Overview

• HetaRAG outlines a hybrid deep-retrieval framework that orchestrates evidence across vector indexes, knowledge graphs, full‑text engines and structured databases, with partial code posted on GitHub.
• A proposed L1–L5 capability framework categorizes enterprise RAG tasks by modality and complexity and reports evaluations of platforms including LangChain, Azure AI Search, OpenAI and Corvic AI.
• MIXRAG introduces a mixture‑of‑experts graph retriever with dynamic routing and a query‑aware GraphEncoder to reduce noise, claiming state‑of‑the‑art results across graph tasks with code planned upon acceptance.
• Think‑on‑Graph 3.0 presents a multi‑agent approach that builds and refines a heterogeneous graph index during query‑time via dual evolution of the query and subgraph, with experiments showing outperformance of baselines.
• Influence Guided Context Selection proposes a Contextual Influence Value to rank retrieved passages via their estimated contribution at inference, offers a surrogate predictor to cut compute, and releases code for replication.