rag · 知识库活文档

  • 更新：R115: RECAST 证据路由+ExperienceIndex+Memory Portability 3 件 RAG 主分类入档+EAL-Bench Defense 节点+Jay 工业 A/B 数据。

1. 现状全景

R115 现状全景（接续 R114-R82 累计 34 轮）：① 运行时 168+ 件（R114 165+ + R115 +3 件 RAG 主分类 net-new：RECAST 2610.10507 + ExperienceIndex 2610.10091 + Memory Portability 2609.05339，Oct 7-9 提交、Oct 8-9 建卡）② 综述 67+ 件（R114 续立）③ 知识结构 42+ 件 ④ RAG 安全 60+ 面（R114 59+ + R115 +1 件 EAL-Bench 2609.01836 授权漂移评测 5+2 LLM × 3 场景）⑤ Memory 47+ 条腿（R114 45+ + R115 +2 件 ExperienceIndex 2610.10091 + Memory Portability 2609.05339）⑥ RAG 71+ 范式（R114 69+ + R115 +2 件 RECAST 2610.10507 + Memory Portability 2609.05339，ExperienceIndex 主分类 agent 不计）⑦ 多模态 RAG 45+（R114 续立）⑧ 63 维张力（R114 62 维 + R115 新增 1 维：RAG 记忆系统可移植性维度 Memory Portability + 续立 8 维）⑨ 范式辩论 R87-R115 累计 32 极 ⑩ 检索精度/脆弱性/评测体系多轨续立 ⑪ 协议级门控 R108-R115 7 轨 ⑫ R114 fresh web_search 70-90% 失败率多源印证续立 + R115 新增 Jay Oct 9 CSDN 工程实践数据 + Substack theaiengineer 2026 Memory 三层架构印证续立。

R115 增量密度「中」——3 件 RAG 主分类 net-new paper_card（RECAST 2610.10507 + ExperienceIndex 2610.10091 + Memory Portability 2609.05339，Oct 7-9 arXiv 提交、Oct 8-9 建卡）+ 1 件 Defense 轴邻接级正式入档（EAL-Bench 2609.01836，授权漂移评测基准）+ 1 件 Jay CSDN Oct 9 工业视角（Agentic RAG + 多模态 RAG + 12-Factor Agents + MCP/A2A 工程实践数据）+ 1 件 Substack theaiengineer 2026 Memory 三层架构续立 + 1 次 fresh web_search（Progress Software Agentic RAG + Lyzr + Squirro + Atlan + Turingpost 续立 + 60+ 格式 + Agentic RAG SoK）。

OpenClaw 选型 v3.45 65 维 + 八角张力 63 维（R115 +1 维）。work-queue 2026-10-09 12:00 确认：RAG 主题无高价值待深度解读缺口（Top 15 全空 · 3 件均为非 RAG 主分类）；待更新/缺失主题活文档 = 0（全部最新）；RAG 主题增量为中密度活文档滚动更新而非紧急缺口响应。

2. 关键工作与脉络

2.1 R115 增量：3 件 RAG 主分类 net-new + 1 件 Defense 轴邻接级正式入档 + 1 件 Jay 工业视角 + 1 件 Substack 续立

2.1.1 RECAST · arXiv:2610.10507v1 通过自适应证据路由学习计算正确的上下文

  • 来源：paper_cards/1717-2610-10507.md（Oct 8-9 建卡 · Oct 7 arXiv 提交）+ inbox/tom/2026-10-09-agent-rag-longcontext-radar.md 高价值 #1
  • paper_card 1717 · 主分类 rag ✅（RAG 主分类 net-new · R114 无此条目）· 副分类 agent · 形态 method · 标签 rag agent benchmark systems
  • 作者：Yilun Hao, Craig Boutilier 等（Google DeepMind）
  • TLDR 锚点：Large language models are increasingly applied to tasks grounded in long, heterogeneous information sources. Conventional Retrieval-Augmented Generation (RAG) relies on fixed similarity-based retrieval, while agentic variants adapt queries and tool use but remain largely retrieval-centric. However, in many tasks, the evidence required for a solution is not explicitly present in any single source item. Instead, it must be derived through filtering, aggregation, or computation across multiple source items. In this work, we introduce RECAST (Routing Evidence through Computation, Access, and Synth[esis])...
  • 方案：① 自适应证据路由（adaptive evidence routing）② 学习式策略（learnable policy）③ 联合优化三要素：计算工具（computation）+ 访问顺序（access）+ 综合方式（synthesis）
  • 意义：落在 Reasoning 轴（多跳/结构化推理）+ Efficiency 轴（智能路由）；改变"固定相似度检索"二分格局——从策略学习层面统一"计算+访问+综合"三要素；与 R114 EC-RAG 2610.08674（事件链结构化推理）+ R112 EnSI-RAG 2608.21252（实体结构索引）+ R112 MatRAG 2610.01767（多跳 QA）共同构成 "RAG Reasoning 轴结构化推理四节点"
  • ⚠️ 限制：具体 benchmark 数字（与其他 multi-doc RAG 方法的详细对比）需读原文确认；与 R114 UNREAL 2610.08463（统一表征层）的边界互补性需进一步研究（UNREAL 表征层统一 vs RECAST 策略层路由）
  • 可信度：⭐⭐⭐⭐（Google DeepMind 出品；arXiv v1 新鲜；策略学习框架有学术贡献）

2.1.2 ExperienceIndex · arXiv:2610.10091v1 基于 artifact 经验积累的索引框架

  • 来源：paper_cards/1716-2610-10091.md（Oct 8-9 建卡 · Oct 7 arXiv 提交）+ inbox/tom/2026-10-09-agent-rag-longcontext-radar.md 高价值 #2
  • paper_card 1716 · 主分类 agent ✅ · 副分类 rag（强邻接）· 形态 method · 标签 agent memory
  • 作者：Peter Baile Chen, Jacob Andreas, Samuel Madden 等（MIT / Microsoft / UPenn）
  • TLDR 锚点：Knowledge-intensive tasks require answering many questions by reasoning about a shared corpus of artifacts (e.g., court cases, or scientific literature). As humans interact with these corpora, they naturally accumulate experiential knowledge about artifacts, enabling them to quickly identify the complete set of relevant artifacts for each new task. However, existing AI agents lack appropriate memory solutions to build or reuse such artifact-grounded experience, leading to lower answer quality and higher online cost. Existing memory solutions extract and reuse information from prior task-solv[in[g]]
  • 方案：① artifact-grounded 经验积累（哪些 artifact 值得检索）② 在线质量提升 + 成本下降 ③ 首个系统解决"agent 跨任务 artifact 经验积累"的工作
  • 意义：落在 交互性轴（Memory 范式第三条路补强）+ Reasoning 轴（结构化经验驱动检索）；与 R114 δ-mem 2605.12357（第三条路 8×8 OSAM）+ R114 UNREAL 2610.08463（统一 RAG/Long-Context 表征）+ R115 RECAST 2610.10507（策略路由）共同构成 "RAG 记忆层四路径"——经验积累 + 固定矩阵 + 模型内部统一 + 策略路由；与 Source Learning 2610.02150（从检索到能力）同属"经验驱动"方向
  • ⚠️ 限制：主分类 agent（不算 RAG 主分类 net-new 数字）；具体 benchmark 数字（质量提升百分比、成本下降幅度）需读原文确认
  • 可信度：⭐⭐⭐⭐（Jacob Andreas 团队出品；MIT/Microsoft/UPenn 联合出品）

2.1.3 Memory Portability · arXiv:2609.05339v1 Agent 记忆在模型升级后的可移植性受控研究

  • 来源：paper_cards/1238-2609-05339.md（Oct 9 确认入库 · 2026-09 arXiv 提交）+ inbox/tom/2026-10-09-rag-e1prep.md 增量 4
  • paper_card 1238 · 主分类 rag ✅（RAG 主分类 net-new · R114 无此条目）· 副分类 agent · 形态 method · 标签 rag agent
  • 作者：未知（academy.dair.ai 收件 + S2 enrich）
  • TLDR 锚点：Findings highlight the necessity of direction-specific migration testing, strict embedding space isolation, and the retention of source histories for memory repair in memory migrations.
  • 方案：① 方向特定的迁移测试（direction-specific migration testing）② 严格的 embedding space 隔离（strict embedding space isolation）③ 源历史保留用于 memory repair（source histories retention）
  • 意义：落在 交互性轴（Memory 系统可移植性维度新增）+ Defense 轴（embedding space 隔离 ≈ 零信任）；与 R114 δ-mem 2605.12357（在线记忆状态矩阵）+ ExperienceIndex 2610.10091（artifact 经验积累）+ Agentic-ZTA 2610.05782（零信任嵌入）共同构成 "RAG Memory/Defense 轴记忆系统四维"——固定矩阵 + 经验积累 + 可移植性 + 零信任嵌入；R115 新增"RAG 记忆系统可移植性维度"
  • ⚠️ 限制：具体受控研究的实验规模、具体模型升级路径（哪个底模升级到哪个底模）、memory repair 具体方法论需读原文确认
  • 可信度：⭐⭐⭐⭐（受控研究方法论严谨；具体三个关键发现：方向特定迁移测试 + embedding space 隔离 + 源历史保留）

