Third-party cyber evaluations involving OpenAI models
openai.com·14h ago
AI Research
arXiv CS.AI·May 21·6 min read
DOTRAG: Retrieval-Time Reasoning Along Paths
TL;DR
Graph-based retrieval-augmented generation (GraphRAG)—systems that fetch relevant information from structured knowledge graphs before answering questions—typically use fixed heuristics to retrieve context, causing them to miss multi-step logical paths or include irrelevant data. DotRAG reformulates retrieval itself as a reasoning process that dynamically traces paths through the graph based on query logic, without requiring model retraining.
✦ Why It Matters
Engineers building knowledge-graph-backed QA systems can adopt training-free retrieval that adapts to complex queries without retraining models.
Key Takeaways
- Replace fixed retrieval heuristics with dynamic reasoning-guided path exploration in graphs.
- Eliminate separate retrieve-then-reason stages to reduce irrelevant context accumulation.
- Apply DotRAG without retraining on existing knowledge graphs immediately.
- Test on multi-hop reasoning benchmarks to validate query-specific path discovery.
Full Summary
Graph Retrieval-Augmented Generation combines knowledge graphs (structured data showing relationships between entities) with language models to answer complex questions. Traditional GraphRAG systems use a two-stage approach: retrieve relevant nodes using heuristics like keyword matching, then feed that context to a reasoning model.
This separation fails on multi-hop queries—questions requiring traversal through multiple relationship steps—because the retrieval stage cannot anticipate which paths the reasoning stage will need. DotRAG proposes a unified framework that treats retrieval itself as a reasoning task.
Instead of pre-selecting context, DotRAG dynamically explores graph paths during retrieval, guided by query semantics. The method requires no training, making it immediately applicable to existing graphs.
By integrating reasoning into the retrieval process, DotRAG reduces accumulated irrelevant context and improves accuracy on multi-hop reasoning tasks compared to traditional retrieve-then-reason baselines.