2016

  6. [14]

    Federal Office of Public Health (BLV). 2025. Schweizer Ernährungsempfehlun- gen. https://www.blv.admin.ch/blv/en/home/lebensmittel-und- ernaehrung/ernaehrung/empfehlungen-informationen/schweizer- ernaehrungsempfehlungen.html. Accessed June 26, 2025

    2025

  7. [15]

    Hua Feng, Xiujuan Xiong, Zhuo Chen, Qunying Xu, Zhongwei Zhang, Nan Luo, and Yongning Wu. 2023. Prevalence and influencing factors of food allergy in global context: a meta-analysis. , 320–352 pages

    2023

  8. [16]

    FoodStruct Nutrition & Medical Content Team. 2025. FoodStruct: Encyclopedia of Food & Nutrition. https://foodstruct.com/. Accessed: 2025-07-09

    2025

  9. [17]

    Google DeepMind. 2025. Gemma3: A family of lightweight, state-of-the-art open multimodal models with 128K context window, supporting 140+ languages and function calling. https://deepmind.google/models/gemma/gemma-3/. Released March 12, 2025

    2025

  10. [18]

    Qingyu Guo, Fuzhen Zhuang, Chuan Qin, Hengshu Zhu, Xing Xie, Hui Xiong, and Qing He. 2020. A survey on knowledge graph-based recommender systems. IEEE Transactions on Knowledge and Data Engineering 34, 8 (2020), 3549–3568

    2020

  11. [19]

    Saransh Kumar Gupta, Lipika Dey, Partha Pratim Das, and Ramesh Jain

  12. [20]

    Haoyu Han, Yu Wang, Harry Shomer, Kai Guo, Jiayuan Ding, Yongjia Lei, Ma- hantesh Halappanavar, Ryan A Rossi, Subhabrata Mukherjee, Xianfeng Tang, et al. 2024. Retrieval-augmented generation with graphs (graphrag). arXiv preprint arXiv:2501.00309 (2024)

    2024 arXiv

  13. [21]

    Steven Haussmann, Oshani Seneviratne, Yu Chen, Yarden Ne’eman, James Codella, Ching-Hua Chen, Deborah L McGuinness, and Mohammed J Zaki

  14. [22]

    Aidan Hogan, Eva Blomqvist, Michael Cochez, Claudia d’Amato, Gerard De Melo, Claudio Gutierrez, Sabrina Kirrane, José Emilio Labra Gayo, Roberto Navigli, Sebastian Neumaier, et al. 2021. Knowledge graphs. ACM Computing Surveys (Csur) 54, 4 (2021), 1–37

    2021

  15. [23]

    Hsin-Ling Hsu and Jengnan Tzeng. 2025. DAT: Dynamic alpha tuning for hybrid retrieval in retrieval-augmented generation. arXiv preprint arXiv:2503.23013 (2025)

    2025 arXiv

  16. [24]

    Yujia Hu, Tuan-Phong Nguyen, Shrestha Ghosh, and Simon Razniewski

  17. [25]

    Shih-Hong Huang, Ya-Fang Lin, Zeyu He, Chieh-Yang Huang, and Ting-Hao Ken- neth Huang. 2024. How does conversation length impact User’s satisfaction? A case study of length-controlled conversations with LLM-powered Chatbots. In Extended Abstracts of the CHI Conference on Human...

    2024

  18. [26]

    International Diabetes Federation. 2025. Global Diabetes Data & Insights. https: //diabetesatlas.org/data-by-location/global/ Data by location: Global; accessed 7 July 2025

    2025

  19. [27]

    Ponrawin Kansaksiri, Pongpipat Panomkhet, and Natthanet Tantisuwichwong

  20. [28]

    Furukawa

    Yuki Kataoka, Sachiko Yamamoto-Kataoka, Ryuhei So, and Toshi A. Furukawa

  21. [29]

    Vamsi Krishna Kommineni, Birgitta König-Ries, and Sheeba Samuel. 2024. From human experts to machines: An LLM supported approach to ontology and knowl- edge graph construction. arXiv:2403.08345 [cs.CL] https://arxiv.org/abs/2403. 08345

    2024 arXiv

  22. [30]

