Chain of Exploration: Refining Fuzzy Recipe Queries

• Chain of Exploration (CoE) is a mechanism that converts ambiguous recipe queries into structured, machine-actionable conditions within ChefMind.
• It employs a five-level progressive search strategy—from exact matching to broad keyword matching—to dynamically refine fuzzy user inputs.
• Empirical results on the Xiachufang dataset demonstrate that CoE significantly reduces unprocessed queries, boosting overall system accuracy.

Chain of Exploration (CoE) is a chain-based intent-refinement mechanism that, in the current literature, is defined most explicitly as the front-end component of ChefMind for handling ambiguous recipe queries. In that formulation, CoE converts fuzzy natural-language requests into structured, machine-actionable conditions for downstream retrieval, chiefly by progressively exploring several semantic interpretations of user intent before passing the refined result to a Knowledge Graph, Retrieval-Augmented Generation, and an LLM (Fu et al., 22 Sep 2025). The acronym is not stable across recent arXiv work: other papers use CoE for “Chain of Execution,” “Chain of Evidence,” “Chain-of-Events,” “Chain-of-Experts,” and other distinct constructs, so the meaning of CoE is paper-specific rather than uniform across domains (Chen et al., 24 Oct 2025, Liu et al., 2 May 2026).

1. Definition and problem setting

In ChefMind, Chain of Exploration is introduced as the mechanism that addresses fuzzy user intent in recipe recommendation. Its role is summarized directly as follows: CoE “refines ambiguous queries into structured conditions,” “dynamically interprets and refines ambiguous user queries into structured conditions,” serves as the “entry point for fuzzy demand parsing,” and converts “abstract user inputs into quantifiable screening conditions” (Fu et al., 22 Sep 2025). The target queries are described as “fuzzy demands,” “ambiguous queries,” “abstract user inputs,” and “fuzzy user intent.”

The paper characterizes the ambiguity problem along several axes. It includes lexical ambiguity, underspecification, latent-preference uncertainty, and semantic mismatch between natural language and recipe records. Examples given in the description include short or vague requests, flavor- or scenario-oriented language rather than exact dish names, and phrases such as “healthy comfort food,” which may not map cleanly to fixed recipe keywords or dish titles (Fu et al., 22 Sep 2025). This suggests that, within ChefMind, CoE is not a general reasoning doctrine but a domain-specific mediation layer between free-form human intent and structured retrieval.

The same paper also makes clear what CoE is not. It is not described as an interactive clarification loop with the user, nor as a pure prompt-engineering analogue of chain-of-thought. The “chain” is internal system exploration over possible semantic interpretations of a query rather than a multi-turn dialogue (Fu et al., 22 Sep 2025).

2. Core mechanism

The activation condition for CoE is given by a binary fuzzy-demand judgment rule:

$$\text{fuzzy}(Q) = \begin{cases} 1, & \text{if ambiguous terms} \in Q \lor |Q| < 5 \ 0, & \text{otherwise} \end{cases}$$

Here $Q$ is the user query. If $\text{fuzzy}(Q)=1$, CoE is triggered; otherwise, the query bypasses CoE and proceeds directly to KG screening (Fu et al., 22 Sep 2025).

Once activated, CoE uses what the paper calls “five-level progressive search logic.” These five levels are exact name matching for specific dish queries, ingredient similarity matching based on available ingredients, quick home-style dish retrieval, cuisine and flavor matching for culinary preferences, and broad keyword matching for comprehensive coverage (Fu et al., 22 Sep 2025). The paper presents this as a layered exploration-and-backoff strategy: highly specific interpretations are attempted first, while progressively broader semantic views of the request are used when the query is less explicit.