2.1.4 EAL-Bench · arXiv:2609.01836v1 Agent 记忆是内生授权洗钱的表层载体

  • 来源：paper_cards/1187-2609-01836.md（Oct 9 确认入库 · 2026-09 arXiv 提交）+ inbox/tom/2026-10-09-rag-e1prep.md 增量 3
  • paper_card 1187 · 主分类 agent · 副分类 rag · 形态 method（benchmark 形态评分）· 标签 agent security benchmark · S2被引 3 · OpenAlex W7207550553 · DOI 10.48550/arxiv.2609.01836
  • TLDR 锚点：This work evaluates five LLMs as memory writers and two as executors across procurement, cybersecurity, and finance and introduces EAL-Bench, which measures how accurately persistent memory preserves evolving authorization state and whether errors propagate to downstream unauthorized actions.
  • 方案：① 5 LLM 记忆写入者 + 2 LLM 执行者 ② 3 场景（采购/网络安全/金融）③ EAL-Bench 测量"持久记忆保留授权状态演化的准确性 + 错误是否级联传播至下游未授权操作"
  • 意义：落在 Defense 轴（授权状态完整性评测基准新增）+ 交互性轴（Memory 系统授权安全）；与 R114 RAG-PIBench 2610.08571（提示注入检测基准）+ R113 Bounded Provisional 2610.05826（入库前时间窗口防御）+ R113 RAGScope 2609.39075（统计证据门控）+ R113 Agentic-ZTA 2610.05782（RAG 嵌入零信任）+ R112 Trojan Hippo Bench + R112 VirusCascade + R113 imMRAG 2610.01871（在库黑盒数据窃取）共同构成 "RAG Defense 轴九节点评测体系"（R115 新增第 9 节点：授权漂移评测 EAL-Bench）；与 RAG-PIBench 形成"提示注入检测 + 授权漂移检测"双评测基准
  • ⚠️ 限制：主分类 agent（不算 RAG 主分类 net-new 数字）；具体 F1/AUC 数字、实验规模需读原文确认；与 RAG-PIBench（提示注入检测）的边界与互补性需进一步研究
  • 可信度：⭐⭐⭐⭐（benchmark 形态；EAL-Bench 具体场景+量化方法；S2被引 3）

2.1.5 Jay CSDN Oct 9 工程视角（R115 工业视角续立）

  • 来源：inbox/jay/2026-10-09T0820-jay-csdn-rag-agentic-multimodal-substack.md
  • 核心 4 条： 1. 手写 RAG 检索增强生成系统（CSDN/AtomGit 源码）—— 完整流水线 文档加载→分块→向量化→向量存储(FAISS)→混合检索(BM25+向量)→重排序→LLM 生成→对话记忆管理；chunk_size 256-512；chunk_overlap 10%-20%；⚠️ 混合检索召回率 +17%（待核实）；RAG 评估三件套 Hit Rate@K + Faithfulness + Answer Relevance；LLM-as-Judge；生产部署三路径（FAISS+BGE → LangChain+Chroma+GPT → Milvus+vLLM） 2. 企业知识库问答系统（Milvus 2.3 + bge-large-zh-v1.5 + 通义灵码 + 阿里云 OSS）—— P95 <50ms；350 token/s；知识查询 15 分钟→28 秒；混合检索（0.7×向量 + 0.3×关键词）；A/B 成本实测（5TB、日均 1 万次查询）：月 6305→4200 元（↓33%） 3. Agentic RAG ADK 框架—— 四阶段工作流 查询重写→智能检索→质量评估→条件分支；查询重写（规则+LLM 双路）；多路召回（BGE-M3 + jieba BM25 + RRF + Cross-Encoder）；质量评估（4 维度 · 80% 阈值 PASS/FAIL）；条件分支（质量达标直接生成 + 不达标触发 SerpAPI 补救） 4. 多模态 RAG 工程化 Pipeline（Dify + CLIP + Whisper + Qwen-VL）—— CLIP ViT-B/16 轻量化（参数量 ↓37% · 推理延迟 ↓28%）；Whisper 流式（3-5 秒窗口 + 1.5 秒重叠）；FAISS + HNSW 混合索引；A/B 测试：Top-1 准确率 58.3%→82.5%（+24.2pp）；跨模态召回率 61.7%→89.2%（+27.5pp）
  • 意义：与 R114 EC-RAG 2610.08674 + R114 UNREAL 2610.08463 形成"学术方法 + 工程实测"双轨印证；A/B 数据（+24.2pp / +27.5pp / ↓33% 成本）是 R115 工业视角的最重要数字
  • ⚠️ 重要限制：混合检索召回率 +17% / 通义灵码 +23% 两个数字 R113-R115 三次出现均"⚠️ 待核实"；禁止引用为事实
  • 可信度：⭐⭐⭐（工业实践；A/B 数据有数字支撑）

2.1.6 Substack theaiengineer 2026 Memory 三层架构（R115 续立）

  • 来源：inbox/tom/2026-10-09-agent-rag-longcontext-radar.md Substack 线索（theaiengineer.substack.com）+ R114 fresh web_search 续立
  • 核心洞察： 1. Memory 架构演进：2026 年 memory 已从"RAG+向量库"演化为三层架构——in-context memory（上下文窗口内）+ agentic memory management（Agent 驱动管理）+ shared memory infrastructure（共享记忆基础设施） 2. 核心玩家：Mem0 / Letta / Zep 已成 agent memory 赛道核心玩家 3. 新基准：LongMemEval 成为 memory 评测新基准（替代旧有的 Memorybank 等基准） 4. 关键洞察：上下文窗口变大并没有消灭 memory 需求，而是改变了"什么放 context、什么放检索"的权衡——与 R114 δ-mem 论文结论一致
  • 意义：与 R114 δ-mem 2605.12357（在线记忆状态矩阵）+ R114 UNREAL 2610.08463（统一 RAG/Long-Context 表征）+ R115 ExperienceIndex 2610.10091（artifact 经验积累）+ R115 Memory Portability 2609.05339 形成"学术定义+工业架构"双轨印证；Mem0/Letta/Zep 三层架构与 R114 §2.9 Memory 活文档对齐
  • 可信度：⭐⭐⭐（Substack 技术专栏；行业趋势判断）

2.2 R114-R112 锚点续立（摘要）

R114：RAG-PIBench arXiv:2610.08571v1（Defense）+ UNREAL arXiv:2610.08463v1（Efficiency）+ EC-RAG arXiv:2610.08674v1（Reasoning）+ Agentic AutoRAG arXiv:2610.08452v1（Efficiency+Interactivity）+ δ-mem arXiv:2605.12357v2 NeurIPS 2026（强邻接 · DOI 10.48550/arxiv.2605.12357）+ RAG Foundation Survey arXiv:2312.10997（Gao et al. S2被引 4207）+ DeCoPrune arXiv:2609.39096/Structuring MoE arXiv:2610.07332/Sensor-Language-Action arXiv:2610.08244/DMAD arXiv:2610.02188 4 件邻接级。

R113：Bounded Provisional Visibility 2610.05826v1（Defense · ⭐⭐⭐⭐）+ AI-Decision Checkpoints 2610.06207v1（BPM · ⭐⭐⭐）+ Source Learning 2610.02150v1（agent·rag · ⭐⭐⭐⭐）+ Agentic-ZTA 2610.05782v1（RAG ZTA · ⭐⭐⭐⭐）+ When Does Selection Replace Extraction 2609.34227（Jev typed · ⭐⭐⭐⭐）+ UndoBench 2610.05622（⭐⭐⭐⭐）+ jay 10-07 CSDN 工业视角。

R112：Mapping the RAG Landscape 2610.01936 + Walking the Embedding Space 2610.01871 + MatRAG 2610.01767 + Temporal Validity 2606.26511 + AutoMem 2607.01224 + A-TMA 2607.01935 + ZooClaw-FashionSigLIP2 2606.27708 七件 + 9 件邻接级。

2.3 R115 工业视角 fresh 信号补强（续立）

R114 续立：70-90% RAG 失败率（Ragaboutit/Blits.ai/dev.to）+ 88% AI Agent 失败（Digital Applied）+ 60% 新部署含评测（starmorph 2026 vs 30% in 2025）+ Re-Ranking 25-40% + Stanford 10%+ 幻觉率 + Atlan text-to-SQL +38%。

R115 新增：① Jay CSDN Oct 9 工程实测（手写 RAG 流水线 + 企业 RAG Milvus 2.3 5TB 月成本 6305→4200 元 ↓33% + Agentic RAG ADK 框架 + 多模态 RAG Dify A/B Top-1 58.3%→82.5% +24.2pp / 跨模态召回 61.7%→89.2% +27.5pp + 12-Factor Agents）② Substack theaiengineer Memory 三层架构续立（见 §2.1.6）③ RAG Reimagined 5 Breakthroughs（2026 主方向 = Agentic RAG）④ Google Gemini Embedding 2 ⑤ Progress Software Agentic RAG 2026 AI Excellence Award 续立 ⑥ RAGPerf arXiv:2603.10765v1 续立 ⑦ fresh web_search 5 源（Progress+Lyzr+Squirro+Atlan+Turingpost · 60+ 格式 · 2026 主导模式 Agentic RAG）。

3. 共识、争议与开放问题

3.1 共识 219-222（R115 新预备）+ 共识 215-218（R114 续立）+ 共识 212-214（R113 续立）+ 共识 209-211（R112 续立）

219. RAG 自适应证据路由与策略学习共识预备（R115 新预备 · RECAST 2610.10507 + EC-RAG 2610.08674 + EnSI-RAG 2608.21252 + MatRAG 2610.01767 → "自适应路由+事件链+实体结构+多跳"四路径结构化推理）。

220. RAG 记忆系统可移植性与生产模型升级共识预备（R115 新预备 · Memory Portability 2609.05339 三个关键发现 → RAG 生产系统升级模型时的强制流程）。

221. RAG 经验驱动检索与跨任务 artifact 共识预备（R115 新预备 · ExperienceIndex + Source Learning → RAG 检索"静态规则"→"经验驱动 + 持续学习"）。

222. RAG Defense 轴九节点标准化体系（EAL-Bench 加入）共识预备（R115 新预备 · EAL-Bench + RAG-PIBench + Bounded Provisional + imMRAG + RAGScope + Agentic-ZTA + Trojan Hippo Bench + VirusCascade + Reasoning-Language Alignment → RAG Defense 轴"五节点"→"九节点"标准化体系）。