    Agnieszka Ławrynowicz, Anna Wróblewska, Weronika T Adrian, Bartosz Kul- czyński, and Anna Gramza-Michałowska. 2022. Food recipe ingredient substitu- tion ontology design pattern. Sensors 22, 3 (2022), 1095

    2022

  23. [31]

    Zhenfeng Lei, Anwar Ul Haq, Adnan Zeb, Md Suzauddola, and Defu Zhang. 2021. Is the suggested food your desired?: Multi-modal recipe recommendation with demand-based knowledge graph. Expert Systems with Applications 186 (2021), 115708

    2021

  24. [32]

    Xianming Li and Jing Li. 2023. AnglE-optimized Text Embeddings.arXiv preprint arXiv:2309. 71 (2023)

    2023

  25. [33]

    2023), 536–538

    Beyond the Pass Mark: Accuracy of ChatGPT and Bing in the National Medical Licensure Examination in Japan.JMA Journal 6, 4 (16 Oct. 2023), 536–538. https://doi.org/10.31662/jmaj.2023-0043

    2023 doi

  26. [34]

    Qidong Liu, Xiangyu Zhao, Yuhao Wang, Yejing Wang, Zijian Zhang, Yuqi Sun, Xiang Li, Maolin Wang, Pengyue Jia, Chong Chen, et al. 2024. Large language model enhanced recommender systems: Taxonomy, trend, application and future. arXiv preprint arXiv:2412.13432 (2024)

    2024 arXiv

  27. [35]

    Ya Lu, Thomai Stathopoulou, Maria F Vasiloglou, Lillian F Pinault, Colleen Kiley, Elias K Spanakis, and Stavroula Mougiakakou. 2020. goFOODTM: an artificial intelligence system for dietary assessment. Sensors 20, 15 (2020), 4283

    2020

  28. [36]

    Jinge Ma, Xiaoyan Zhang, Gautham Vinod, Siddeshwar Raghavan, Jiangpeng He, and Fengqing Zhu. 2024. MFP3D: Monocular Food Portion Estimation Leveraging 3D Point Clouds. arXiv preprint arXiv:2411.10492 (2024)

    2024

  29. [37]

    Priyanka Mandikal and Raymond Mooney. 2024. Sparse Meets Dense: A Hybrid Approach to Enhance Scientific Document Retrieval. arXiv:2401.04055 [cs.IR] https://arxiv.org/abs/2401.04055

    2024 arXiv

  30. [38]

    Bingchen Liu and Xin Li. 2025. Large Language Models for Knowledge Graph Embedding Techniques, Methods, and Challenges: A Survey. arXiv preprint arXiv:2501.07766 (2025)

    2025 arXiv

  31. [39]

    Dexon Mckensy-Sambola, Miguel Ángel Rodríguez-García, Francisco García- Sánchez, and Rafael Valencia-García. 2021. Ontology-based nutritional recom- mender system. Applied Sciences 12, 1 (2021), 143

    2021

  32. [40]

    Microsoft Research. 2024. Phi-4: A 14B-parameter language model with 16K context window. https://huggingface.co/microsoft/phi-4. Released under the MIT License; arXiv technical report also available

    2024

  33. [41]

    Weiqing Min, Chunlin Liu, Leyi Xu, and Shuqiang Jiang. 2022. Applications of knowledge graphs for food science and industry. Patterns 3, 5 (2022)

    2022

  34. [42]

    Mistral AI. 2025. Mistral-Small-3.2-24B-Instruct-2506: minor update improving instruction-following, reducing repetition, and enhancing function calling. https: //huggingface.co/mistralai/Mistral-Small-3.2-24B-Instruct-2506. Apache2.0 license; knowledge cut-off 2023-10-01; con...

    2025

  35. [43]

    Javier Marın, Aritro Biswas, Ferda Ofli, Nicholas Hynes, Amaia Salvador, Yusuf Aytar, Ingmar Weber, and Antonio Torralba. 2021. Recipe1m+: A dataset for learning cross-modal embeddings for cooking recipes and food images. IEEE Transactions on Pattern Analysis and Machine Intel...

