RouteDRAG: Learning to Route and Control Agentic Execution for Multimodal Document RAG
Summary
- Researchers introduce RouteDRAG, a learning-based framework designed to route and control agentic execution for multimodal document RAG.
- RouteDRAG features a Unified Document Index, Adaptive Path Routing, Dynamic Agent Execution, and Sequential Policy Learning.
- The framework selectively assigns queries to agentic execution, avoiding unnecessary computation for tasks that do not benefit from it.
- RouteDRAG achieves an average accuracy improvement of 3.47 percentage points over shorter-path-only execution across two benchmarks while assigning only 79.3% of queries to agents.
Why it matters
Optimizing agentic retrieval-augmented generation reduces unnecessary computation costs and latency while improving accuracy for complex multimodal document understanding tasks.