R114-R112 共识续立：R114 215-218（表征层统一证据选取 / 提示注入检测标准化 / 视频事件链结构化推理 / Agentic 超参可解释）+ R113 212-214（持续 ingestion 防御 / BPM / 评测从可选到必选）+ R112 209-211（多模态检索质量 / 多语言鲁棒性 / 记忆成本工程化）。

3.2 争议 191（R115 新预备）+ 争议 190（R114 续立）+ 争议 188-189（R113 续立）+ 争议 186-187（R112 续立）

191. RAG 表征层统一 vs 策略路由 vs 第三条路 vs 经验积累四路径边界争议预备（R115 新预备 · UNREAL 2610.08463 表征层统一 vs RECAST 2610.10507 策略路由 vs δ-mem 2605.12357 第三条路 8×8 OSAM vs ExperienceIndex 2610.10091 经验积累 → 四条路径的边界与互补性仍未厘清；不同应用场景的优劣待验证）。

R114-R112 争议续立：R114 190（70-90% 失败率口径统一）+ R113 188-189（协议级 vs 模型级 / 统计 vs 宣传）+ R112 186-187（多模态 vs 单模态 / 多语言对齐 vs 统一协议）。

3.3 开放 261-263（R115 新预备）+ 开放 259-260（R114 续立）+ 开放 257-258（R113 续立）+ 开放 254-256（R112 续立）

261. RAG 自适应证据路由学习样本效率开放预备（R115 新预备）：RECAST 2610.10507 学习式策略 → 策略学习的样本效率与可解释性是否能推广到 RAG 检索的更多场景（嵌入模型选型 / chunk 策略 / reranker 训练）？

262. RAG 记忆系统迁移工程化开放预备（R115 新预备）：Memory Portability 2609.05339 三个关键发现（方向特定迁移测试 + embedding space 隔离 + 源历史保留）→ RAG 生产系统升级模型时的标准流程尚未形成；不同底模间的迁移成本与精度损失未量化。

263. RAG Defense 轴九节点标准化体系工业采纳率开放预备（R115 新预备）：EAL-Bench 加入后九节点标准化体系（提示注入 + 授权漂移 + 入库前 + 在库黑盒 + 协议级门控 + 零信任嵌入 + Trojan Hippo + VirusCascade + Reasoning-Language Alignment）→ 工业部署中哪些节点被广泛采用？哪些仍停留在学术研究阶段？

R114-R112 开放续立：R114 259-260 + R113 257-258 + R112 253-256。

3.4 趋势 328-330（R115 新预备）+ 趋势 325-327（R114 续立）+ 趋势 322-324（R113 续立）+ 趋势 318-321（R112 续立）

328. RAG 自适应证据路由与策略学习工业化趋势预备（R115 新预备）：RECAST 2610.10507 + Agentic AutoRAG 2610.08452 + Agentic RAG SoK 2026 → RAG 系统从"固定相似度检索"向"自适应策略学习"演进；Agentic RAG 在企业 AI 平台成为 2026 主导模式（Atlan+Lyzr+Squirro+Progress Software+Turingpost 多源印证）。

329. RAG 记忆系统可移植性与生产模型升级标准化趋势预备（R115 新预备）：Memory Portability 2609.05339 + δ-mem 2605.12357 + ExperienceIndex 2610.10091 + theaiengineer 三层架构 → 2026-2027 RAG 记忆系统"可移植性"成为生产系统标配；模型升级标准流程（方向特定迁移测试 + embedding space 隔离 + 源历史保留）将进入工业实践。

330. RAG Defense 轴九节点标准化体系成形趋势预备（R115 新预备）：EAL-Bench 2609.01836 授权漂移 + RAG-PIBench 2610.08571 提示注入检测 + Bounded Provisional 2610.05826 入库前防御 + imMRAG 2610.01871 在库黑盒 + RAGScope 2609.39075 协议级门控 + Agentic-ZTA 2610.05782 零信任嵌入 + Trojan Hippo Bench + VirusCascade + Reasoning-Language Alignment 2610.03136 → RAG Defense 轴从"五节点"→"九节点"标准化体系成形；OWASP LLM Top 10 + Agentic AI Security 工程化映射进入体系化阶段。

R114-R112 趋势续立：R114 325-327（表征层统一与轻量化 / Defense 五节点 / 多模态结构化推理深化）+ R113 322-324 + R112 317-321。

4. State-of-the-Art 综述（R115 · 多轮深综合）

4.1 现状全景

2026 年 RAG 从"search enhancer"升级为"企业 AI 架构基石 + Agent Memory 第一原语 + Memory 系统可移植性新维度"。R115 工业视角升级（接续 R114）：① 70-90% 失败率多源印证 + R114 争议 190 续立 ② 60% 新部署含系统化评测（starmorph 2026 vs 2025 30%）③ Re-Ranking 25-40% ④ Stanford 10%+ 幻觉率 + Atlan +38% ⑤ UndoBench 故障恢复解耦 ⑥ Bounded Provisional + imMRAG 双节点 ⑦ AI-Decision Checkpoints + Context Engineering as OS ⑧ Source Learning ⑨ R114 新增：RAG-PIBench 4,876/F1=0.896 + UNREAL<500K + EC-RAG 视频事件链 + AutoRAG failure attribution 10v30 + δ-mem 2605.12357 NeurIPS 26 arXiv 入档 + RAG Foundation Survey 2312.10997 4207 S2 被引 ⑩ R115 新增：RECAST 自适应证据路由 + ExperienceIndex artifact 经验积累 + Memory Portability 模型升级可移植性 + EAL-Bench 授权漂移评测 + Jay CSDN 工业实测数据（Top-1 +24.2pp / 跨模态召回 +27.5pp / 成本 ↓33%） + Substack theaiengineer 2026 Memory 三层架构。

学术侧，R109 Mapping the RAG Landscape 2610.01936 提出四轴分类法（效率/防御/交互/推理）。R111 Q139 闭合确认 RAG Landscape 团队（VIT+MBZUAI）≠ ImmRAG 团队（Maria Carmen Jica et al.）——跨洲/跨研究圈；Defense 轴"轴+实证"独立支撑。R114 锚入 RAG Foundation Survey 2312.10997 Gao et al. 4207 S2 被引形成"RAG 综述 1+1"双锚——2023 三阶段范式（Naive/Advanced/Modular）+ 2026 四轴分类（效率/防御/交互/推理）。R115 进一步强化"RAG 综述双锚"在 RAG 范式演进中的学术权威性。

R115 学术增量：① RECAST 2610.10507v1（Google DM · 自适应证据路由 · Reasoning 轴新节点）② ExperienceIndex 2610.10091v1（MIT/MS/UPenn · artifact 经验记忆 · 交互性轴新节点）③ Memory Portability 2609.05339v1（受控研究 · 交互性轴新维度）④ EAL-Bench 2609.01836（授权漂移评测 · Defense 轴第 9 节点）⑤ Jay CSDN Oct 9 工程实测（Top-1 +24.2pp / 跨模态召回 +27.5pp / 成本 ↓33%）⑥ Substack theaiengineer Memory 三层架构续立 ⑦ 5 fresh web_search 来源（Progress+Lyzr+Squirro+Atlan+Turingpost · 2026 主导模式 Agentic RAG）。

4.2 关键工作脉络（按四轴 + 三层）

效率轴（R115 续立 + +2 + RECAST 策略路由）：① R112 MatRAG 2610.01767；② R107 Periodic Weak Spots 2609.36322；③ R106 LazyGraphRAG（Microsoft 0.1% 成本）；④ R106 BM25 Wins at Scale（23,088 queries）；⑤ R104 Corpus2Skill 2604.14572 + Knowledge-as-Skill 2609.25991；⑦ R110 Pipeline/Agentic/GraphRAG/LazyGraphRAG 选型决策树；⑧ R110 δ-mem 第三路径备选 → R114 δ-mem 2605.12357 NeurIPS 2026 正式入档；⑨ R111 Agent Stack 2026 Agent Guardrails；⑩ R112 CLIMB 2610.03421 + PostgreSQL-V 2.0 2608.15994 / ACRONYM 2609.03712 / Aker + pgvector 0.8.0 + Vector Databases Are Dying；⑪ R113 Bounded Provisional 2610.05826 协议级门控新范式；⑫ R114 UNREAL 2610.08463 <500K model-native 统一 RAG/Long-Context；⑬ R114 Agentic AutoRAG 2610.08452 推理驱动 RAG 超参多目标优化；⑭ R115 RECAST 2610.10507 策略路由与自适应证据路由（Google DM）。效率轴共识：架构层（LazyGraphRAG）→ 表征层（MatRAG）→ 推理控制（CLIMB）→ 协议级门控（Bounded Provisional）→ 分层 token 经济（OpenViking）→ VecDB 开源替代（pgvector 0.8.0）→ 模型内部统一证据选取（UNREAL <500K）→ 第三条路 8×8 OSAM（δ-mem 4.87M/0.12%）→ Agentic 超参优化（AutoRAG failure attribution）→ 自适应策略路由（RECAST）。