    2021

  36. [44]

    Morris, Brandon Duderstadt, and Andriy Mulyar

    Zach Nussbaum, John X. Morris, Brandon Duderstadt, and Andriy Mulyar. 2024. Nomic Embed: Training a Reproducible Long Context (8192) Text Embedder. https://www.nomic.ai/blog/posts/nomic-embed-text-v1. Open -source model with open weights, training code, and data; 8192 -token c...

    2024

  37. [45]

    Ollama Contributors. 2025. Ollama: A lightweight, extensible framework for run- ning large language models locally. https://github.com/ollama/ollama. GitHub repo (MITlicense)

    2025

  38. [46]

    Ishita Padhiar, Oshani Seneviratne, Shruthi Chari, Dan Gruen, and Deborah L McGuinness. 2021. Semantic modeling for food recommendation explanations. In Abdur Rahman L. et al. 2021 IEEE 37th International Conference on Data Engineering Workshops (ICDEW) . IEEE, 13–19

    2021

  39. [47]

    Maria Panagiotou, Ioannis Papathanail, Lubnaa Abdur Rahman, Lorenzo Brigato, Natalie S Bez, Maria F Vasiloglou, Thomai Stathopoulou, Bastiaan E de Galan, Ulrik Pedersen-Bjergaard, Klazine van der Horst, et al . 2023. A Complete AI- Based System for Dietary Assessment and Perso...

    2023

  40. [48]

    Paweł Niszczota and Iga Rybicka. 2023. The credibility of dietary advice for- mulated by ChatGPT: Robo-diets for people with food allergies. Nutrition 112 (2023), 112076

    2023

  41. [49]

    Ioannis Papathanail, Lubnaa Abdur Rahman, Lorenzo Brigato, Natalie S Bez, Maria F Vasiloglou, Klazine van der Horst, and Stavroula Mougiakakou. 2023. The nutritional content of meal images in free-living conditions—automatic assessment with goFOODTM. Nutrients 15, 17 (2023), 3835

    2023

  42. [50]

    Ioannis Papathanail, Jana Brühlmann, Maria F Vasiloglou, Thomai Stathopoulou, Aristomenis K Exadaktylos, Zeno Stanga, Thomas Münzer, and Stavroula Mougiakakou. 2021. Evaluation of a novel artificial intelligence system to moni- tor and assess energy and macronutrient intake in...

    2021

  43. [51]

    Valentina Ponzo, Ilaria Goitre, Enrica Favaro, Fabio Dario Merlo, Maria Vittoria Mancino, Sergio Riso, and Simona Bo. 2024. Is Chatgpt an effective tool for providing dietary advice? Nutrients 16, 4 (2024), 469

    2024

  44. [52]

    thinking

    Qwen Team, Alibaba Cloud. 2025. Qwen3: Think Deeper, Act Faster. https: //qwenlm.github.io/blog/qwen3/. Introduces 8 open -weighted models (dense & MoE) ranging from 0.6B–235B params, supports 128K context window, hybrid "thinking" mode, Apache2.0 license

    2025

  45. [53]

    Ilias Papastratis, Dimitrios Konstantinidis, Petros Daras, and Kosmas Dimitropou- los. 2024. AI nutrition recommendation using a deep generative model and ChatGPT. Scientific Reports 14, 1 (2024), 14620

    2024

  46. [54]

    Shaina Raza, Mizanur Rahman, Safiullah Kamawal, Armin Toroghi, Ananya Raval, Farshad Navah, and Amirmohammad Kazemeini. 2024. A comprehensive review of recommender systems: Transitioning from theory to practice. arXiv preprint arXiv:2407.13699 (2024)

    2024

  47. [55]

    Ricardo Rei, Craig Stewart, Ana C Farinha, and Alon Lavie. 2020. COMET: A neural framework for MT evaluation. arXiv preprint arXiv:2009.09025 (2020)

    2020 arXiv

  48. [56]

    Nils Reimers, Iryna Gurevych, and the Sentence -Transformers Team. 2020. all-MiniLM-L6-v2: Efficient sentence embedding model (384 -dim vectors) fine-tuned on 1 billion pairs. https://huggingface.co/sentence-transformers/all- MiniLM-L6-v2. Apache2.0 license; trained via contra...

