A developer (E. Zimmermann) introduces CoreQuarry, an Apache 2.0-licensed engine that exposes document structure as a queryable "textual graph" using positional and structural operators, reviving coordinate-addressing ideas from Gonnet's Unstructured Data Bases (1983) and HyTime (ISO/IEC 10744:1992). Agents can retrieve text by position, containment, and overlap instead of relying on arbitrary chunk boundaries, letting the LLM reason over semantically assembled context.
A Reddit post recaps Anthropic's published contextual retrieval numbers, arguing the largest RAG improvement came from adding keyword search rather than a better model. Chunk-level context descriptions, hybrid keyword search, and reranking cut retrieval failures progressively by 35%, 49%, and 67%. Practitioners can apply these as retrieval-pipeline techniques independent of model choice.
A Reddit post recaps Anthropic's published contextual retrieval numbers, arguing the largest RAG improvement came from adding keyword search rather than a better model. Chunk-level context descriptions, hybrid keyword search, and reranking cut retrieval failures progressively by 35%, 49%, and 67%. Practitioners can apply these as retrieval-pipeline techniques independent of model choice.
TopoAgent replaces linear planning with dynamic graph evolution to improve multimodal scientific reasoning. It decomposes complex queries into visually-grounded atoms organized as a topological graph, addressing issues like visual-semantic misalignment and long-context hallucinations.
TopoAgent replaces linear planning with dynamic graph evolution to improve multimodal scientific reasoning. It decomposes complex queries into visually-grounded atoms organized as a topological graph, addressing issues like visual-semantic misalignment and long-context hallucinations.
A new arXiv paper introduces DocNavRAG, a document-structured Graph RAG approach that organizes document hierarchies and cross-region relations into a navigable graph. It exposes graph operations for locating, navigating, expanding, and fetching, enabling agents to assemble evidence across sections and documents more effectively than fixed-traversal GraphRAG or weakly structured agentic RAG.
A new arXiv paper introduces DocNavRAG, a document-structured Graph RAG approach that organizes document hierarchies and cross-region relations into a navigable graph. It exposes graph operations for locating, navigating, expanding, and fetching, enabling agents to assemble evidence across sections and documents more effectively than fixed-traversal GraphRAG or weakly structured agentic RAG.
An arXiv paper introduces the Wiki Foundation Model (WFM), a scalable, agent-native alternative to sparse graph embeddings for long-term agent memory and multi-hop reasoning. WFM combines a Wiki Graph schema, query-conditioned attentive aggregation, and a custom NCCL boundary exchange protocol. Evaluations across five benchmarks show strong performance plus 10.5x distributed training acceleration.
An arXiv paper introduces the Wiki Foundation Model (WFM), a scalable, agent-native alternative to sparse graph embeddings for long-term agent memory and multi-hop reasoning. WFM combines a Wiki Graph schema, query-conditioned attentive aggregation, and a custom NCCL boundary exchange protocol. Evaluations across five benchmarks show strong performance plus 10.5x distributed training acceleration.
An arXiv paper introduces NS-ST-GraphRAG, a neuro-symbolic spatio-temporal GraphRAG framework targeting retrieval-augmented generation over long-form literary narratives. It combines ontology-guided extraction, deterministic constraint checking, dual temporal coordinates, spatial scene attributes, and dynamic sub-graph retrieval to handle evidence spread across chapters and evolving character relations.
An arXiv paper introduces NS-ST-GraphRAG, a neuro-symbolic spatio-temporal GraphRAG framework targeting retrieval-augmented generation over long-form literary narratives. It combines ontology-guided extraction, deterministic constraint checking, dual temporal coordinates, spatial scene attributes, and dynamic sub-graph retrieval to handle evidence spread across chapters and evolving character relations.