防御轴（R115 深化 + +1 形成九节点标准化）：① R108 RAGScope 2609.39075；② R109 Walking the Embedding Space 2610.01871；③ R109 AAAI 2026 MrM；④ R108 RoPE 2609.39929；⑤ R110 OWASP LLM Top 10 2026 映射；⑥ R111 Q139 闭合 + Agent Stack 2026 Agent Guardrails；⑦ R112 Trojan Hippo Bench + VirusCascade + Trustworthy Agentic AI + Reasoning-Language Alignment；⑧ R113 Bounded Provisional 2610.05826 入库前防御；⑨ R113 Agentic-ZTA 2610.05782 RAG 嵌入零信任；⑩ R114 RAG-PIBench 2610.08571 提示注入检测基准 4,876 + DistilBERT F1=0.896；⑪ R115 EAL-Bench 2609.01836 授权漂移评测 5+2 LLM × 3 场景。防御轴共识：关键词规则 → 统计证据门控（RAGScope）→ 零散防御 → OWASP 系统化映射 → R114 五节点标准化体系（Bounded Provisional + Trojan Hippo Bench + VirusCascade + Reasoning-Language Alignment + Trustworthy Agentic AI + Agentic-ZTA + RAG-PIBench）→ R115 九节点标准化体系成形（+ EAL-Bench 授权漂移 + imMRAG 在库黑盒）→ "提示注入检测 + 授权漂移检测"双评测基准。

交互性轴（R115 续立 + +3 形成十一轨竞争）：① R112 Mem0 v3 SDK ADD-only + LoCoMo 91.6%；② R101 MemoryAthena 1505；③ R109 AutoMem 2607.01224（2-4×）；④ R102 Superposition 2609.29845；⑤ R99 Designer-RSI 2609.22086 + δ-mem + SELF-INDEX；⑥ R106 JAM + Stashbird + EngramRAG + R112 JAM v2；⑦ R110 记忆三层架构；⑧ R111 Agent Stack 2026 L1/L2/L3；⑨ R112 Mem++ 2610.02002 + OpenViking / VikingRAG / VikingMem；⑩ R113 Jev 选 vs 抽取 memory 2609.34227 预注册研究；⑪ R113 AI-Decision Checkpoints 2610.06207 BPM 设计方法论；⑫ R114 δ-mem 2605.12357 NeurIPS 2026 第三条路 8×8 OSAM 4.87M/0.12%；⑬ R114 DeCoPrune 2609.39096 KV-Cache 压缩 + Structuring MoE 2610.07332 expert routing-agent 操作自然对齐；⑭ R115 ExperienceIndex 2610.10091 artifact 经验积累（MIT/Microsoft/UPenn · Jacob Andreas 团队）；⑮ R115 Memory Portability 2609.05339 模型升级可移植性受控研究；⑯ R115 theaiengineer 2026 Memory 三层架构（in-context + agentic + shared infrastructure · Mem0/Letta/Zep + LongMemEval 新基准）。交互性轴共识：六轨竞争 + 按场景分三层 + 分层文件系统 L0/L1/L2 + 非破坏性历史记录 + 可计量基础设施计费 + Jev 选 vs 抽取预注册 + BPM 一等设计要素 + 第三条路 OSAM（δ-mem）+ MoE-Agent 共设计 + artifact 经验积累（ExperienceIndex） + 可移植性维度（Memory Portability）+ theaiengineer 三层架构工业定义 十一轨。

研发轴/Reasoning 轴（R115 续立 + +1 RECAST 策略路由）：① R112 MatRAG 2610.01767；② R107 Better Retrieval Worse Robustness 2608.22872 + Keyword vs Semantic Search 2609.37749；③ R107 EnSI-RAG 2608.21252；④ R106 SAGE 2609.30192；⑤ R104 RAG 26 篇 6 年演进时间线；⑥ R111 abstract Reasoning 轴定义；⑦ R112 Gradient Flow RAG Reimagined 5 Breakthroughs（2026 主方向 = Agentic RAG）+ Reasoning-Language Alignment 2610.03136；⑧ R113 Source Learning 2610.02150 从检索到能力的范式跃迁；⑨ R114 EC-RAG 2610.08674 视频事件链 training-free；⑩ R114 Sensor-Language-Action 2610.08244 语义桥接；⑪ R115 RECAST 2610.10507 自适应证据路由（Google DM · "检索文档"→"推导证据"）。研发轴共识：多跳/长文档 RAG 推理从"经验堆叠"→ "结构化推理"（SAGE + EnSI-RAG + MatRAG + EC-RAG）；2026 年 Agentic RAG 成为工业落地主方向（Atlan + Lyzr + Squirro + Progress Software + Turingpost 多源印证）；"检索答案"→"建立能力"范式转变获学术支撑（Source Learning）；多模态 RAG 从"模态堆叠"→"事件级 + 实体级 + 动作级 + 多跳级"结构化推理；"检索文档"→"推导证据"（RECAST 2610.10507）。

4.3 共识与争议

新共识（14 项）：R115 219-222（自适应证据路由 / 记忆可移植性 / 经验驱动检索 / Defense 九节点）+ R114 215-218（表征层统一 / 提示注入检测标准化 / 视频事件链 / Agentic 超参可解释）+ R113 212-214（持续 ingestion 防御 / BPM / 评测从可选到必选）+ R112 209-211（多模态检索质量 / 多语言鲁棒性 / 记忆成本工程化）。

新争议（6 项）：R115 191（四路径边界）+ R114 190（失败率口径）+ R113 188-189（协议级 vs 模型级 / 统计 vs 宣传）+ R112 186-187（多模态 vs 单模态 / 多语言对齐 vs 统一协议）。

新开放（10 项）：R115 261-263（自适应路由样本效率 / 记忆迁移工程化 / Defense 九节点工业采纳率）+ R114 259-260 + R113 257-258 + R112 254-256。

新趋势（13 项）：R115 328-330（自适应证据路由工业化 / 记忆可移植性标准化 / Defense 九节点成形）+ R114 325-327（表征层统一与轻量化 / Defense 五节点 / 多模态结构化推理深化）+ R113 322-324 + R112 318-321。

4.4 开放问题

① RAG 四轴分类法正交性论证与覆盖回检。② RAG 时间维度失效系统性根因量化。③ RAG 多模态安全差异化设计。④ RAG Defense 轴 OWASP 映射粒度。⑤ RAG 工业部署工程化（70-90% 失败率真实工程问题 R113-R115 续立 + 口径统一方法论）。⑥ Agent Stack ≠ LLM Stack + Agent Guardrails 独立成学科具体实施路径量化。⑦ RAG 持续 ingestion 安全协议标准化（R113 续立）。⑧ RAG 工业部署失败根因可重现性（R113 续立）。⑨ RAG 多模态检索质量控制方法的统一框架（R112 续立）。⑩ RAG 多语言鲁棒性评测标准化（R112 续立）。⑪ RAG 记忆基础设施计费模式与成本优化（R112 续立）。⑫ RAG 开源 VecDB 替代专有 SaaS 迁移路径工程化（R112 续立）。⑬ RAG 表征层统一（UNREAL）与第三条路（δ-mem）边界 R114 续立。⑭ RAG Agentic 超参优化可解释性（failure attribution 推广性）R114 续立。⑮ RAG Defense 轴五节点标准化体系（RAG-PIBench 为主）的工业采纳率 R114 续立。⑯ RAG 综述 1+1 双锚（2023 三阶段范式 + 2026 四轴分类）的协同与冲突 R114 续立。⑰ RAG 自适应证据路由学习样本效率 R115 新增（RECAST 2610.10507）。⑱ RAG 记忆系统迁移工程化 R115 新增（Memory Portability 2609.05339）。⑲ RAG Defense 轴九节点标准化体系工业采纳率 R115 新增（EAL-Bench 加入）。⑳ RAG 四路径边界（UNREAL 表征层 vs RECAST 策略路由 vs δ-mem 第三条路 vs ExperienceIndex 经验积累）R115 新增。

4.5 趋势

① RAG 自适应证据路由与策略学习工业化趋势（R115 新增）——RECAST 2610.10507 + Agentic AutoRAG 2610.08452 + Agentic RAG SoK 2026 + Atlan + Lyzr + Squirro + Progress Software + Turingpost 多源印证 = RAG 系统从"固定相似度检索"向"自适应策略学习"演进；Agentic RAG 在企业 AI 平台成为 2026 主导模式。② RAG 记忆系统可移植性与生产模型升级标准化趋势（R115 新增）——Memory Portability 2609.05339 + δ-mem 2605.12357 + ExperienceIndex 2610.10091 + theaiengineer 2026 Memory 三层架构 = RAG 记忆系统"可移植性"成为生产系统标配；模型升级标准流程将进入工业实践。③ RAG Defense 轴九节点标准化体系成形趋势（R115 新增）——EAL-Bench + RAG-PIBench + Bounded Provisional + imMRAG + RAGScope + Agentic-ZTA + Trojan Hippo Bench + VirusCascade + Reasoning-Language Alignment = RAG Defense 轴从"五节点"→"九节点"标准化体系成形。④ RAG 表征层统一与轻量化趋势（R114 续立）。⑤ RAG Defense 轴五节点标准化体系成形趋势（R114 续立 → R115 九节点）。⑥ RAG 多模态结构化推理深化趋势（R114 续立）。⑦ RAG 协议级门控标准化趋势（R113 续立）。⑧ RAG 作为业务流程一等设计要素趋势（R113 续立）。⑨ RAG 评测体系从可选到必选趋势（R113 续立）。⑩ RAG 多模态检索质量控制趋势（R112 续立）。⑪ RAG 多语言鲁棒性工业化趋势（R112 续立）。⑫ RAG 记忆基础设施分层与成本工程化趋势（R112 续立）。⑬ RAG 开源向量数据库替代 SaaS 趋势（R112 续立）。⑭ RAG Defense 轴立库 + OWASP 系统化映射工程化趋势（R110-R115 续立）。⑮ RAG 从"检索+推理"→"检索+推理+时效感知"三元论（R109 续立）。⑯ RAG 多模态化从"技术堆叠"→"专属安全+质量控制+物理世界表征基准"（R109+R112 续立）。⑰ RAG 综述从"单一 Agentic"→"四轴系统化 + 2023 三阶段范式双锚"+正交性争议（R109+R110+R111+R114+R115 续立）。⑱ RAG memory 范式从"五轨竞争"→"八轨+三层架构+分层文件系统+非破坏性+Jev 选 vs 抽取+第三条路 OSAM+MoE-Agent 共设计" → R115 十一轨（+ ExperienceIndex + Memory Portability + theaiengineer 三层架构）。⑲ RAG 检索质量优化从"现象观测"→"位置编码根因理论化"（R108 续立）。⑳ RAG 工业部署从"实验"→"production-critical infrastructure"+选型决策树+70-90% 失败率多源印证+评测体系从可选到必选+口径统一方法论（R109-R115 续立）。㉑ RAG 效率优化从"全量"→"按场景选型"+第三路径+分层 token 经济+开源 VecDB 替代+协议级门控+模型内部统一+第三条路 OSAM+Agentic 超参可解释优化+自适应策略路由（R104-R115 续立）。㉒ Agent Stack ≠ LLM Stack + Agent Guardrails 独立成学科趋势（R111-R115 续立）。㉓ RAG 四路径边界收敛趋势（R115 新增）——UNREAL（表征层统一）+ RECAST（策略路由）+ δ-mem（第三条路 8×8 OSAM）+ ExperienceIndex（经验积累）四条路径从竞争走向场景化分工。㉔ RAG Memory 三层架构工业定义趋势（R115 新增）——theaiengineer 2026：in-context memory + agentic memory management + shared memory infrastructure；Mem0/Letta/Zep + LongMemEval 标准化。㉕ Jay CSDN 工程实测数据工业化趋势（R115 新增）——Top-1 +24.2pp / 跨模态召回 +27.5pp / 成本 ↓33% 印证 RAG 工程化实践价值。