This five-level structure is the most concrete basis for the name Chain of Exploration. It explores dish identity, ingredient availability, home-style convenience, cuisine or flavor preference, and general topical keywords as alternative semantic lenses on the same request. Afterward, the outputs are “integrated via matching scores,” and deduplication is applied, “ensuring diversity” (Fu et al., 22 Sep 2025). The paper does not provide the matching-score formula or deduplication algorithm, but it does state that the result is converted into structured KG query conditions.

3. Architecture and end-to-end workflow

ChefMind embeds CoE in a larger hybrid architecture combining CoE, a Knowledge Graph, Retrieval-Augmented Generation, and an LLM. The paper’s implied end-to-end workflow is: a raw query enters CoE; CoE decides whether the query is fuzzy; clear queries go directly to KG screening; fuzzy queries are refined through the five exploration levels; the outputs are merged and deduplicated; the refined interpretation is converted into KG query conditions; KG retrieves candidate recipes; RAG retrieves contextual text fragments; and the LLM integrates KG and RAG outputs into final recommendations (Fu et al., 22 Sep 2025).

The downstream modules are described more formally than CoE itself. The KG is implemented in Neo4j and stores Recipe, Ingredient, and Keyword nodes with relations such as “CONTAINS” and “HAS_KEYWORD.” RAG is implemented over Milvus with 768-dimensional vectors. Its similarity function is given as

$$\text{Sim}(\mathbf{v}_q, \mathbf{v}_d) = \frac{\mathbf{v}_q \cdot \mathbf{v}_d}{|\mathbf{v}_q| \cdot |\mathbf{v}_d|}$$

where $\mathbf{v}_q$ is described as the 768-dimensional vector of a candidate recipe and $\mathbf{v}_d$ is the vector of a text fragment in the RAG corpus (Fu et al., 22 Sep 2025). Although this is not a CoE-specific equation, it clarifies how CoE’s refined conditions affect subsequent retrieval and enrichment.

The paper is explicit that CoE outputs structured conditions rather than final recommendations. In the architecture description, for fuzzy demands, “CoE generates refinement logic converted into KG query conditions” (Fu et al., 22 Sep 2025). This makes CoE an upstream semantic-conditioning module rather than the final reasoning or generation engine.

4. Empirical performance and system-level effects

ChefMind is evaluated on the Xiachufang dataset and 129 human-annotated queries across 13 batches, comparing LLM+KG, LLM+RAG, and the full ChefMind system (Fu et al., 22 Sep 2025). The reported overall scores are 6.4 for LLM+KG, 6.7 for LLM+RAG, and 8.7 for ChefMind. The dimension-wise averages are Accuracy 6.2, Relevance 6.0, Completeness 6.5, Clarity 6.8 for LLM+KG; Accuracy 6.5, Relevance 6.3, Completeness 7.0, Clarity 7.1 for LLM+RAG; and Accuracy 8.5, Relevance 8.8, Completeness 8.6, Clarity 9.0 for ChefMind (Fu et al., 22 Sep 2025).

The strongest evidence for CoE’s practical value comes from handling fuzzy demands. Unprocessed-query counts are reported as 33/129 = 25.6% for LLM+KG, 22/129 = 17.1% for LLM+RAG, and 2/129 = 1.6% for ChefMind (Fu et al., 22 Sep 2025). The paper highlights fuzzy-query-heavy batches 3 and 7: in batch 3, ChefMind leaves 1 query unprocessed, compared with 5 for LLM+KG and 4 for LLM+RAG; in batch 7, ChefMind again leaves 1 unprocessed, compared with 4 and 3 respectively (Fu et al., 22 Sep 2025). The authors interpret this as robustness in handling fuzzy demands.

These results are system-level rather than module-isolated. The paper states that gains in accuracy and relevance “reflect the improvement of Chain of Exploration on intent parsing and the collaborative value of Knowledge Graph and Retrieval-Augmented Generation,” but it does not provide a no-CoE hybrid ablation that keeps KG, RAG, and the LLM fixed while removing only CoE (Fu et al., 22 Sep 2025). A plausible implication is that CoE is central to the observed robustness, but the exact contribution of CoE alone is not directly measured.

