01 · Enterprise Knowledge Graph Architecture

企业知识图谱（KG）架构

Research record	Details
Authors	Sanjeeve Kumar Gajadi
Published	2026-10-01
Sources	openalex
Focus	enterprise knowledge graph architecture, ontology & governance, entity resolution & provenance, graph-enhanced RAG and AI agents 企业知识图谱（KG）架构、本体与治理、实体解析与溯源、图增强检索增强生成（RAG）与智能体

English

The paper offers useful high-level design and governance guidance relevant to building persistent graph memory in production settings but appears to be architectural rather than experimental, so it is worth skimming for design patterns and operational concerns rather than deep methodological insight.

中文

该论文在构建生产化持久化图记忆方面提供有价值的高层设计与治理指导，但更偏架构层面而非实证方法学，因此适合略读以获取设计范式与运营性注意事项，而非深入方法细节。

Research synopsis

English

The paper identifies the problem of fragmented enterprise information spread across applications, documents, metadata, and processes, which hinders relationship-centric analysis. It proposes an Enterprise Knowledge Graph (KG) Architecture that assembles a governed relationship layer connecting business entities, semantics, lineage, processes, policies, and events. The framework emphasizes integration of ontology management, entity resolution, graph storage, provenance, security, query services, graph analytics, and pipeline components (including citations of SAP products) and explicitly mentions support for Retrieval-Augmented Generation (RAG) and AI agents. The stated contribution is a system-level architecture that enables enterprise search, impact analysis, data discovery, decision intelligence, and graph-enhanced RAG while preserving ownership, access control, semantic quality, and lifecycle governance.

中文

本文指出企业信息分散在应用、文档、元数据和流程中，阻碍基于关系的分析。论文提出一种企业知识图谱（KG）架构，将业务实体、语义、血缘、流程、策略和事件连接到受治理的关系层。该框架强调本体管理、实体解析、图存储、溯源、安全、查询服务、图分析与数据摄取等要素，并明确提到对检索增强生成（RAG）和人工智能智能体的支持。论文的贡献是提供一个系统级架构，以在保留所有权、访问控制、语义质量与生命周期治理的同时，支持企业搜索、影响分析、数据发现、决策智能与图增强RAG。

Thesis relevance

English

Overlap: both the thesis and this paper center on knowledge graph usage to improve search and RAG, and both highlight entity resolution, provenance, and governance concerns. Differences and limitations: the candidate is an enterprise-level architecture paper focused on systems integration and governance (including vendor platforms) and does not present empirical evaluation, incremental agentic graph construction, or experiments on multi-hop QA and agentic paradigms described in the thesis. Complementarity: the architecture’s treatment of ontology management, provenance, security, and lifecycle governance can inform design decisions for the thesis’s graph memory pipeline and operational constraints when deploying local, privacy-preserving agentic RAG.

中文

重合点：论文与本论文都聚焦于利用知识图谱（KG）改进搜索和RAG，并强调实体解析、溯源与治理问题。差异与局限：候选论文是偏向企业级的架构性工作，侧重系统集成与治理（包含供应商生态），未提供实证评估，也未涉及基于智能体的增量图构建或对论文所关注的多跳问答与智能体范式的实验比较。互补性：其对本体管理、溯源、安全与生命周期治理的讨论可为论文中图记忆流水线的设计与部署约束（例如隐私与访问控制）提供参考。

Writing and related work

English

Position this paper in related-work as an engineering and governance reference rather than an empirical baseline. Cite it for practical concerns (ontology management, provenance, access control) when arguing for production-readiness of persistent graph memory, and note that its claims require operationalization and experiments to validate retrieval/reasoning benefits.

中文

在相关工作中将该论文作为工程与治理方面的参考，而非实证基线。引用其在本体管理、溯源与访问控制上的实践性建议，以支持持久化图记忆的生产就绪性论点，并指出其主张需要通过实作与实验来验证对检索/推理性能的实际影响。

Method and evaluation

English

Adopt components of the architecture when implementing the thesis prototype: integrate explicit provenance metadata for extracted triples, enforce access-control and ownership tags on graph nodes/edges, and use an ontology-management layer to constrain relation types. For evaluation, measure how provenance, governance checks, and access-control filtering affect multi-hop retrieval accuracy, bridge-entity identification, and end-to-end latency compared to unsecured graph memory and vector RAG baselines.