5. 关键参考清单（R115 增量版）

R115 增量：① RECAST arXiv:2610.10507v1 paper_card 1717（Google DM · 自适应证据路由 · "检索文档"→"推导证据"· 主分类 rag · ⭐⭐⭐⭐ · Reasoning 轴结构化推理新节点）② ExperienceIndex arXiv:2610.10091v1 paper_card 1716（MIT/Microsoft/UPenn · Jacob Andreas · artifact 经验记忆 · 主分类 agent · 副分类 rag · 强邻接 · ⭐⭐⭐⭐ · 交互性轴新节点）③ Memory Portability arXiv:2609.05339v1 paper_card 1238（受控研究 · 模型升级可移植性 · 主分类 rag · ⭐⭐⭐⭐ · 交互性轴新维度 · 3 关键发现：方向特定迁移测试 + embedding space 隔离 + 源历史保留）④ EAL-Bench arXiv:2609.01836 paper_card 1187（5+2 LLM × 3 场景 · 授权漂移评测 · S2被引 3 · 主分类 agent · 副分类 rag · ⭐⭐⭐⭐ · Defense 轴九节点标准化体系新增第 9 节点）⑤ Jay CSDN Oct 9 工程视角（手写 RAG 完整流水线 + 企业 RAG 成本分析 + Agentic RAG ADK + 多模态 RAG Dify Pipeline + 12-Factor Agents + 七层企业 AI 架构 · Top-1 +24.2pp / 跨模态召回 +27.5pp / 成本 ↓33% · ⚠️ 召回率 +17% / 通义灵码 +23% 待原论文核验）⑥ Substack theaiengineer 2026 Memory 三层架构续立（in-context + agentic + shared infrastructure · Mem0/Letta/Zep + LongMemEval 新基准）⑦ R115 fresh web_search 5 源（Progress Software Agentic RAG + Lyzr + Squirro + Atlan + Turingpost · 60+ 格式 · 八模式 · 2026 主导模式 Agentic RAG）。

R114 增量续立：① RAG-PIBench arXiv:2610.08571v1 paper_card 1700（⭐⭐⭐⭐）② UNREAL arXiv:2610.08463v1 paper_card 1701（⭐⭐⭐⭐）③ EC-RAG arXiv:2610.08674v1 paper_card 1699（⭐⭐⭐⭐）④ Agentic AutoRAG arXiv:2610.08452v1 paper_card 1706（⭐⭐⭐⭐）⑤ δ-mem arXiv:2605.12357v2 NeurIPS 2026（⭐⭐⭐⭐）⑥ RAG Foundation Survey arXiv:2312.10997 paper_card 438（⭐⭐⭐⭐⭐）⑦ 4 件邻接级（DeCoPrune + Structuring MoE + Sensor-Language-Action + DMAD）。

R113 增量续立：① Bounded Provisional Visibility arXiv:2610.05826v1 paper_card 1687（⭐⭐⭐⭐）② AI-Decision Checkpoints arXiv:2610.06207v1 paper_card 1686（⭐⭐⭐）③ Source Learning arXiv:2610.02150v1 paper_card 1684（⭐⭐⭐⭐）④ Agentic-ZTA arXiv:2610.05782v1 paper_card 1688（⭐⭐⭐⭐）⑤ Jev 选 vs 抽取 arXiv:2609.34227 paper_card 1698（⭐⭐⭐⭐）⑥ UndoBench arXiv:2610.05622（⭐⭐⭐⭐）⑦ jay 10-07 CSDN 工业视角 + R113 fresh web_search 5 源。

R112 续立锚点（2 件 RAG 主分类 net-new + 1 件强邻接 + 9 件邻接级）：CLIMB 2610.03421 + Reasoning-Language Alignment 2610.03136 + World Embedding Benchmark 2610.03632 + OpenViking 三件套 + Mem++ + Trustworthy Agentic AI + JAM v2 + LawCompass + PostgreSQL-V 2.0 / ACRONYM + Trojan Hippo Bench + VirusCascade + jay 10-06 tech-brief 工业视角。

R111 7 件锚点续立：Mapping the RAG Landscape 2610.01936 + Walking the Embedding Space 2610.01871 + MatRAG 2610.01767 + Temporal Validity 2606.26511 + AutoMem 2607.01224 + A-TMA 2607.01935 + ZooClaw-FashionSigLIP2 2606.27708。

CVE 14 条（R114 沿用 R113）：CVE-2025-68664 + CVE-2026-20805 + CVE-2026-21858 + CVE-2026-22778 + CVE-2026-26030 + CVE-2026-3172 + CVE-2026-45829 + CVE-2026-4810 + CVE-2026-49468 + CVE-2026-50548 + CVE-2026-50549 + CVE-2026-54309 + CVE-2026-55255 + CVE-2026-56274。

DOI 21 件（R115 19 + 新增 2）：10.48550/arxiv.2606.26511 + 10.48550/arxiv.2606.27708 + 10.48550/arxiv.2607.01224 + 10.48550/arxiv.2607.01935 + 10.48550/arxiv.2607.26760 + 10.48550/arxiv.2609.39075 + 10.48550/arxiv.2610.02150 + 10.48550/arxiv.2312.10997（R114） + 10.48550/arxiv.2605.12357（R114） + 10.48550/arxiv.2609.01836（R115 新增 · EAL-Bench DOI） + 10.48550/arxiv.2609.05339（R115 新增 · Memory Portability；DOI 待 OpenAlex 补充）。

v1/V1/补充 URL 覆盖：

R115 v1 新增：http://arxiv.org/abs/2610.10507v1 http://arxiv.org/abs/2610.10091v1 https://arxiv.org/abs/2609.01836 https://arxiv.org/abs/2609.05339v1 https://arxiv.org/pdf/2609.01836 https://huggingface.co/papers/2609.01836 https://huggingface.co/papers/2609.05339 https://blog.csdn.net/weixin_43726381/article/details/161983901 https://blog.csdn.net/weixin_53920044/article/details/148875486 https://blog.csdn.net/FuncWander/article/details/160794992 https://codearts.csdn.net/6a92d6c63bda720d4b3799c5.html https://gitcode.csdn.net/6a08f72f10ee7a33f2730f92.html https://xingyun3d.csdn.net/69f019380a2f6a37c5a67fba.html https://adg.csdn.net/695331d35b9f5f31781bc374.html https://developer.jdcloud.com/article/4382 https://www.lyzr.ai/blog/agentic-rag https://squirro.com/squirro-blog/state-of-rag-genai https://atlan.com/know/what-is-rag https://www.turingpost.com/p/ragtypes`

R114 v1 新增：http://arxiv.org/abs/2610.08463v1 http://arxiv.org/abs/2610.08571v1 http://arxiv.org/abs/2610.08674v1 http://arxiv.org/abs/2610.08452v1 https://arxiv.org/abs/2605.12357v2 https://arxiv.org/abs/2605.12357 https://arxiv.org/abs/2312.10997 https://arxiv.org/pdf/2605.12357 https://arxiv.org/pdf/2312.10997 https://huggingface.co/papers/2605.12357 https://huggingface.co/papers/2312.10997 https://alphasignalai.substack.com/p/rag-and-long-context-arent-enough`

R113 v1 续立：http://arxiv.org/abs/2610.05826v1 http://arxiv.org/abs/2610.06207v1 https://arxiv.org/abs/2610.02150 http://arxiv.org/abs/2610.05782v1 https://arxiv.org/abs/2609.34227 https://arxiv.org/abs/2610.05622 https://arxiv.org/abs/2606.16903`

R113 fresh web_search URL 续立：https://dev.to/gabrielanhaia/70-of-enterprise-rag-deployments-fail-before-production-heres-what-kills-them-26ml https://pub.towardsai.net/mastering-agentic-rag-3-architecture-patterns-for-production-grade-ai-system-with-examples-03f799a3cbd0 https://workativ.com/hr/blog/agentic-rag https://www.cake.ai/blog/why-90-of-agentic-rag-projects-fail-and-how-cake-changes-that https://www.digitalapplied.com/blog/88-percent-ai-agents-never-reach-production-failure-framework https://blog.starmorph.com/blog/rag-techniques-compared-best-practices-guide https://duynguyenngoc.com/posts/advanced-rag-techniques-2026 https://atlan.com/know/what-is-rag https://investors.progress.com/news-releases/news-release-details/progress-agentic-rag-named-2026-ai-excellence-award-winner https://arxiv.org/html/2603.10765v1

