Culinary Crossroads: A RAG Framework for Enhancing Diversity in Cross-Cultural Recipe Adaptation

1. Problem Statement

The paper addresses Cross-Cultural Recipe Adaptation (CCRA), a task where a recipe from a source culture (e.g., Latin American) is transformed to align with the culinary norms and dietary preferences of a target culture (e.g., Spanish) while preserving the dish's essence.

While existing methods focus on generating high-quality adapted recipes, they largely overlook output diversity. In creative tasks like cooking, a single source recipe often has multiple valid cultural adaptations (e.g., substituting nopales with zucchini, chard, or eggplant).
The authors identify a critical gap in Retrieval-Augmented Generation (RAG):

• The Paradox: Even when RAG systems are provided with diverse, culturally relevant retrieved contexts, they fail to generate diverse outputs.
• The Limitation: Standard RAG frameworks tend to rely heavily on a narrow subset of the retrieved context, ignoring the variety of information available. Consequently, they often produce less diverse results than "closed-book" Large Language Models (LLMs) that rely solely on internal knowledge.

2. Methodology: CARRIAGE

To address these limitations, the authors propose CARRIAGE (Cultural-Aware Recipe Retrieval Augmented Generation), a plug-and-play, training-free framework designed to enhance diversity in both retrieval and generation stages.

The framework operates through four key components:

1. Query Rewriting:

   • Goal: Address semantic gaps caused by cultural naming differences (e.g., a dish named differently in the target culture).
   • Mechanism: The source recipe title is rewritten based on content and prompted to culturally adapt the title for the target audience before retrieval. This ensures the retrieval system finds culturally appropriate candidates that might otherwise be missed.

2. Diversity-Aware Re-ranking:

   • Goal: Prevent the retrieval of overly similar recipes.
   • Mechanism: An extension of the Maximal Marginal Relevance (MMR) algorithm. The ranking score balances relevance to the query against similarity to:
     • Already selected documents ($S$).
     • Recipes previously generated by the RAG model ($H$).
   • Formula: $Score(D_i) = \max [\lambda \cdot Rel(D_i) - (1-\lambda) \cdot \max_{D_j \in S \cup H} Sim(D_i, D_j)]$.

3. Dynamic Context Organization:

   • Goal: Force the LLM to utilize different parts of the retrieved context across multiple generation attempts.
   • Mechanism: Instead of feeding all $k$ retrieved recipes to the LLM, a sliding window of size $w$ is applied. For the $t$-th generation, the model only sees a specific subset $\{D_{tw+1}, \dots, D_{(t+1)w}\}$. This prevents the model from fixating on the same top-ranked documents.

4. Contrastive Context Injection:

   • Goal: Explicitly prevent the LLM from repeating previous outputs.
   • Mechanism: The system retrieves the LLM's own previously generated outputs for the same source recipe and injects them as a "contrastive context." The prompt explicitly instructs the LLM to avoid generating results similar to these past outputs.

3. Key Contributions

• First RAG Framework for Diverse CCRA: This is the first work to explicitly design a RAG framework to generate highly diverse outputs for cross-cultural adaptation, moving beyond single-answer optimization.
• Novel Evaluation Metric: Introduction of Recipe Cultural Appropriateness Score (CultureScore), a BERT-based classifier that predicts the country of origin of a recipe based on ingredients and title, serving as a proxy for cultural alignment.
• Comprehensive Metrics: A multi-dimensional evaluation framework covering:
  • Diversity: Lexical (Unique-n), Semantic (Sentence-BERT cosine distance), and Ingredient diversity.
  • Quality: Source Recipe Preservation (BERTScore) and Cultural Appropriateness (CultureScore).
• Empirical Discovery: The finding that standard RAG often yields lower diversity than closed-book LLMs because it fails to leverage contextual diversity effectively.

4. Experimental Results

The study was conducted on the RecetasDeLaAbuel@ dataset, adapting recipes from 7 Latin American countries to Spanish culinary preferences.

• Pareto Efficiency: CARRIAGE achieves Pareto efficiency, meaning it improves diversity without sacrificing quality (or vice versa). It outperforms both closed-book LLMs (LLaMA3.1, Qwen2.5, Gemma2) and standard RAG baselines.
• Diversity vs. Quality Trade-off:
  • Closed-book LLMs: High preservation, low diversity.
  • Standard IR/RAG: High diversity (in retrieval) but low preservation or low utilization of that diversity in generation.
  • CARRIAGE: Successfully balances high diversity (Lexical, Semantic, Ingredient) with strong cultural appropriateness and source preservation.
• Context Utilization Analysis:
  • Standard RAG models rely on only 1–2 context recipes in ~76% of generations, even when diverse contexts are provided.
  • CARRIAGE increases the contextual diversity utilization by over 40%, forcing the model to draw from a broader range of retrieved information.
• Metric Correlation: The study found a positive correlation between diversity and cultural appropriateness, but a negative correlation between diversity and source preservation (staying too close to the original limits diversity).

5. Significance and Impact

• Theoretical Insight: The paper challenges the assumption that providing diverse context in RAG automatically leads to diverse generation. It highlights a "utilization gap" where LLMs in RAG settings default to the most relevant (and often similar) parts of the context.
• Practical Application: CARRIAGE offers a scalable, training-free solution for applications requiring creative variation, such as personalized diet planning, cultural education, and creative writing, where multiple valid answers are preferred over a single "correct" one.
• Future Direction: It paves the way for RAG systems in open-ended, creative domains that require balancing factual grounding with subjective variation, moving beyond the traditional "single correct answer" paradigm of RAG.