A developer released CDRAG (Clustered Dynamic RAG), an open-source extension of Clustered RAG that routes queries through LLM-selected document clusters before retrieval. On 100 legal questions from the legal RAG bench dataset, it reports +12% faithfulness and +8% overall quality over standard RAG. Code and a writeup are on GitHub.
CDRAG (Clustered Dynamic RAG) improves retrieval-augmented generation by first clustering embedded documents into semantic groups, then using an LLM to generate keywords per cluster for guided retrieval. It addresses the problem of standard cosine retrieval missing higher-level semantic relevance, demonstrated on legal QA tasks.
CDRAG (Clustered Dynamic RAG) improves retrieval-augmented generation by first clustering embedded documents into semantic groups, then using an LLM to generate keywords per cluster for guided retrieval. It addresses the problem of standard cosine retrieval missing higher-level semantic relevance, demonstrated on legal QA tasks.
A new arXiv paper introduces ORDER (Optimal Routing for Dynamic Evidence Retrieval), a framework that conditions both indexing and retrieval on the incoming query instead of using one fixed RAG configuration. It clusters queries, learns per-cluster chunking and filtering strategies, and routes at inference via nearest-centroid assignment. Evaluations on heterogeneous historical archives show consistent gains over naive baselines and strong state-of-the-art RAG systems.
A new arXiv paper introduces ORDER (Optimal Routing for Dynamic Evidence Retrieval), a framework that conditions both indexing and retrieval on the incoming query instead of using one fixed RAG configuration. It clusters queries, learns per-cluster chunking and filtering strategies, and routes at inference via nearest-centroid assignment. Evaluations on heterogeneous historical archives show consistent gains over naive baselines and strong state-of-the-art RAG systems.
An arXiv paper introduces RAGCompass, a retriever router for retrieval-augmented generation that routes per query by reading retrieved passage sets rather than query-specific features, enabling generalization to unseen retrievers and corpora. It improves answer F1 by 28% relative to the best fixed retriever and outperforms fusion baselines, query-only routers, and a prompted 8B LLM. For practitioners, it offers per-query retriever selection without retriever-specific parameters.
An arXiv paper introduces RAGCompass, a retriever router for retrieval-augmented generation that routes per query by reading retrieved passage sets rather than query-specific features, enabling generalization to unseen retrievers and corpora. It improves answer F1 by 28% relative to the best fixed retriever and outperforms fusion baselines, query-only routers, and a prompted 8B LLM. For practitioners, it offers per-query retriever selection without retriever-specific parameters.
A developer reports that adding a re-ranking pass improved RAG retrieval accuracy more than switching embedding models. The second-stage cross-encoder re-scores top results using full query context, catching relevant documents that basic vector similarity misses.
A developer reports that adding a re-ranking pass improved RAG retrieval accuracy more than switching embedding models. The second-stage cross-encoder re-scores top results using full query context, catching relevant documents that basic vector similarity misses.
A neural network retrieval head that transforms query embeddings before retrieval, enabling the system to weight embedding dimensions according to their predictive value for a specific corpus. This improves hit rate, completeness, and faithfulness in RAG systems by reducing noise from irrelevant dimensions.
A neural network retrieval head that transforms query embeddings before retrieval, enabling the system to weight embedding dimensions according to their predictive value for a specific corpus. This improves hit rate, completeness, and faithfulness in RAG systems by reducing noise from irrelevant dimensions.