Hybrid + Graph RAG wins
Enterprise search is moving from vector-only RAG to hybrid pipelines and graph-based document flows—teams report better recall and explainability by combining dense retrieval with sparse signals and dynamic stitching of multi-document context. Weaviate engineers laid out query-routing and decomposition tactics while MIT pushed a beginner-friendly RAG primer to standardize best practices. (x.com) (x.com)
Weaviate published an "Advanced RAG Techniques" e‑book that enumerates pre‑retrieval tactics like query transformation, query decomposition and query‑routing as well as retrieval levers such as hybrid search, metadata filtering, and embedding fine‑tuning. (weaviate.io)) Weaviate’s product docs and recipes show multi‑vector indexing and a Query Agent that parses complex questions, routes subqueries across collections, composes filters, and returns synthesized answers with citations for legal and enterprise content. (docs.weaviate.io)) Engineering notes and open‑source modules tied to Weaviate describe multi‑query generation plus RAGFusion (Reciprocal Rank Fusion) and multi‑query composition to stitch multi‑document context before passing compact, scored evidence to the LLM. (deepwiki.com)) Academic and challenge results quantify the hybrid case: a LiveRAG 2025 submission combined BM25 + E5 dense retrieval and reported MAP improvements from 0.523 to 0.797 after neural re‑ranking (52% relative gain) though with reranker latency rising from ~1.74s to ~84s per query. (arxiv.org)) Earlier retrieval research showed hybrid dense+sparse fusion can drive very large recall gains—one study reported up to a 580% increase in Recall@10 on MS MARCO when simple weighted fusion was applied versus dense‑only retrieval. (researchgate.net)) MIT materials have pushed beginner‑facing RAG teaching into mainstream coursework (MIT 6.S079 "Embeddings, RAG, and Vector Databases", March 14, 2024) while MIT authors also published PipeRAG (KDD ’25) and related systems work that focus on adaptive pipelines to cut RAG latency in production. (dsg.csail.mit.edu)) Enterprise write‑ups and partner case studies — including Box’s blog on a Weaviate Query Agent integration and practitioner guides for production Generative Search — document live deployments that combine hybrid retrieval, graph/schema routing, and re‑ranking to improve grounding and explainability in customer-facing search. (blog.box.com))
