ZeroRAG
A decentralized routing layer for RAG document retrieval that performs semantic query routing over a peer-to-peer overlay network. It requires no global index or centralized coordination; nodes maintain only lightweight local state (neighbor advertisement sets and query caches) and achieve self-organization and self-healing through best-first walks and dynamic rewiring.
Timeline (last 90 days)
ZeroRAG is a decentralized routing layer for RAG document retrieval that performs semantic query routing over an overlay network of peer nodes, requiring neither a global data structure nor centralized orchestration.
In ZeroRAG, each node maintains only two kinds of lightweight local state: a bounded set of neighbor advertisements and a cache of recently routed queries.
ZeroRAG uses a best-first walk to forward queries step by step and continuously rewires the overlay network during the process, causing neighbor relationships to evolve toward semantic relatedness.
ZeroRAG's recall improvement depends on network self-organization driven by accumulated queries, so cold-start quality directly determines whether the system can enter a virtuous cycle.
The paper evaluates ZeroRAG using three real workloads and two decentralized baselines.
ZeroRAG outperforms both baselines on both recall and network message overhead.
ZeroRAG approaches the recall and LLM answer accuracy of centralized document storage.
ZeroRAG can self-heal after catastrophic node failures because there is no global index acting as a single point of dependency.
ZeroRAG's experiments scale to 100,000 nodes, validating the feasibility of this approach in large-scale peer-to-peer networks.
ZeroRAG suggests a different RAG deployment paradigm: data does not have to move, and retrieval capability can be distributed through a network protocol.
All Facts (10)
ZeroRAG is a decentralized routing layer for RAG document retrieval that performs semantic query routing over an overlay network of peer nodes, requiring neither a global data structure nor centralized orchestration.
50%UnverifiedIn ZeroRAG, each node maintains only two kinds of lightweight local state: a bounded set of neighbor advertisements and a cache of recently routed queries.
50%UnverifiedZeroRAG uses a best-first walk to forward queries step by step and continuously rewires the overlay network during the process, causing neighbor relationships to evolve toward semantic relatedness.
50%UnverifiedZeroRAG's recall improvement depends on network self-organization driven by accumulated queries, so cold-start quality directly determines whether the system can enter a virtuous cycle.
50%UnverifiedThe paper evaluates ZeroRAG using three real workloads and two decentralized baselines.
50%UnverifiedZeroRAG outperforms both baselines on both recall and network message overhead.
50%UnverifiedZeroRAG approaches the recall and LLM answer accuracy of centralized document storage.
50%UnverifiedZeroRAG can self-heal after catastrophic node failures because there is no global index acting as a single point of dependency.
50%UnverifiedZeroRAG's experiments scale to 100,000 nodes, validating the feasibility of this approach in large-scale peer-to-peer networks.
50%UnverifiedZeroRAG suggests a different RAG deployment paradigm: data does not have to move, and retrieval capability can be distributed through a network protocol.
50%