中文

在实现论文原型时可采用该架构的组件：为抽取的三元组集成显式溯源元数据，为图节点/边加上访问控制与所有权标签，并使用本体管理层约束关系类型。评估方面，应衡量溯源、治理检查与访问控制过滤对多跳检索准确性、桥实体识别及端到端延迟的影响，并将其与不受约束的图记忆和向量RAG基线比较。

Future directions

English

Extend the architecture’s ideas to an agentic, incrementally assembled graph memory by specifying interfaces between intelligent agents and the governance layer, and empirically evaluate trade-offs between governance constraints and retrieval/reasoning performance on benchmarks such as MultiHop-RAG. Investigate automated policies for when agents may add or redact graph facts under access and provenance constraints.

中文

将该架构思想扩展到基于智能体的增量组装图记忆，明确智能体与治理层之间的接口，并在如 MultiHop-RAG 的基准上实证评估治理约束与检索/推理性能之间的权衡。研究在访问与溯源约束下，智能体添加或编辑图事实的自动化策略。

---

02 · IMicAP: an intelligent microbial application platform for knowledge-driven microbiome research

IMicAP：用于知识驱动微生物组研究的智能微生物应用平台

Research record	Details
Authors	Chaoyu Zhu, Xing Wang, Lihui Feng, Mingzhang Xu, Zhengkun Huang, Weijie Chen, Lei Liu
Published	2026-10-01
Sources	openalex
Focus	knowledge graph construction, LLM-based triple extraction, microbiome knowledge integration, multi-hop reasoning 知识图谱（KG）构建, 基于大型语言模型的三元组抽取, 微生物组知识整合, 多跳推理

English

Relevant for practical examples of heterogeneous KG population, biomedical LLM extraction, and QA system engineering, but it is an application paper without explicit treatment of agentic graph memory, incremental assembly, or detailed entity-resolution experiments central to the thesis.

中文

对异构 KG 填充、生物医学 LLM 抽取与问答系统工程提供了实用示例，但属于应用型工作，未在摘要中针对智能体图记忆、增量装配或论文核心的实体消歧实验展开具体论述，故建议略读。

Research synopsis

English

The paper presents IMicAP, an AI-driven, open-access web platform integrating 12 microbial databases and over 35 million articles to support microbiome research. IMicAP offers five modules—Knowledge Query, Knowledge Graph for multi-hop reasoning and hypothesis generation, Intelligent Question-Answering, Genome Browser, and 16S rRNA sequencing analysis—built from curated microbe–disease and microbe–small molecule relations extracted with biomedical language models. The platform emphasizes context-aware exploration of microbial relations and claims to address fragmented data, limited disease scope, and incomplete metabolomic coverage in prior resources. Future work aims to expand coverage and improve reasoning scalability.

中文

该文介绍了 IMicAP，一个面向微生物组研究的 AI 驱动开放访问 Web 平台，整合了 12 个微生物数据库与超过 3500 万篇文献。IMicAP 提供五大模块：知识查询、用于多跳推理与假设生成的知识图谱（KG）模块、智能问答模块、基因组浏览器和 16S rRNA 测序分析模块，所用微生物—疾病与微生物—小分子关系由生物医学语言模型抽取并经过人工整理。平台旨在支持语境感知的微生物关系探索，解决数据分散、疾病范围受限与代谢组信息缺失等问题，并计划扩展覆盖范围与提升推理可伸缩性。

Thesis relevance

English

Overlap: IMicAP is directly relevant on KG construction, LLM-based triple extraction, and multi-hop reasoning—areas the thesis cites as highly relevant. Differences: IMicAP is an applied, domain-specific platform assembling a KG from external databases and literature, whereas the thesis focuses on incremental, agent-driven graph memory constructed from the system’s own retrieval and inference activity. Limitations: the abstract does not describe agentic RAG, persistent agent memory, incremental graph assembly, or empirical comparisons of entity-resolution strategies and agentic paradigms that are central to the thesis.