R115 fresh web_search URL 新增：https://investors.progress.com/news-releases/news-release-details/progress-agentic-rag-named-2026-ai-excellence-award-winner https://www.lyzr.ai/blog/agentic-rag https://squirro.com/squirro-blog/state-of-rag-genai https://atlan.com/know/what-is-rag https://www.turingpost.com/p/ragtypes`

R112-R106 v1 沿用：http://arxiv.org/abs/2606.26511v1 http://arxiv.org/abs/2607.01935v1 http://arxiv.org/abs/2610.01767v1 http://arxiv.org/abs/2610.01936v1 https://arxiv.org/abs/2610.01871v1 https://arxiv.org/abs/2605.29640v1 https://arxiv.org/abs/2609.11390 https://arxiv.org/abs/2610.02002v1 https://arxiv.org/abs/2610.03136v1 https://arxiv.org/abs/2610.03421v1 https://arxiv.org/abs/2610.03632v1 https://arxiv.org/abs/2609.22712 https://arxiv.org/abs/2610.01027v1 https://arxiv.org/abs/2608.15994 https://arxiv.org/abs/2609.03712 https://arxiv.org/abs/2605.01970v4 https://arxiv.org/abs/2609.38270v1 http://arxiv.org/abs/2608.21252v1 http://arxiv.org/abs/2608.22752v1 http://arxiv.org/abs/2608.22872v1 http://arxiv.org/abs/2609.37673v1 http://arxiv.org/abs/2609.37749v1 http://arxiv.org/abs/2609.39075v1 http://arxiv.org/abs/2609.39929v1 https://arxiv.org/abs/2609.23551 https://arxiv.org/abs/2609.33485 https://arxiv.org/abs/2609.35629 https://arxiv.org/abs/2609.27213 https://arxiv.org/abs/2609.30192 https://arxiv.org/abs/2609.30904 https://arxiv.org/abs/2609.34242 https://arxiv.org/abs/2609.17012 https://arxiv.org/abs/2609.19969 https://arxiv.org/abs/2609.35427 https://arxiv.org/abs/2609.37559` 以及 R107-R106 历史 URL Archive。

R107-R106 补充 URL：https://arxiv.org/abs/2401.00396 https://arxiv.org/abs/2511.07328 https://arxiv.org/abs/2608.21252 https://arxiv.org/abs/2608.22752 https://arxiv.org/abs/2608.22872 https://arxiv.org/abs/2608.24053 https://arxiv.org/abs/2609.16409 https://arxiv.org/abs/2609.35427 https://arxiv.org/abs/2609.37559 https://arxiv.org/abs/2609.37673 https://arxiv.org/abs/2609.37749 https://agent.csdn.net/6a52fd2f662f9a54cb8e8279.html https://blog.csdn.net/chen1415886044/article/details/166736574 https://blog.csdn.net/qq_33957603/article/details/165118957 https://blog.csdn.net/yonggeit/article/details/166773439 https://cobusgreyling.substack.com/p/ragtruth https://github.com/Tencent/WeMM-Embedding https://github.com/griver/Q-RAG https://huggingface.co/papers/2609.36322 https://hyper.ai/en/papers/2608.24053 https://local-ai-zone.github.io/guides/best-ai-reranker-models-ultimate-ranking-2026.html https://medium.com/coding-nexus/wemm-embedding-the-open-source-multimodal-embedding-model-built-for-1-billion-requests-a-day-e8cd1cbb518f https://ojs.aaai.org/index.php/AAAI/article/view/40726/44687 https://slash-digital.io/en/insights/rag-2026 https://www.elastic.co/search-labs/blog/jina-reranker-35-legal-medical-structured-data https://www.emergentmind.com/topics/jina-reranker-v3 https://www.linkedin.com/posts/muhammad-arslan-b591b1285_research-aiagents-agenticai-activity-7474154906019438592-erLD https://www.percona.com/blog/vector-distance-vector-similarity-search-in-percona-server-for-mysql` https://arxiv.org/abs/2511.07328v2 https://blog.csdn.net/a2875254060/article/details/156463704 https://blog.csdn.net/a28752405254060/article/details/156463704 https://github.com/ParticleMedia/RAGTruth https://open.substack.com/pub/alexewerlof/p/owasp-top-10-ai-llm-agents https://open.substack.com/pub/alexewerlof/p/owasp-top-10-ai-llm-agents?r=18prwo https://ragaboutit.com

R112 新增 URL：https://github.com/volcengine/OpenViking https://developers.redhat.com/articles/2026/04/23/deploy-openviking-openshift-ai-improve-ai-agent-memory https://pub.towardsai.net/the-agent-context-wars-three-battles-at-different-layers-487dab110ea8 https://gradientflow.substack.com/p/rag-reimagined-5-breakthroughs-you https://aishwaryasrinivasan.substack.com/p/all-you-need-to-know-about-rag-in https://1bench.dev/databases/vector https://www.designveloper.com/blog/best-open-source-vector-database https://mem0.ai/blog/state-of-ai-agent-memory-2026 https://www.databricks.com/blog/long-context-rag-performance-llms`

URL Archive (R82-R106 Historical References · 紧凑格式)