    2020

  49. [57]

    Thomas Rosemann, Andrea Bachofner, and Oliver Strehle. 2024. Kardiovaskuläre Erkrankungen in der Schweiz – Prävalenz und Versorgung [Cardiovascular diseases in Switzerland - Prevalence and care]. Praxis (Bern 1994) 113, 3 (March 2024), 57–66

    2024

  50. [58]

    John W Ratcliff, David E Metzener, et al. 1988. Pattern matching: The gestalt approach. Dr. Dobb’s Journal 13, 7 (1988), 46

    1988

  51. [59]

    Santiago Ruiz-Rincón and Ixent Galpin. 2024. SnapChef: AI-powered Recipe Suggestions. (2024)

    2024

  52. [60]

    Fatemeh Sarani Rad, Rasha Hendawi, Xinyi Yang, and Juan Li. 2024. Personalized Diabetes Management with Digital Twins: A Patient-Centric Knowledge Graph Approach. Journal of Personalized Medicine 14, 4 (March 2024), 359. https: //doi.org/10.3390/jpm14040359

    2024 doi

  53. [61]

    Sola S Shirai, Oshani Seneviratne, Minor E Gordon, Ching-Hua Chen, and Debo- rah L McGuinness. 2021. Identifying ingredient substitutions using a knowledge graph of food. Frontiers in Artificial Intelligence 3 (2021), 621766

    2021

  54. [62]

    Beata Sińska and Alicja Kucharska. 2023. Dietary guidelines in diabetes–why are they so difficult to follow? Pediatric Endocrinology Diabetes and Metabolism 29, 3 (2023), 125–127

    2023

  55. [63]

    Konstantinos Rouskas, Mary Guela, Marianna Pantoura, Ioannis Pagkalos, Maria Hassapidou, Elena Lalama, Andreas FH Pfeiffer, Elise Decorte, Veronique Cor- nelissen, Saskia Wilson-Barnes, et al . 2025. The Influence of an AI-Driven Personalized Nutrition Program on the Human Gut...

    2025

  56. [64]

    Swiss Federal Commission on Nutrition (NAE). 2025. Nährwertdaten- banken. https://www.nae.admin.ch/nae/en/home/fachinformationen/ naehrwertdatenbanken.htm. Accessed June 26, 2025

    2025

  57. [65]

    Swiss Medical Network. n.d.. Diabetology. https://www.swissmedical.net/en/ diabetology Accessed: 2025-06-24

    2025

  58. [66]

    Chihiro Taguchi, Seiji Maekawa, and Nikita Bhutani. 2025. Efficient Context Selection for Long-Context QA: No Tuning, No Iteration, Just Adaptive-𝑘. arXiv preprint arXiv:2506.08479 (2025)

    2025

  59. [67]

    Katherine Thornton, Kenneth Seals-Nutt, Mika Matsuzaki, and Damion Dooley

  60. [68]

    Guangzhi Sun, Xiao Zhan, and Jose Such. 2024. Building better ai agents: A provocation on the utilisation of persona in llm-based conversational agents. In Proceedings of the 6th ACM Conference on Conversational User Interfaces . 1–6

    2024

  61. [69]

    Department of Agriculture, Agricultural Research Service

    U.S. Department of Agriculture, Agricultural Research Service. 2019. FoodData Central. https://fdc.nal.usda.gov/. Launched April 2019; metadata updated April 21,2025

    2019

  62. [70]

    Francesco Vitali, Rosario Lombardo, Damariz Rivero, Fulvio Mattivi, Pietro Franceschi, Alessandra Bordoni, Alessia Trimigno, Francesco Capozzi, Giovanni Felici, Francesco Taglino, et al. 2018. ONS: an ontology for a standardized descrip- tion of interventions and observational...