中文

重合点：IMicAP 在 KG 构建、基于大型语言模型的三元组抽取与多跳推理方面与论文主题直接相关，这些均列为论文的高相关研究方向。差异：IMicAP 是面向领域的应用型平台，主要从外部数据库与文献构建 KG，而论文侧重于由智能体检索与推理活动增量构建的图记忆。局限：摘要未说明 agentic RAG、持久化的智能体记忆、增量图装配，或论文关注的实体消歧策略与多种智能体范式的实证比较。

Writing and related work

English

Position IMicAP as an application case for KG population and domain-specific LLM extraction methods rather than a methodological advance on agentic graph memory. Cite it for engineering choices in heterogeneous data integration, biomedical extraction pipelines, and QA interfaces, but verify extraction provenance and accuracy before using it as a methodological precedent.

中文

将 IMicAP 视为 KG 填充与领域化 LLM 抽取方法的应用案例，而非关于智能体图记忆的理论性方法贡献。可在异构数据整合、生物医学抽取流水线与问答接口的工程选择上引用，但在将其作为方法论先例前应核验抽取的溯源与准确性。

Method and evaluation

English

Examine IMicAP’s reported pipeline for triple extraction and provenance tracking to inform the thesis’s KG population and confidence-scoring design. Compare a prebuilt, curated KG (as in IMicAP) versus the thesis’s incremental agent-assembled graph by measuring multi-hop retrieval accuracy, bridge-entity recovery, and latency when reusing previously discovered paths. If possible, obtain or recreate subsets of IMicAP’s relations to test entity-resolution strategies (embedding similarity vs. LLM-as-judge) in a domain-specific setting.

中文

查阅 IMicAP 的三元组抽取流程与溯源记录，以借鉴论文中 KG 填充与置信度评分的设计思路。将预构建的、经整理的 KG（如 IMicAP）与论文中由智能体增量组装的图进行对比，通过多跳检索准确率、桥接实体恢复率和重用先前发现路径时的延迟来量化差异。如可能，获取或重建 IMicAP 的部分关系子集，以在该领域中测试不同实体消歧策略（嵌入相似度 vs. LLM 评判者）。

Future directions

English

A natural follow-up is to integrate agentic RAG and persistent query-aware graph memory into IMicAP to study incremental KG growth and reuse in a biomedical domain. Another direction is to benchmark entity-resolution strategies on IMicAP’s extracted relations and measure how structured memory accumulation affects multi-hop QA verifiability and latency.

中文

自然的后续方向是将 agentic RAG 与持久化的查询感知图记忆集成到 IMicAP 中，以研究在生物医学领域中 KG 的增量增长与重用。另一个方向是对 IMicAP 抽取的关系进行实体消歧策略基准测试，并衡量结构化记忆积累对多跳问答的可验证性与延迟影响。

---

03 · MEMO: Multi-Level Entity-Aware Memory for Streaming Video Understanding

MEMO：用于流式视频理解的多层次实体感知记忆

Research record	Details
Authors	Yinying Li, Yuqian Fu, Yulin Dai, Jingyu Gong, Tianwen Qian, Xiaoling Wang
Published	2026-09-30
Sources	openalex
Focus	streaming video understanding, structured memory, entity-level representation, query-time retrieval 流式视频理解、结构化记忆、实体级表示、查询时检索

English

Relevant for ideas about entity-aware, multi-level indexing and selective evidence recall that can inform storage and retrieval design, but the paper focuses on visual streaming rather than incremental KG construction, entity-resolution for textual surface forms, or agentic RAG — so a targeted skim is sufficient to extract architectural lessons.

中文

本文在实体感知的多层索引与选择性证据召回方面具有启发意义，可为存储与检索设计提供参考，但其针对视觉流数据，未涉及增量式 KG 构建、文本表面形式的实体解析或智能体 RAG，因此可有针对性地略读以提取架构性经验。

Research synopsis

English

The paper addresses long-horizon streaming video understanding, where retaining fine-grained visual semantics over long temporal spans is challenging. The authors propose MEMO, a training-free, plug-and-play framework that constructs multi-level, entity-aware structured memory by partitioning the stream into semantically coherent chunks. Each chunk stores lightweight global and entity-level representations as retrieval indices while keeping high-resolution visual content separately for on-demand recall. At inference, MEMO performs query-specific retrieval over these structured memories and selectively retrieves high-resolution evidence for downstream multimodal LLM reasoning. Experiments on StreamingBench and OVO-Bench are reported to improve multiple base models and achieve state-of-the-art performance.

