01 · Modern Web Development Using Retrieval-Augmented Generation (RAG): A Comprehensive Review

面向现代 Web 开发的检索增强生成（RAG）综述

Research record	Details
Authors	Rajat Parab, Aryan Sharma, Vishal Shrivastava, Vibhakar Pathak, Rakesh Ranjan
Published	2026-10-06
Sources	openalex
Focus	Retrieval-Augmented Generation, RAG tooling and infrastructure, agentic RAG and Graph RAG, vector databases and embeddings 检索增强生成（RAG）, RAG 工具链与基础设施, 智能体 RAG 与 Graph RAG, 向量数据库与嵌入

English

Recommended for skimming as a broad, up-to-date survey useful for background, tooling choices, and locating recent work on agentic and Graph RAG, but it lacks the focused empirical and methodological depth needed for the thesis’s core research questions.

中文

建议略读：该综述能提供背景、工具选择以及关于智能体 RAG 与 Graph RAG 的最新工作线索，但缺乏对论文核心研究问题所需的针对性实证与方法深度。

Research synopsis

English

This review surveys the application of Retrieval-Augmented Generation (RAG) in modern web development, framing RAG as a remedy to LLMs’ context and hallucination limits. It synthesizes the retrieval pipeline (chunking, embeddings, vector indexing, similarity search), popular vector databases and embedding models, and orchestration frameworks (LangChain, LlamaIndex, Haystack, LangGraph). The paper categorizes recent directions including hybrid search, agentic RAG, Graph RAG, and corrective/adaptive approaches, and discusses engineering trade-offs of latency, retrieval quality and evaluation. It concludes by identifying open research directions for building reliable retrieval-grounded web applications.

中文

该综述评估了检索增强生成（RAG）在现代 Web 开发中的应用，认为 RAG 可以缓解大规模语言模型在上下文窗口与幻觉方面的局限。文章综述了检索流水线（分片、嵌入、向量索引、相似性检索）、主流向量数据库与嵌入模型，以及编排框架（LangChain、LlamaIndex、Haystack、LangGraph）。作者对混合检索、智能体 RAG、Graph RAG 以及自适应/纠正型方法进行了分类，并讨论了延迟、检索质量与评估等工程权衡，最后指出构建可验证检索驱动应用的若干开放方向。

Thesis relevance

English

Overlap: the review explicitly covers agentic RAG and Graph RAG, and lists orchestration frameworks and vector DBs relevant to implementation choices in the thesis. Differences and limitations: as a broad web-development review it is high-level and does not present empirical comparisons of persistent structured graph memory, nor focused evaluations of entity resolution or multi-hop retrieval accuracy. Complementarity: useful background and engineering context for tool selection, taxonomy of hybrid/agentic approaches, and pointers to recent work the thesis should cite and contrast with its empirical contributions.

中文

重合点：该综述明确讨论了智能体 RAG 与 Graph RAG，并列举了对论文实现有参考价值的编排框架与向量数据库。差异与局限：作为面向 Web 的广泛综述，它属于高层概述，未提供对持久化结构化图记忆的实证比较，也没有专门评估实体解析或多跳检索精度。互补性：可作为工具选择与工程背景的参考，提供混合/智能体方法的分类，并指向论文应当引用并与之对比的最近工作。

Writing and related work

English

Position this review as background and engineering context rather than a primary technical source. Cite its taxonomy and tooling survey when justifying infrastructure choices (e.g., vector DB, orchestration framework) and to motivate why a focused empirical study on structured graph memory is needed.

中文

将该综述作为背景与工程实践的参考，而非主要技术依据。在说明基础设施选择（如向量数据库、编排框架）时引用其分类与工具清单，并用以论证为什么需要针对结构化图记忆的聚焦实证研究。

Method and evaluation

English

Use the review to extract candidate tooling, embedding models, and evaluation practices to reproduce or compare baselines (e.g., vector DB configurations, similarity metrics). For evaluation design, adopt its discussion of latency vs. retrieval quality trade-offs and include comparable metrics (end-to-end latency, retrieval accuracy). The paper’s taxonomy of hybrid and agentic variants can help enumerate baseline agentic paradigms for experiments.

中文

可从综述中提取候选工具、嵌入模型与评估实践以复现或比较基线（例如向量数据库配置、相似性度量）。在评估设计上采纳其关于延迟与检索质量权衡的讨论，并包含可比指标（端到端延迟、检索准确率）。其对混合与智能体变体的分类有助于列举实验中的基线智能体范式。

Future directions

English

Follow-up directions include focused empirical comparisons between Graph RAG and pure vector RAG on multi-hop benchmarks, systematic evaluation of entity-resolution strategies in persistent graph memory, and development of standardized metrics for graph coherence and reuse efficiency in agentic settings.

中文

后续方向包括在多跳基准上对 Graph RAG 与纯向量 RAG 进行针对性实证比较，系统评估持久化图记忆中的实体解析策略，以及为智能体场景下的图一致性与重用效率制定标准化度量。

---

02 · Bridge-Aware Reinforced Compositional Exploration and Decomposition for Biomedical Multi-Hop Answering

面向生物医学多跳问答的桥接感知强化组合探索与分解

Research record	Details
Authors	Warodom Phungjununt, Kanabadee Srisomboon, Wilaiporn Lee, Akara Prayote, Luepol Pipanmekaporn
Published	2026-10-06
Sources	openalex
Focus	biomedical multi-hop QA, bridge-aware decomposition, reinforcement-learned planner, typed graph traversal 生物医学多跳问答, 桥接感知分解, 强化学习规划器, 类型化图遍历

English

Relevant for planner training, type-aware traversal, and traceability ideas, but its reliance on a curated PrimeKG and biomedical focus make it less directly applicable to the thesis’s central questions about incrementally assembled graph memory and entity-resolution trade-offs.