    2018

  63. [71]

    Kun Wang, Guibin Zhang, Zhenhong Zhou, Jiahao Wu, Miao Yu, Shiqian Zhao, Chenlong Yin, Jinhu Fu, Yibo Yan, Hanjun Luo, et al. 2025. A comprehensive survey in llm (-agent) full stack safety: Data, training and deployment. arXiv preprint arXiv:2504.15585 14, 8 (2025)

    2025 arXiv

  64. [72]

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. 2022. Chain-of-thought prompting elicits reason- ing in large language models. Advances in neural information processing systems 35 (2022), 24824–24837

    2022

  65. [73]

    Semantic Web 15, 4 (2024), 1195–1206

    Reuse of the FoodOn ontology in a knowledge base of food composition data. Semantic Web 15, 4 (2024), 1195–1206

    2024

  66. [74]

    Maria Tziva, SO Negro, Agni Kalfagianni, and Marko P Hekkert. 2020. Un- derstanding the protein transition: The rise of plant-based meat substitutes. Environmental innovation and societal transitions 35 (2020), 217–231

    2020

  67. [75]

    Haoyu Zhang, Jun Liu, Zhenhua Zhu, Shulin Zeng, Maojia Sheng, Tao Yang, Guohao Dai, and Yu Wang. 2024. Efficient and Effective Retrieval of Dense-Sparse Hybrid Vectors using Graph-based Approximate Nearest Neighbor Search.arXiv preprint arXiv:2410.20381 (2024)

    2024 arXiv

  68. [76]

    Jian Zhang, Bifan Wei, Shihao Qi, haiping Zhu, Jun Liu, and Qika Lin. 2025. GKG- LLM: A Unified Framework for Generalized Knowledge Graph Construction. arXiv:2503.11227 [cs.AI] https://arxiv.org/abs/2503.11227

    2025 arXiv

  69. [77]

    Lingfeng Zhong, Jia Wu, Qian Li, Hao Peng, and Xindong Wu. 2023. A Compre- hensive Survey on Automatic Knowledge Graph Construction. ACM Comput. Surv. 56, 4, Article 94 (Nov. 2023), 62 pages. https://doi.org/10.1145/3618295

    2023 doi

  70. [78]

    Pengfei Zhou, Weiqing Min, Chaoran Fu, Ying Jin, Mingyu Huang, Xiangyang Li, Shuhuan Mei, and Shuqiang Jiang. 2025. FoodSky: A food-oriented large language model that can pass the chef and dietetic examinations. Patterns 6, 5 (2025), 101234. https://doi.org/10.1016/j.patter.20...

    2025

  71. [79]

    World Health Organization. 2024. Noncommunicable diseases. https://www. who.int/news-room/fact-sheets/detail/noncommunicable-diseases Fact sheet, last updated 23 December 2024

    2024

  72. [80]

    World Health Organization, Western Pacific Regional Office. 2021. Dietary Restrictions Infographic. Technical poster (digital), https://www.who.int/docs/ default-source/wpro---documents/posters/food-safety/dietary-restrictions- infographic---digital---english-wpro.pdf?sfvrsn=5...

    2021

  73. [85]

    Samuel Zumthurm, Ioannis Papathanail, Lubnaa Abdur Rahman, Lorenzo Brigato, Stavroula Mougiakakou, and Aline Stämpfli. 2025. Reducing meat consumption using a diet-related written prompt and the Swiss food pyramid: A field study. Food Quality and Preference 126 (2025), 105416

    2025

  74. [2019]

    In The Semantic Web–ISWC 2019: 18th International Semantic Web Conference, Auckland, New Zealand, October 26–30, 2019, Proceedings, Part II 18

    FoodKG: a semantics-driven knowledge graph for food recommendation. In The Semantic Web–ISWC 2019: 18th International Semantic Web Conference, Auckland, New Zealand, October 26–30, 2019, Proceedings, Part II 18 . Springer, 146–162

    2019

  75. [2023]

    Procedia Computer Science 225 (2023), 2028–2036

    Smart cuisine: Generative recipe & chatgpt powered nutrition assistance for sustainable cooking. Procedia Computer Science 225 (2023), 2028–2036

    2023

  76. [2024]

    in: a Knowledge Graph for Indian Food

    Building FKG. in: a Knowledge Graph for Indian Food. arXiv preprint arXiv:2409.00830 (2024)

    2024 arXiv

  77. [2025]

    arXiv:2411.04920 [cs.CL] https://arxiv.org/abs/2411.04920

    Enabling LLM Knowledge Analysis via Extensive Materialization. arXiv:2411.04920 [cs.CL] https://arxiv.org/abs/2411.04920

    arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.