R²Adapter: A Hybrid RAG Intelligent Routing Adapter That Cuts Graph Retrieval Overhead by 59%

The Hybrid RAG Dilemma

Retrieval-Augmented Generation (RAG) has become the dominant paradigm for equipping large language models (LLMs) with non-parametric knowledge. By incorporating external retrieval results before generation, RAG effectively mitigates issues like outdated knowledge and hallucinations. However, as real-world use cases grow more complex, traditional RAG approaches are hitting clear capability limits.

According to a recent study published on arXiv (arXiv:2609.02894), Vanilla RAG handles simple queries efficiently but often struggles with relational reasoning or multi-hop reasoning. For example, when a user asks something like "What city is the university attended by the founder of a certain company located in?" — a question requiring multiple chained retrieval steps — simple vector search rarely surfaces all the relevant information in one shot.

To address this, Graph-based RAG emerged as a solution. By building knowledge graphs and traversing entity relationships, it significantly improves multi-hop reasoning. But this capability comes at a cost — graph retrieval introduces higher inference complexity and latency, wasting resources on queries that don't actually need it.

Why a Fixed RAG Strategy Doesn't Work

In practice, query complexity varies enormously. Some questions can be answered with a single retrieval; others require traversing multiple knowledge nodes. This means neither sticking exclusively to Vanilla RAG nor Graph RAG is optimal — the former breaks down on complex reasoning, while the latter burns unnecessary compute on simple questions.

So, could a system dynamically choose the right RAG strategy based on query complexity? That's the core idea behind Hybrid text-graph RAG. However, most existing hybrid approaches rely on heuristic rules or LLM-based routing decisions. Both have serious drawbacks: heuristic rules struggle to cover diverse query patterns, while LLM-based routing introduces additional inference overhead and is heavily dependent on the underlying LLM's capabilities, limiting portability and generalizability.

R²Adapter's Dual Design: Routing and Rewriting

To tackle these pain points, the research team introduced R²Adapter — a lightweight, plug-and-play Routing and Rewriting Adapter. Its core goal is to dynamically distribute queries between Vanilla RAG and Graph RAG, saving resources where possible while deploying graph retrieval only when necessary.

Intelligent Routing: Precisely Filtering Queries That Need Graph Reasoning

The first key capability of R²Adapter is routing. It determines whether each query genuinely benefits from graph-structured reasoning. Only queries that truly stand to gain from graph reasoning are routed to the Graph RAG pipeline; simpler queries are handled by the more efficient Vanilla RAG. This eliminates unnecessary graph retrieval overhead.

The elegance of this design lies in offloading the routing decision from a heavy LLM to a lightweight adapter, reducing latency and reducing dependence on any specific large model.

Query Rewriting: Making Multi-Hop Reasoning Requirements Explicit

The second key capability of R²Adapter is rewriting. For boundary queries where it's uncertain whether graph reasoning is needed, the adapter rewrites them to make any implicit multi-hop reasoning requirements more explicit.

Notably, this rewriting process requires no additional supervision signals to improve retrieval quality. In other words, by making queries more "retrievable," the system further optimizes overall performance without increasing annotation costs. This "routing + rewriting" combination strikes a balance between efficiency and accuracy.

Experimental Results: Graph RAG Usage Reduced by 59%

The research team conducted extensive experiments on three multi-hop QA benchmark datasets. Results show that R²Adapter can reduce Graph RAG usage by up to 59%, while maintaining answer accuracy comparable to using Graph RAG exclusively.

The significance of these numbers is straightforward: more than half of graph retrieval costs are eliminated with virtually no loss in answer quality. For production environments where inference cost and response latency must be controlled, this is a highly attractive gain.

Equally important, R²Adapter is model-agnostic. It can be seamlessly integrated into various Vanilla RAG and Graph RAG pipelines as a general-purpose component, requiring no deep customization for specific LLMs or retrieval architectures. This plug-and-play nature significantly lowers the barrier to deployment.

Implications for Hybrid RAG Engineering Practice

The value of R²Adapter goes beyond a single concrete solution — it represents a design philosophy: use lightweight, specialized components instead of heavyweight general-purpose LLMs for decision-making.

In real-world RAG deployments, there is always a trade-off between cost, latency, and accuracy. The industry has historically tended to "throw more model" at routing problems, but this often yields diminishing returns. R²Adapter demonstrates that a small, purpose-trained adapter can handle intermediate decision tasks like routing both quickly and accurately.

Looking ahead, as RAG systems grow increasingly complex, a similar "divide and conquer" approach — applying different strategies to queries of different complexity, coordinated by efficient components — is likely to become the standard paradigm for building cost-effective RAG applications. For teams building enterprise-grade retrieval-augmented systems, the dynamic routing and query rewriting ideas behind R²Adapter are well worth taking seriously.