5. Interpretability, scope, and limitations

ChefMind attributes part of its interpretability to CoE’s conversion of vague language into “quantifiable screening conditions.” Once these conditions are passed into the KG, the retrieval path becomes inspectable because the KG “provides complete retrieval paths and associated attributes for each result,” enhancing “recommendation interpretability and transparency” (Fu et al., 22 Sep 2025). CoE therefore functions as an interpretability-enabling abstraction layer between fuzzy user language and graph-based retrieval.

The paper also exposes several limitations. The fuzzy-demand classifier is heuristic: ambiguity is triggered by “ambiguous terms” or by query length shorter than five characters, but the vocabulary or detection procedure for “ambiguous terms” is not specified (Fu et al., 22 Sep 2025). The text mentions a “3-step refinement logic,” yet the concrete procedural description that is actually provided is the five-level progressive search logic. Likewise, no scoring formula is given for integrating search results across the five levels, no pseudocode block is provided, and no interactive clarification mechanism is described.

Another limitation is evidential isolation. ChefMind compares the full CoE+KG+RAG+LLM system with LLM+KG and LLM+RAG baselines, but not with an otherwise identical hybrid model lacking CoE (Fu et al., 22 Sep 2025). This means the paper supports the importance of CoE indirectly, through the superiority of the full architecture on fuzzy queries, rather than through a strict module-level ablation.

6. Related constructs and acronym ambiguity

The phrase “Chain of Exploration” should be distinguished from several unrelated expansions of CoE in adjacent literature. “Chain of Execution” converts code execution into step-by-step natural-language traces for reasoning supervision and is explicitly stated not to concern “Chain of Exploration” (Chen et al., 24 Oct 2025). “Chain of Evidence” has been used for at least three different settings: pixel-level visual attribution in iterative retrieval-augmented generation, explicit evidence grounding for efficient video reasoning, and verifiability in autonomous research systems (Liu et al., 2 May 2026, Huang et al., 12 Jan 2026, Meng et al., 25 May 2026). “Chain-of-Events” denotes a training-free multimodal summarization framework guided by a hierarchical event graph (You et al., 6 Mar 2026). “Chain-of-Experts” refers to sequential expert communication inside sparse MoE layers (Wang et al., 23 Jun 2025). “Chain-of-Embedding” denotes hidden-state trajectories used for output-free self-evaluation (Wang et al., 2024). “Chain-of-Explanation” is a concept-circuit explainability framework for deep vision models (Yu et al., 19 Mar 2025).

Several neighboring works are exploration-relevant without using the exact phrase. “Conditionally Optimistic Exploration” introduces a sequential, agent-ordered exploration bonus for cooperative deep MARL, but the phrase “Chain of Exploration” does not appear there (Zhao et al., 2023). “CoEx — Co-evolving World-model and Exploration” develops a hierarchical LLM-agent architecture in which subgoal-driven exploration updates a persistent belief state (Kim et al., 29 Jul 2025). “Explore-Execute Chain” separates exploratory high-level planning from deterministic execution in structured reasoning (Yang et al., 28 Sep 2025). “V-REX” operationalizes multi-step exploratory visual reasoning as a “Chain-of-Questions,” splitting Planning from Following in VLM evaluation (Fan et al., 12 Dec 2025).

Taken together, these works show that “CoE” is an overloaded acronym rather than a stable technical term. In the strict sense of the phrase itself, Chain of Exploration is currently most concretely specified as ChefMind’s mechanism for converting ambiguous recipe requests into structured retrieval conditions through staged semantic exploration (Fu et al., 22 Sep 2025). In a broader methodological sense, the phrase also resonates with adjacent research on exploration-oriented planning, evidence chaining, and structured decomposition, but those works define different objects and should not be conflated with Chain of Exploration proper.