中文

该文在规划器训练、类型化遍历与结果可追溯性方面具有参考价值，但依赖经整理的 PrimeKG 且集中于生物医学领域，因此对论文关于增量组装图记忆和实体解析权衡的核心问题而言适用性较弱，建议略读以抽取可借鉴方法。

Research synopsis

English

The paper addresses low precision and unverifiable outputs of closed-book LMs on biomedical two-hop queries. It proposes BRACED (Bridge-Aware Reinforced Compositional Exploration and Decomposition): a planner trained with Group-Relative Policy Optimization that emits sub-queries, followed by a deterministic type-aware two-hop traversal over a curated PrimeKG so that returned entities are traceable to explicit graph paths. On the BioHopR benchmark the method reports micro-F1 0.4241 and macro-F1 0.6381, a large improvement over supervised fine-tuning of the same planner. The authors also report one language-model call per question from an on-prem 7B model and observe two typed hops already cover most gold answers on their dataset.

中文

本文针对闭卷语言模型在生物医学两跳查询上精度低且缺乏可验证证据的问题，提出了 BRACED（桥接感知强化组合探索与分解）。该方法用 Group-Relative Policy Optimization 训练一个规划器以产生子查询，然后在经过整理的 PrimeKG 上进行确定性、类型感知的两跳遍历，使返回的实体可追溯到明确的图路径。在 BioHopR 基准上，方法报告 micro-F1 为 0.4241、macro-F1 为 0.6381，相对于对同一规划器的有监督微调有显著提升。作者还报告使用一次语言模型调用即可得到答案（基于本地 7B 模型），并指出两跳类型化遍历已覆盖数据集的大部分金标准答案。

Thesis relevance

English

Overlap: both works target multi-hop question answering with explicit graph paths and emphasize traceability of answers to graph traversals. Differences: BRACED relies on a curated PrimeKG and a reinforcement-learned planner with deterministic two-hop traversal, whereas the thesis centers on incrementally constructed graph memory assembled from system interactions and studies entity-resolution strategies and agentic RAG paradigms on local models. Complementarity: BRACED’s planner training and type-aware traversal could inform the thesis’s traversal policies and evaluation of planner-vs-agentic approaches, but its reliance on a fixed KG limits direct applicability to incremental, noisy graph-memory scenarios.

中文

重合点：两者都关注多跳问答、将答案可追溯到图遍历路径，并重视基于图的检索与推理。差异：BRACED 依赖经整理的 PrimeKG 和强化学习训练的规划器并采用确定性两跳遍历；而论文主题强调由系统交互增量构建的图记忆，重点研究实体解析策略以及在本地模型上比较不同的 agentic RAG 范式。互补性：BRACED 的规划器训练和类型感知遍历可为论文中遍历策略与规划器对比提供参考，但其基于固定 KG 的假设限制了在增量、噪声图记忆情境中的直接适用性。

Writing and related work

English

Position this paper as a domain-focused, KG-backed approach that highlights planner training and traceability. When discussing related work, contrast BRACED’s curated-KG, two-hop design and single-call efficiency with incremental graph-memory systems and multi-agent strategies.

中文

将此文作为一种以领域知识图谱为后盾、强调规划器训练与可追溯性的工作来定位。在相关工作中对比 BRACED 的经整理 KG、两跳设计与单次调用效率，与增量图记忆系统和多种智能体策略的差异。

Method and evaluation

English

Reproduce the deterministic type-aware two-hop traversal but run it over an incrementally assembled knowledge graph to measure robustness to fragmented entities. Compare the Group-Relative Policy Optimization planner against the thesis’s four agentic paradigms (ReAct, Plan-and-Execute, Self-Ask, Reflexion) in terms of LM calls, latency, and verifiability. Report bridge-entity identification rates and how performance changes as structured memory accumulates.

中文

复现该论文的类型化确定性两跳遍历，但在增量构建的知识图谱（图记忆）上运行，以评估对实体碎片化的鲁棒性。将 Group-Relative Policy Optimization 训练的规划器与论文中四种智能体范式（ReAct、Plan-and-Execute、Self-Ask、Reflexion）在语言模型调用次数、延迟与可验证性方面进行对比。报告桥接实体识别率以及随着结构化记忆累积性能的变化。

Future directions

English

Extend BRACED by replacing the curated PrimeKG with an incrementally constructed graph memory and evaluate entity-resolution strategies under noisy, heterogeneous surface forms. Explore integrating the planner into multi-agent agentic RAG pipelines to trade off single-call efficiency for progressive, verifiable reasoning.

中文

将 BRACED 的 PrimeKG 替换为增量构建的图记忆，评估在噪声和异构表面形式下的实体解析策略。探索将该规划器集成入多智能体的 agentic RAG 管道，以权衡单次调用效率与逐步、可验证推理之间的关系。

---