Retrieval augmented generation (RAG) combines language models with external corpora to support knowledge-intensive tasks, such as open-domain question answering. Standard RAG systems typically employ a fixed top-k retrieval strategy, retrieving the same number of passages regardless of query needs. This can lead to either insufficient evidence, or the inclusion of irrelevant contexts that lead to a degradation of generation performance. In this work, we conduct an empirical study of how irrelevant retrieved passages affect downstream generation, analyzing their impact across multiple standard generator models. Building on these insights, we propose a lightweight, context-size classification module that dynamically predicts how much context is required based on query-specific needs. We integrate this approach into a full RAG pipeline and demonstrate improved performance over several baselines.

International audience

International audience

Companies use these different approaches to elicit, specify, analyse and validate their requirements in different contexts.The globalization and the rapid development of information technologies sometimes require companies to work together in order to achieve common objectives as quickly as possible.We propose a Unified Requirements Engineering meta-model (UREM) that allows cooperation in the requirements engineering process between heterogeneous RE (Requirement Engineering) models.In this paper, we explore UREM as a pivot meta-model to ensure interoperability between heterogeneous RE models.

Graphs are mathematical structures that provide natural means for complex-data representation. Graphs capture the structure and thus help modeling a wide range of complex real-life data in various domains. Moreover graphs are especially suitable for information visualization. Indeed the intuitive visual-abstraction (dots and lines) they provide is intimately associated with graphs. Visualization paves the way to interactive exploratory data-analysis and to important goals such as identifying groups and subgroups among data and helping to understand how these groups interact with each other. In this paper, we present a graph drawing approach that helps to better appreciate the cluster structure in data and the interactions that may exist between clusters. In this work, we assume that the clusters are already extracted and focus rather on the visualization aspects. We propose an energy-based model for graph drawing that produces an esthetic drawing that ensures each cluster will occupy a separate zone within the visualization layout. This method emphasizes the inter-groups interactions and still shows the inter-nodes interactions. The drawing areas assigned to the clusters can be user-specified (prefixed areas) or automatically crafted (free areas). The approach we suggest also enables handling geographically-based clustering. In the case of free areas, we illustrate the use of our drawing method through an example. In the case of prefixed areas, we first use an example from citation networks and then use another example to compare the results of our method to those of the divide and conquer approach. In the latter case, we show that while the two methods successfully point out the cluster structure our method better visualize the global structure.

Several approaches adopted by the software engineering community rely on the principle of multi-modeling which allows to separate concerns and to model a system as a set of less complex sub-models. Model composition is a crucial activity in Model Driven Engineering (MDE). It is particularly useful when adopting a multi-modeling approach to analyze and design software systems. In previous work, we have defined a view-based UML profile called VUML. In this paper, we describe a formal approach for model composition in which we consider the composition as an algebraic operator on the set of UML-compliant models. We specify the semantics of our composition operator by means of graph transformations. Furthermore, we present a composition scheme interpreted by a two-steps composition process based on two strategies of correspondence and merging. To illustrate our approach, we apply it to the composition of UML class models diagrams into one VUML model class diagram.

No abstract available.