https://academy.dair.ai/papers/does-your-agents-memory-survive-a-model-upgrade-a-controlled-study-of-memory-por-2609.05339 https://aclanthology.org/2025.findings-acl.861 https://aclanthology.org/2026.findings-acl.1619.pdf https://actian.com/blog/databases/how-to-evaluate-vector-databases-in-2026 https://adg.csdn.net/6a68c5bd662f9a54cb956060.html https://agentmarketcap.ai/blog/2026/04/11/agent-memory-architecture-production-2026 https://aiagentssimplified.substack.com/p/the-2026-path-to-learning-ai-agents https://aiamastery.substack.com/p/lesson-44-evaluating-agentic-rag https://aiengineeringinsider.substack.com/p/rag-interview-q-and-a-how-retrieval https://aiexpjourney.substack.com/p/advanced-rag-06-exploring-query-rewriting-23997297f2d1 https://aiquinta.ai/blog/harness-engineering-guide-to-reliable-ai-agents https://aishwaryasrinivasan.substack.com/p/all-you-need-to-know-about-rag-in https://alphacorp.ai/blog/best-vector-databases-for-rag-2026-top-7-picks https://alphasignalai.substack.com https://alphasignalai.substack.com/p/rag-and-long-context-arent-enough https://aminrj.com/posts/rag-security-architecture https://arxiv.org/abs/2005.11401 https://arxiv.org/abs/2402.07867 https://arxiv.org/abs/2409.01666 https://arxiv.org/abs/2410.05983 https://arxiv.org/abs/2412.11854 https://arxiv.org/abs/2501.09136 https://arxiv.org/abs/2502.08826 https://arxiv.org/abs/2504.01840 https://arxiv.org/abs/2504.09554 https://arxiv.org/abs/2505.23052 https://arxiv.org/abs/2506.03401 https://arxiv.org/abs/2506.05176 https://arxiv.org/abs/2507.02962 https://arxiv.org/abs/2507.05257 https://arxiv.org/abs/2509.12541 https://arxiv.org/abs/2510.15253 https://arxiv.org/abs/2510.15253v3 https://arxiv.org/abs/2512.09487 https://arxiv.org/abs/2512.09695v4 https://arxiv.org/abs/2512.14629 https://arxiv.org/abs/2601.02993 https://arxiv.org/abs/2601.07504 https://arxiv.org/abs/2601.07978 https://arxiv.org/abs/2601.12538 https://arxiv.org/abs/2601.23254 https://arxiv.org/abs/2602.00296 https://arxiv.org/abs/2602.03442 https://arxiv.org/abs/2602.03992 https://arxiv.org/abs/2602.06052 https://arxiv.org/abs/2602.08005 https://arxiv.org/abs/2602.11671 https://arxiv.org/abs/2603.00873 https://arxiv.org/abs/2603.00873v1 https://arxiv.org/abs/2603.07379 https://arxiv.org/abs/2603.07379v1 https://arxiv.org/abs/2603.25152 https://arxiv.org/abs/2604.04359 https://arxiv.org/abs/2604.04359v1 https://arxiv.org/abs/2604.08224 https://arxiv.org/abs/2604.09666 https://arxiv.org/abs/2604.14572 https://arxiv.org/abs/2605.01495 https://arxiv.org/abs/2605.16352 https://arxiv.org/abs/2605.21071v4 https://arxiv.org/abs/2605.27123 https://arxiv.org/abs/2605.28120 https://arxiv.org/abs/2606.00610v1 https://arxiv.org/abs/2606.01613 https://arxiv.org/abs/2606.16817 https://arxiv.org/abs/2606.16817v1 https://arxiv.org/abs/2606.18037 https://arxiv.org/abs/2606.26458 https://arxiv.org/abs/2606.26511 https://arxiv.org/abs/2606.26916 https://arxiv.org/abs/2606.27708 https://arxiv.org/abs/2607.08716 https://arxiv.org/abs/2607.17715 https://arxiv.org/abs/2607.17715v1 https://arxiv.org/abs/2607.18152 https://arxiv.org/abs/2607.21596 https://arxiv.org/abs/2607.22042 https://arxiv.org/abs/2607.24748 https://arxiv.org/abs/2607.26497v2 https://arxiv.org/abs/2607.26760 https://arxiv.org/abs/2607.26760v1 https://arxiv.org/abs/2607.29677 https://arxiv.org/abs/2608.00902 https://arxiv.org/abs/2608.00902v1 https://arxiv.org/abs/2608.01526 https://arxiv.org/abs/2608.01526v1 https://arxiv.org/abs/2608.02870 https://arxiv.org/abs/2608.08340 https://arxiv.org/abs/2608.08466 https://arxiv.org/abs/2608.09802 https://arxiv.org/abs/2608.11030 https://arxiv.org/abs/2608.11660 https://arxiv.org/abs/2608.12218 https://arxiv.org/abs/2608.12253 https://arxiv.org/abs/2608.12304 https://arxiv.org/abs/2608.12304v1 https://arxiv.org/abs/2608.12304v1v1 https://arxiv.org/abs/2608.12875 https://arxiv.org/abs/2608.12875v1 https://arxiv.org/abs/2608.13010 https://arxiv.org/abs/2608.13010v1 https://arxiv.org/abs/2608.13120 https://arxiv.org/abs/2608.13160 https://arxiv.org/abs/2608.13160v1 https://arxiv.org/abs/2608.13237 https://arxiv.org/abs/2608.13237v1 https://arxiv.org/abs/2608.13410 https://arxiv.org/abs/2608.13410v1 https://arxiv.org/abs/2608.13545 https://arxiv.org/abs/2608.13547 https://arxiv.org/abs/2608.14036 https://arxiv.org/abs/2608.14054 https://arxiv.org/abs/2608.14054v1 https://arxiv.org/abs/2608.14210 https://arxiv.org/abs/2608.14210v1 https://arxiv.org/abs/2608.14229 https://arxiv.org/abs/2608.15767 https://arxiv.org/abs/2608.15888 https://arxiv.org/abs/2608.16490 https://arxiv.org/abs/2608.16515 https://arxiv.org/abs/2608.16515v1 https://arxiv.org/abs/2608.16536 https://arxiv.org/abs/2608.16536v1 https://arxiv.org/abs/2608.16628 https://arxiv.org/abs/2608.16628v1 https://arxiv.org/abs/2608.16776 https://arxiv.org/abs/2608.16776v1 https://arxiv.org/abs/2608.16885 https://arxiv.org/abs/2608.17050 https://arxiv.org/abs/2608.17536 https://arxiv.org/abs/2608.17536v1 https://arxiv.org/abs/2608.17744 https://arxiv.org/abs/2608.17781 https://arxiv.org/abs/2608.17781v1 https://arxiv.org/abs/2608.17950v1 https://arxiv.org/abs/2608.17960 https://arxiv.org/abs/2608.17960v1 https://arxiv.org/abs/2608.18184 https://arxiv.org/abs/2608.18184v1 https://arxiv.org/abs/2608.18271 https://arxiv.org/abs/2608.18489 https://arxiv.org/abs/2608.18489v1 https://arxiv.org/abs/2608.18607 https://arxiv.org/abs/2608.18607v1 https://arxiv.org/abs/2608.18613 https://arxiv.org/abs/2608.18613v1 https://arxiv.org/abs/2608.18731 https://arxiv.org/abs/2608.18852 https://arxiv.org/abs/2608.19269 https://arxiv.org/abs/2608.19758 https://arxiv.org/abs/2608.19799 https://arxiv.org/abs/2608.19857 https://arxiv.org/abs/2608.19857v1 https://arxiv.org/abs/2608.20246 https://arxiv.org/abs/2608.20246v1 https://arxiv.org/abs/2608.20281 https://arxiv.org/abs/2608.20281v1 https://arxiv.org/abs/2608.20317 https://arxiv.org/abs/2608.20335 https://arxiv.org/abs/2608.21252v1 https://arxiv.org/abs/2608.21450 https://arxiv.org/abs/2608.22752v1 https://arxiv.org/abs/2608.23252 https://arxiv.org/abs/2608.23252v1 https://arxiv.org/abs/2608.24040 https://arxiv.org/abs/2608.24040v1 https://arxiv.org/abs/2608.25500 https://arxiv.org/abs/2608.25500v1 https://arxiv.org/abs/2608.25625 https://arxiv.org/abs/2608.25625v1 https://arxiv.org/abs/2608.25735 https://arxiv.org/abs/2608.25986 https://arxiv.org/abs/2608.25986v1 https://arxiv.org/abs/2608.26091 https://arxiv.org/abs/2608.26091v1 https://arxiv.org/abs/2608.26530 https://arxiv.org/abs/2608.26530v1 https://arxiv.org/abs/2608.26623 https://arxiv.org/abs/2608.26836 https://arxiv.org/abs/2608.26836v1 https://arxiv.org/abs/2608.27448 https://arxiv.org/abs/2608.27448v1 https://arxiv.org/abs/2608.27455 https://arxiv.org/abs/2608.27455v1 https://arxiv.org/abs/2608.27809 https://arxiv.org/abs/2608.27809v1 https://arxiv.org/abs/2608.28389 https://arxiv.org/abs/2608.28389v1 https://arxiv.org/abs/2608.31082 https://arxiv.org/abs/2608.31139 https://arxiv.org/abs/2609.01343 https://arxiv.org/abs/2609.01601 https://arxiv.org/abs/2609.01657 https://arxiv.org/abs/2609.03005 https://arxiv.org/abs/2609.03756 https://arxiv.org/abs/2609.05110 https://arxiv.org/abs/2609.05736 https://arxiv.org/abs/2609.06702 https://arxiv.org/abs/2609.08126 https://arxiv.org/abs/2609.08365 https://arxiv.org/abs/2609.08572 https://arxiv.org/abs/2609.08832 https://arxiv.org/abs/2609.09875 https://arxiv.org/abs/2609.10539 https://arxiv.org/abs/2609.10745 https://arxiv.org/abs/2609.11209 https://arxiv.org/abs/2609.11294 https://arxiv.org/abs/2609.11390 https://arxiv.org/abs/2609.11561 https://arxiv.org/abs/2609.11572 https://arxiv.org/abs/2609.11646 https://arxiv.org/abs/2609.11808 https://arxiv.org/abs/2609.13463 https://arxiv.org/abs/2609.14302 https://arxiv.org/abs/2609.14320 https://arxiv.org/abs/2609.15126 https://arxiv.org/abs/2609.15578 https://arxiv.org/abs/2609.15800 https://arxiv.org/abs/2609.15830 https://arxiv.org/abs/2609.15991 https://arxiv.org/abs/2609.16591 https://arxiv.org/abs/2609.16816 https://arxiv.org/abs/2609.16818 https://arxiv.org/abs/2609.17320 https://arxiv.org/abs/2609.17652 https://arxiv.org/abs/2609.17653 https://arxiv.org/abs/2609.18063 https://arxiv.org/abs/2609.19656 https://arxiv.org/abs/2609.20423 https://arxiv.org/abs/2609.20715 https://arxiv.org/abs/2609.22086 https://arxiv.org/abs/2609.22319 https://arxiv.org/abs/2609.24220 https://arxiv.org/abs/2609.25053 https://arxiv.org/abs/2609.25267 https://arxiv.org/abs/2609.25636 https://arxiv.org/abs/2609.25804 https://arxiv.org/abs/2609.25853 https://arxiv.org/abs/2609.25991 https://arxiv.org/abs/2609.26489 https://arxiv.org/abs/2609.26774 https://arxiv.org/abs/2609.26781 https://arxiv.org/abs/2609.29845 https://arxiv.org/abs/2609.31002 https://arxiv.org/abs/2609.31415 https://arxiv.org/abs/2609.32049 https://arxiv.org/abs/2609.32259 https://arxiv.org/abs/2609.32722 https://arxiv.org/abs/2609.33403 https://arxiv.org/abs/2609.34227 https://arxiv.org/abs/2609.34385 https://arxiv.org/abs/2609.34392 https://arxiv.org/abs/2609.35530 https://arxiv.org/abs/2609.36138 https://arxiv.org/abs/2609.36314 https://arxiv.org/abs/2609.36322 https://arxiv.org/abs/2609.36333 https://arxiv.org/abs/2609.36636 https://arxiv.org/abs/2609.36730 https://arxiv.org/abs/2609.37147 https://arxiv.org/abs/2609.37687 https://arxiv.org/abs/2609.37690 https://arxiv.org/abs/2609.37863 https://arxiv.org/abs/2609.38349 https://arxiv.org/abs/2609.38660 https://arxiv.org/abs/2609.38721 