Dietary recommendation systems provide personalized food and nutrition suggestions based on individual health status and preferences. Traditional approaches have primarily relied on rule-based methods using expert knowledge and predefined guidelines, often lacking personalization and accessibility. Recent advancements in artificial intelligence have transformed this field through deep learning and knowledge graphs (KGs). Chen et al [13] defined food recommendation as constrained question answering over a food KG, integrating user preferences with health guidelines. PPKG [14] is a cancer-specific dietary recommendation system that combines a KG with time-aware long short-term memory networks to dynamically adjust recommendations for cancer prevention and rehabilitation. Ma et al [15] proposed a nutrition-related KG method based on graph convolutional networks to enhance food recommendation diversity while promoting more healthy eating habits. LLMs have further advanced dietary recommendation capabilities. ChatDiet [16] integrates personal and population models with causal inference to enhance personalization and explainability. Jin et al [17] demonstrated the clinical effectiveness of a GPT-based dietary recommendation system for patients with hemodialysis, successfully reducing serum potassium levels. Kopitar et al [18] created a generative artificial intelligence system for personalized inpatient meal planning that incorporates electronic health records and clinical guidelines.

However, we still do not have methods to integrate MFH principles with LLMs for dietary recommendation. Our framework bridges this gap by combining MFH KG with LLMs to develop more comprehensive and culturally informed dietary recommendations.

KGs serve as structured representations of factual knowledge, providing an excellent complement to parameterized language models [19,20]. With the emergence of LLMs, research has evolved along 2 complementary trajectories: leveraging LLMs to construct KGs and enhancing LLMs with KG integration. Several approaches have emerged for LLM-assisted KG construction. Extract-define-canonicalize [21] addresses scalability through open information extraction (OpenIE) and post hoc canonicalization. BEAR [22] leverages LLMs for zero-shot knowledge extraction in service-oriented domains. LLM-TIKG [23] applies few-shot learning for constructing threat intelligence KGs, while Yang et al [24] presented an approach for medical ontology expansion using LLMs to generate competency questions.

Researchers have also explored integrating KGs to enhance LLMs’ capabilities. GraphGPT [25] combines LLMs with graph structural knowledge through instruction tuning. GaLM [26] transforms KGs into text representations to improve reasoning while reducing hallucinations. Think-on-Graph [27] is a training-free framework that uses iterative beam search to enable interactive reasoning, while StructGPT [28] enhances zero-shot reasoning through an iterative Reading-then-Reasoning approach. Our work builds upon these dual research trajectories by proposing a framework that leverages LLMs for KG construction and uses KGs to enhance the capabilities of LLMs specifically within the domain of TCM, thereby providing reliable dietary recommendations grounded in MFH principles.

To construct an MFH KG, we systematically constructed a corpus from diverse sources, including ancient and classical literature, modern TCM literature, food and herbs information, regulatory documents, research papers, and TCM formulations. This process results in a corpus of 1359 relevant documents. Table 1 shows the statistical distribution of documents across different source categories.

To improve the efficiency and accuracy of entity-relation triple extraction, each document was segmented into passages of 300 tokens. These passages were then input into LLMs with a carefully designed prompt (Textbox S1 in Multimedia Appendix 1) to resolve pronominal references by incorporating information from document titles and neighboring passages. Additionally, irrelevant content within passages was filtered out to ensure domain-specific focus.

Based on the collected corpus, we constructed an MFH KG G, as shown in Figure 2, which illustrates the detailed process. Formally, the graph is defined as a triple G=(E,R,F), where E represents the set of entities, R represents the set of relations, and F⊆E×R×E represents the fact set. Each fact in F is represented as a quadruple (h,r,t,s), where h is the head entity, r is the relation, t is the tail entity, and s is the confidence score reflecting the likelihood that the triple (h,r,t) holds true. For instance, consider the quadruple (h: Insomnia and Vivid Dreams, r: Related Constitution, t: Yin-Deficient Constitution, s: 0.87). This quadruple indicates that Insomnia and Vivid Dreams are associated with the Yin-Deficient Constitution, with a confidence score of 0.87.

