Dynamic RAG: Adaptive Retrieval for Generation
- Dynamic RAG is a retrieval-augmented generation paradigm that adaptively determines when and what to retrieve during the generation process.
- It leverages mechanisms such as query-adaptive routing, dynamic memory stores, and selective fallback to optimize evidence selection in real time.
- This approach enhances multi-hop reasoning by continuously aligning retrieval with the evolving contextual needs of the language model.
Searching arXiv for recent Dynamic RAG papers to ground the article. Dynamic Retrieval-Augmented Generation (Dynamic RAG) denotes a family of retrieval-augmented generation architectures in which retrieval is no longer a single fixed preprocessing step, but an adaptive mechanism coupled to the state, needs, or operating conditions of the generation pipeline. In the tutorial framing of "Dynamic and Parametric Retrieval-Augmented Generation" (Su et al., 7 Jun 2025), Dynamic RAG adaptively determines when to retrieve and what to retrieve during the LLM’s generation process, in contrast to static retrieve-then-generate pipelines that retrieve once from the initial query and condition on that evidence throughout decoding. Across the recent literature, the term has broadened beyond token-level retrieval scheduling to include query-adaptive routing over large corpora, dynamically optimized cooperation between retrieval and generation modules, adaptive reranking and evidence budgeting, dynamic memory stores, and selective fallback from static text retrieval to external tools or APIs (Su et al., 7 Jun 2025).
1. Conceptual scope and definition
Dynamic RAG is defined most directly as a departure from the conventional static pipeline in which an external retriever first selects documents from the user’s initial query and the LLM then conditions on those retrieved passages for the entire response (Su et al., 7 Jun 2025). The central claim of the dynamic paradigm is that the model’s information needs evolve during generation, especially in multi-hop