Efficient training-free query routing framework for federated retrieval-augmented generation

RAGSpot is a training-free query router for federated retrieval-augmented generation (FedRAG). Each corpus publishes only the centroids and radii of a K-Means clustering of its documents, computed in a shared routing space. At query time the router compares the query with these clusters, keeps the corpora whose radius-normalized distance is close to the best one, searches them with their own retrieval encoders, and fuses the per-corpus results with a two-level reciprocal rank fusion (document rank within a corpus and corpus rank). No router training is needed, and corpora can join or leave the federation by registering or removing their metadata. Contents of the archive: ragspot/ - Python implementation of RAGSpot and of the three baseline routers compared in the paper (C-FedRAG broadcast, RAGRoute learned MLP router, ReSLLM LLM-prompted router); the offline pipeline (document embedding, FAISS indexing, cluster metadata generation, query encoding); the comparison experiment (python -m ragspot.compare) and the hyperparameter sweep scripts; the motivation experiments on uniform versus corpus-specific encoders and on router retraining cost; the trained RAGRoute checkpoints; and the BEIR queries, relevance judgments, and cluster metadata of the eight corpora that make up the six-corpus federation and its two extensions (NFCorpus, SciFact, ArguAna, SciDocs, FiQA, Webis-Touche-2020, DBpedia-Entity, HotpotQA). The twelve-corpus federation of the revised paper is defined in the same configuration and adds NQ, FEVER, Climate-FEVER, and MS MARCO. 90 unit tests pin down the interfaces the experiments rely on. workloads/ - Scripts that construct the two workloads added in the revised paper: the twelve-corpus federation (per-corpus document sampling to at most 50,000 documents, query selection, composite-query pairing, and the generation-experiment subset, all seeded and recorded in stats.json) and the naturally phrased composite queries (rewriting of concatenated query pairs with Qwen2.5-7B-Instruct under greedy decoding, with the prompt used and a tool for the manual preservation audit with Wilson confidence intervals). Source documents are not included. All corpora come from the BEIR benchmark (https://github.com/beir-cellar/beir) and are downloaded with the links given in the READMEs; the FAISS indexes and embeddings are rebuilt from them deterministically (seed 42) with the included pipeline. The corpora remain subject to their original licenses. Requirements: Python 3.10-3.12, PyTorch, sentence-transformers, faiss-cpu, scikit-learn (see pyproject.toml). The rewriting script additionally needs transformers with a GPU able to run a 7B model.