We used LLM-driven OpenIE to extract triples h, r, t from each passage within the corpus. Specifically, we first integrated predefined relation definitions into a prompt (Textbox S2 in Multimedia Appendix 1), guiding the LLM to identify and extract potential relations. These candidate relations, along with existing ones, were subsequently processed using another prompt (Textbox S3 in Multimedia Appendix 1), where the LLM determined whether to retain new relations or filter out redundant ones. This process resulted in the construction of a relation set rsel, which was incorporated into OpenIE prompts, instructing the LLM to extract entity-relation triples. The relations of extracted triples were strictly constrained to the relation set rsel.

Specifically, during the OpenIE, we instructed the LLM to extract attributes of food entities and dish entities (Textboxes S4 and S5 in Multimedia Appendix 1). For food entities, LLMs referenced their source passages to update attributes, including name, primary effects, consumption methods, indications, and contraindications. Similarly, for dish entities, attributes such as name, ingredient composition, cooking method, and dietary benefits were updated.

The automated construction of the MFH KG inevitably introduced noise, incompleteness, and factual inaccuracies, making absolute factual correctness unattainable. To mitigate these issues, we adopted UKG reasoning to measure the confidence of extracted triples and complete missing triples. Specifically, we first used unKR [30], an open-source Python library for UKG reasoning, to learn vector representations of entities and relations. Based on these representations, we applied the triple confidence measurement method [31], integrating heterogeneous evidences to measure confidence of each triple. For each given head entity h and relation r, candidate tail entities were enumerated to construct new triples absent from the current graph. The triples with confidence scores larger than 0.85 were added to the MFH KG.

Given a user query Q, we first constructed a prompt (Textbox 1) to guide the LLM to extract a set of relevant keywords E from Q. We then encoded both the keywords and all entities from the MFH KG into dense vector representations, denoted as HR and HG, respectively. Considering that SBERT [32] has demonstrated superior performance over BERT [33] and RoBERTa [34] on semantic textual similarity tasks while maintaining lower computational overhead [35], we adopted SBERT for embedding both the extracted keywords and entities. Subsequently, we computed the cosine similarities between the embeddings of each keyword and all entities. For each keyword, we retrieved the top-10 most similar entities from the MFH KG based on cosine similarity scores. These candidate entities were then presented to the LLM, which was instructed to identify the matched entity (Textbox 2). The selected entity corresponding to each keyword constitutes the query entity set EQ={e1,e2,… ,en}, which serves as the starting point for subsequent reasoning path retrieval.

This study used path-based retrieval from the MFH KG rather than retrieving contextual subgraphs for linked entities. This is motivated by the advantages reasoning paths offer in generating MFH-based dietary recommendations. Specifically, reasoning paths provide structured knowledge by capturing direct relations between query entities and answer entities. For example, the reasoning path “(Insomnia and Vivid Dreams, Related Constitution, Qi-deficient Constitution); (Qi-deficient Constitution, Suitable Food, Codonopsis)” enables LLMs to clearly infer the relations between a user’s symptoms, their constitution, and the recommended food, thereby enhancing the interpretability. In contrast, subgraph-based retrieval in long-hop reasoning tasks often introduces irrelevant information [36]. This not only leads to significant computational overhead but also hinders LLMs from effectively extracting relevant information from the retrieval results.

Built upon the previous work [37], we first provided the LLM with a structured prompt (Textbox 3) listing all relations in G. The LLM then generated multiple relation paths  {prel1, prel2, …, prelk}, where each preli is an ordered sequence of relations r1,r2,…,rl. These relation paths serve as guides for retrieving knowledge essential to answering user queries. Given that the MFH KG contained considerably fewer relations compared with large-scale KGs such as FreeBase [38], we did not fine-tune the LLM for relation path generation. Instead, we used few-shot prompting to enhance the LLM’s ability to generate meaningful relation paths.

Subsequently, starting from each entity in the EQ, we followed each generated relation path preli to retrieve a set of reasoning paths Prea​={prea1, prea2, …, pream} from G, where each reasoning path preai={(e1,r1,e2),(e2,r2,e3),…,(el,rl,el+1)} consists of multiple triples ej,rj,ej+1 and represents an instance of one of the relation paths.

In our initial experiment, we noticed that the reasoning paths in Prea​ contained redundant and irrelevant information. To enhance the quality of these reasoning paths, we computed a score for each reasoning path in Prea​ by considering both the entity importance and the triple confidence within the reasoning path. These scores enabled us to filter out less relevant reasoning paths.