中文

论文针对长时间跨度的流式视频理解问题，强调在远时域保留细粒度视觉语义的困难。作者提出了 MEMO——一个无需训练、可即插即用的框架，通过将视频流划分为语义一致的块来构建多层次、实体感知的结构化记忆。每个块以轻量级的全局与实体级表示作为检索索引，同时将高分辨率视觉内容单独保留以便按需召回。推理时，MEMO 对结构化记忆执行查询特定的检索，选择性地检索高分辨率证据用于下游多模态 LLM 推理。作者在 StreamingBench 和 OVO-Bench 上的实验显示对多种基础模型有一致改进并实现了最新性能。

Thesis relevance

English

Overlap: both works emphasize explicit, entity-centric structured memories and query-time selective retrieval to preserve fine-grained evidence. Differences: MEMO focuses on visual streaming data and chunk-level indices rather than constructing an explicit relational Knowledge Graph (KG) or supporting LLM-driven agentic RAG with persistent graph memory. Limitations relative to the thesis: MEMO does not describe incremental KG construction, typed relations, or extensive entity-resolution across surface forms, so it does not address graph-coherence or multi-hop path traversal evaluation. Complementarity: MEMO’s separation of lightweight indices and retained high-resolution evidence and its entity-aware chunking could inform efficient retrieval and storage design for the thesis’s graph memory.

中文

重合点：两者都强调以实体为中心的结构化记忆与查询时的选择性检索，以保留细粒度证据。差异：MEMO 侧重于视觉流数据和块级索引，而非构建显式的关系型知识图谱（KG）或支持基于 LLM 的智能体检索增强生成（agentic RAG）持久图记忆。相对于论文的局限：MEMO 未描述增量式 KG 构建、有类型的关系或跨异名的实体解析，因此未涉及图一致性或多跳路径遍历的评估。互补性：MEMO 将轻量索引与高分辨率证据分离、以及实体感知的切分方法，可为论文中图记忆的高效检索和存储设计提供参考。

Writing and related work

English

Position MEMO as a systems-level example that argues for entity-aware, structured memory and selective evidence recall. In related work, contrast MEMO’s chunk-and-index approach with symbolic KG-based memory and explicitly note that MEMO is training-free and multimodal-LLM-ready.

中文

将 MEMO 作为强调实体感知结构化记忆与选择性证据召回的系统级示例。在相关工作中对比 MEMO 的块-索引方法与基于符号 KG 的记忆，并明确指出 MEMO 无需训练且可与多模态 LLM 直接配合。

Method and evaluation

English

Consider adopting MEMO’s design of lightweight global and entity-level indices with separate high-resolution evidence storage to reduce retrieval latency while preserving verifiability. Evaluate chunking strategies and index granularity as variables when constructing the graph memory (e.g., chunk → candidate node grouping). Measure the latency/accuracy trade-offs of index-only retrieval followed by on-demand evidence fetch, and test cross-modal entity alignment techniques to bridge visual/textual entity mentions.

中文

可借鉴 MEMO 将轻量级全局与实体级索引与独立的高分辨率证据存储相结合的做法，以在降低检索延迟的同时保留可验证性。在构建图记忆时将切分策略与索引粒度作为可调变量（例如由块到候选节点分组）进行评估。衡量仅索引检索随后按需获取证据的延迟/准确率折中，并尝试跨模态实体对齐方法以连接视觉与文本的实体表述。

Future directions

English

Integrate MEMO-style entity-aware indices with an explicit KG layer that stores typed edges and confidence scores, allowing multi-hop traversal over recalled evidence. Explore cross-modal entity resolution mechanisms to merge visually and textually grounded entity mentions into coherent graph nodes. Experimentally evaluate whether index+on-demand retrieval accelerates agentic reasoning loops in local LLM-based pipelines.

中文

将 MEMO 风格的实体感知索引与存储有类型边和置信度的显式 KG 层相结合，从而支持基于召回证据的多跳遍历。研究跨模态实体解析机制，将视觉和文本基础的实体表述合并为一致的图节点。实验性地评估索引加按需检索是否能加速在本地 LLM 管线中的智能体推理循环。

---