The importance of each entity in the reasoning path was calculated using a weighted PageRank [39] algorithm. Traditional PageRank ignored the confidence of triples, which could lead to equivalent treatment of false and true relations, thus causing inaccuracies in entity importance. To overcome this limitation, we constructed a subgraph using entities from the Prea​ along with their one-hop neighbors and incorporated confidence scores of triples as edge weights in the PageRank computation. The weighted PageRank formula is defined as:

where PR(ei) represents the weighted PageRank score of entity ei, d is the damping factor, and N is the total number of entities in the graph. The term wj→i denotes the edge weight from entity ei to ej, derived from the confidence score of the corresponding triples. When multiple relations exist between ei and ej with different confidence scores, their influence was aggregated by averaging these scores. For scoring the reasoning paths, we integrated entity importance and triple confidence scores to formulate the following path-scoring function:

where wi-1→i represents the confidence score of the relation from entity ei to ej, n denotes the total number of entities in the path, and PR(ej) is the weighted PageRank score of entity ej within the reasoning path. This scoring methodology captured both the confidence of triples within the reasoning path and the importance of the entities involved, striking a balance between path reliability and informational importance. Finally, we ranked the reasoning paths based on their scores and retained only the top-k reasoning paths for subsequent inference.

We used prompt engineering to improve the accuracy and relevance of generated answers of LLM. Prompt engineering involves constructing carefully designed prompts to guide LLMs toward generating high-quality and task-specific responses. Specifically, we constructed a prompt template (Textbox 4) comprising four key components: (1) task specification, (2) user query, (3) retrieved reasoning paths, and (4) attributes of food and dish entities. This structured prompt facilitates the integration of external knowledge, improves the LLMs’ comprehension of MFH principles, and supports the generation of more accurate responses.

To quantitatively assess the effectiveness and accuracy of the model’s dietary recommendations, we used 2 evaluation metrics: Hits@1 and F₁-score. Hits@1 measures the proportion of queries where the top-ranked predicted recommendation matches a ground truth answer. This metric evaluates the model’s ability to prioritize the most relevant dietary recommendation. F₁-score captures the harmonic mean of precision (P) and recall (R), offering a balanced evaluation of both aspects. This metric is particularly appropriate in scenarios where multiple valid recommendations may exist for a single query. It is defined as follows:

where precision and recall are defined as:

In this context, TP (true positives) represents the number of dietary recommendations correctly suggested, FP (false positives) indicates inappropriate dietary recommendations incorrectly suggested, and FN (false negatives) refers to appropriate dietary recommendations that the model failed to suggest.

While automated evaluation methods are widely applied in natural language processing tasks, they have certain limitations in the context of dietary recommendations. Automated metrics struggle to assess the effectiveness of recommendations, specifically whether the suggested food and their combinations align with MFH principles. Moreover, user acceptability and explainability of the generated dietary recommendations are difficult to accurately assess using automated methods. Therefore, we incorporated human evaluation, integrating subjective ratings from both experts and general users to provide a more comprehensive assessment of the generated dietary recommendations.

Specifically, we assessed the generated dietary plans based on the following five evaluation criteria: (1) rationality, (2) explainability, (3) user acceptability, (4) personalization, and (5) consistency. Rationality assesses whether the recommended dietary plan adhered to both TCM principles and modern nutritional standards. Explainability evaluates whether the recommendations included clear TCM-based justifications. A higher score indicates that the system provided reasonable and understandable explanations rather than simply outputting ingredient combinations. User acceptability measures how well the recommendations aligned with users’ subjective preferences, including taste, cultural adaptability, and feasibility. Personalization assesses whether the system provided tailored recommendations based on an individual’s constitution and health conditions. Consistency examines whether the model produced stable recommendations under similar inputs.

Among these evaluation criteria, rationality, explainability, user acceptability, and personalization were assessed using a 5-point Likert scale (1‐5, where 1 represented the lowest score and 5 represented the highest). Consistency was evaluated using a binary judgment (Yes=5 points and No=0 points).

To conduct the evaluation, we recruited 5 experts with backgrounds in TCM or nutrition, along with 20 general users. Participants followed a structured human evaluation guideline to rate the generated dietary recommendations. The evaluation was conducted using a survey system, where assessors were provided with the user queries along with the corresponding dietary recommendations generated by LLMs. Finally, we computed the mean scores for each evaluation criterion.