https://arxiv.org/abs/2609.39075 https://arxiv.org/abs/2609.39102 https://arxiv.org/abs/2609.39154 https://arxiv.org/abs/2609.39788 https://arxiv.org/abs/2609.39841 https://arxiv.org/abs/2609.39903 https://arxiv.org/abs/2609.39929 https://arxiv.org/abs/2609.39982 https://arxiv.org/abs/2609.40222 https://arxiv.org/abs/2609.40316 https://arxiv.org/abs/2610.00437 https://arxiv.org/abs/2610.00544 https://arxiv.org/abs/2610.00559 https://arxiv.org/abs/2610.00812 https://arxiv.org/abs/2610.01092 https://arxiv.org/abs/2610.01428 https://arxiv.org/abs/2610.01767 https://arxiv.org/abs/2610.01871 https://arxiv.org/abs/2610.01936 https://arxiv.org/abs/2610.02150 https://arxiv.org/abs/2610.02196 https://arxiv.org/abs/2606.10953 https://arxiv.org/html/2510.11541v2 https://arxiv.org/html/2601.07978 https://arxiv.org/html/2604.09666v1 https://arxiv.org/html/2604.14572v3 https://arxiv.org/html/2607.24748 https://arxiv.org/html/2607.24748v1 https://arxiv.org/html/2608.21252v1 https://arxiv.org/html/2609.25853 https://arxiv.org/html/2609.25991v1 https://arxiv.org/pdf/2602.03442 https://arxiv.org/pdf/2603.07379 https://arxiv.org/pdf/2607.17715 https://arxiv.org/pdf/2607.26760 https://arxiv.org/pdf/2607.29677 https://arxiv.org/pdf/2608.00902 https://arxiv.org/pdf/2608.01526 https://arxiv.org/pdf/2608.09802 https://arxiv.org/pdf/2608.12304 https://arxiv.org/pdf/2608.13010 https://arxiv.org/pdf/2608.13237 https://arxiv.org/pdf/2608.13410 https://arxiv.org/pdf/2608.18184 https://arxiv.org/pdf/2608.18489 https://arxiv.org/pdf/2608.19269 https://arxiv.org/pdf/2608.19857 https://arxiv.org/pdf/2608.21252 https://arxiv.org/pdf/2608.22752 https://arxiv.org/abs/2607.01224 https://arxiv.org/abs/2607.01935 https://arxiv.org/pdf/2608.23252 https://arxiv.org/pdf/2608.25500 https://arxiv.org/pdf/2608.25625 https://arxiv.org/pdf/2608.25757 https://arxiv.org/pdf/2608.25735 https://arxiv.org/pdf/2608.25986 https://arxiv.org/pdf/2608.26091 https://arxiv.org/pdf/2608.26530 https://arxiv.org/pdf/2608.26836 https://arxiv.org/pdf/2608.27448 https://arxiv.org/pdf/2608.27455 https://arxiv.org/pdf/2608.27809 https://arxiv.org/pdf/2608.28389 https://arxiv.org/pdf/2608.31139 https://arxiv.org/pdf/2609.01343 https://arxiv.org/pdf/2609.01601 https://atlan.com/know/advanced-rag-techniques https://atlan.com/know/best-ai-agent-memory-frameworks-2026 https://bhavishyapandit9.substack.com https://blog.bytebytego.com/p/graphrag-how-ai-answers-questions https://blog.compendialabs.org/posts/2026-09-07-al-agent-memory-model-upgrades https://blog.csdn.net/a28752405254060/article/details/156463704 https://blog.csdn.net/weixin_43726381/article/details/161983901 https://blog.csdn.net/weixin_53920044/article/details/148875486 https://blog.csdn.net/FuncWander/article/details/160794992 https://blog.starmorph.com/blog/rag-techniques-compared-best-practices-guide https://blog.turingpost.com/p/ragtypes https://buzzgrewal.medium.com/ai-agents-dont-need-vector-search-anymore-inside-the-agentic-search-stack-replacing-rag-in-2026-58efcabe4f6f https://c3.unu.edu/projects/ai/deepresearch/demo_research-report_searchengine2.html https://ceur-ws.org/Vol-4211/paper05.pdf https://claudiostamile.substack.com/p/agent-memory-is-not-rag-a-practical https://codefarm.in/blog/gen-ai/long-context-vs-rag-2026 https://codearts.csdn.net/6a92d6c63bda720d4b3799c5.html https://commandcode.ai/guides/rag-in-2026 https://cruxdigits.nl/blog/rag-tables-bm25-vs-vector-search https://cruxdigits.nl/blog/rag-vs-graphrag-2026 https://cycode.com/blog/owasp-top-10-agentic-applications https://datanucleus.dev/rag-and-agentic-ai/agentic-rag-enterprise-guide-2026 https://dev.to/varun_pratapbhardwaj_b13/5-ai-agent-memory-systems-compared-mem0-zep-letta-supermemory-superlocalmemory-2026-benchmark-59p3 https://docs.oracle.com/en-us/iaas/releasenotes/services/conteng https://doi.org/10.48550/arxiv.2607.26760 https://easychair.org/cfp/SynIRgy2026 https://engineersguide.substack.com/p/best-vector-databases-rag https://evoailabs.medium.com/metis-the-dawn-of-native-memory-in-foundation-models-de4ddb55cef1 https://futureagi.com/blog/agentic-rag-systems-2025 https://futureagi.com/blog/llm-agent-architectures-core-components/ https://futureagi.com/blog/rag-architecture-llm-2025 https://futureagi.com/blog/rag-summarization https://genai.owasp.org/resource/owasp-top-10-for-agentic-applications-for-2026 https://getfeather.store/theory/agentic-search-vs-vector-database-2026 https://gitcode.csdn.net/6a0eb9cb10ee7a33f274216e.html https://gitcode.csdn.net/6a08f72f10ee7a33f2730f92.html https://github.com/Ayanami0730/arag https://github.com/DavidZWZ/Awesome-RAG-Reasoning https://github.com/OpenBMB/UltraRAG https://github.com/RUC-NLPIR/Awesome-Long-Horizon-Agents https://github.com/Shubhamsaboo/awesome-llm-apps https://github.com/YennNing/MC-Search https://github.com/YunjiaXi/Awesome-Search-Agent-Papers https://github.com/dukesun99/Corpus2Skill https://github.com/edinburghnlp/mmlongbench https://github.com/langchain-ai/langgraph/releases https://github.com/llm-lab-org/multimodal-rag-survey https://github.com/modelcontextprotocol https://github.com/neulab/think-before-you-link https://github.com/omkarbhad/agentforge https://github.com/piterkai/nova-retrieve https://github.com/run-llama/ExtractBench https://github.com/umwyf/CRITICL https://github.com/weitianxin/Awesome-Agentic-Reasoning https://github.com/zilliztech/deep-searcher https://github.com/ziqiwang0908/RAGRouter-Bench https://goteleport.com/blog/owasp-top-10-agentic-applications https://harness.io/blog/ai-deployment-in-production-orchestrate-llms-rag-agents https://huggingface.co/Qwen/Qwen3-Embedding-8B https://huggingface.co/blog/hotchpotch/mlateon-multilingual-colbert-hakari-bench https://huggingface.co/blog/state-of-open-models-summer-2026 https://huggingface.co/datasets/neulab/merlin-rare https://huggingface.co/datasets/paiteq-ai/rag-bench-2026q2 https://huggingface.co/nvidia/nemotron-colembed-vl-8b-v2 https://huggingface.co/papers/2501.09136 https://huggingface.co/papers/2601.07504 https://huggingface.co/papers/2602.03442 https://huggingface.co/papers/2603.07379 https://huggingface.co/papers/2604.14572 https://huggingface.co/papers/2606.16817 https://huggingface.co/papers/2607.26760 https://huggingface.co/papers/2607.29677 https://huggingface.co/papers/2608.12304 https://huggingface.co/papers/2608.12875 https://huggingface.co/papers/2608.13010 https://huggingface.co/papers/2608.13237 https://huggingface.co/papers/2608.13410 https://huggingface.co/papers/2608.16515 https://huggingface.co/papers/2608.16536 https://huggingface.co/papers/2608.16628 https://huggingface.co/papers/2608.16776 https://huggingface.co/papers/2608.17536 https://huggingface.co/papers/2608.17781 https://huggingface.co/papers/2608.17960 https://huggingface.co/papers/2608.18184 https://huggingface.co/papers/2608.18489 https://huggingface.co/papers/2608.18607 https://huggingface.co/papers/2608.18613 https://huggingface.co/papers/2608.19857 https://huggingface.co/papers/2608.20246 https://huggingface.co/papers/2608.20281 https://huggingface.co/papers/2608.21252 https://huggingface.co/papers/2608.22752 https://huggingface.co/papers/2608.23252 https://huggingface.co/papers/2608.25500 https://huggingface.co/papers/2608.25625 https://huggingface.co/papers/2608.25735 https://huggingface.co/papers/2608.25986 https://huggingface.co/papers/2608.26091 https://huggingface.co/papers/2608.26530 https://huggingface.co/papers/2608.26623 https://huggingface.co/papers/2608.26836 https://huggingface.co/papers/2608.27448 https://huggingface.co/papers/2608.27455 https://huggingface.co/papers/2608.27809 https://huggingface.co/papers/2608.28389 https://huggingface.co/papers/2608.31082 https://huggingface.co/papers/2608.31139 https://huggingface.co/papers/2609.01343 https://huggingface.co/papers/2609.01601 https://huggingface.co/papers/2609.24220 https://huggingface.co/papers/2609.25053 https://huggingface.co/papers/2609.25636 https://huggingface.co/papers/2609.25853 https://huggingface.co/papers/2609.26489 https://huggingface.co/papers/2609.26781 https://huggingface.co/papers/2609.29845 https://huggingface.co/papers/2608.17050 https://huggingface.co/cohere/rerank-v3.5 https://huggingface.co/zeroentropy/zembed-1-embedding https://iclr.cc/virtual/2026/oral/10010224 https://jamwithai.substack.com/p/the-2026-roadmap-production-aiml https://jobsbyculture.com/jobs/ai-engineer-memory-systems https://kanerika.com/blogs/rag-vs-agentic-rag https://kenhuangus.substack.com/p/announcing-the-10-part-series-the https://kenhuangus.substack.com/p/how-ai-agents-actually-remember-inside https://kenhuangus.substack.com/p/the-kv-cache-wars https://lilianweng.github.io/posts/2026-07-04-harness/ https://llms3.com/blog/when-the-benchmarks-stopped-agreeing-july-2026 https://lyzr.ai/blog/agentic-rag https://mc-search-project.github.io https://medium.com/@Micheal-Lanham/pipeline-rag-vs-agentic-rag-vs-knowledge-graph-rag-what-actually-works-and-when-47a26649a457 https://mem0.ai/blog/ai-memory-benchmarks-in-2026 https://mem0.ai/blog/state-of-ai-agent-memory-2026 https://mem0.ai/research https://memoryarena.github.io https://micheallanham.substack.com/p/comparative-analysis-of-rag-architectures https://microsoft.com/en-us/research/blog/echoverse-deep-evolving-environments-for-computer-use-agents/ https://microsoft.com/en-us/research/blog/evolib-turning-experience-into-evolving-knowledge/ https://microsoft.com/en-us/research/blog/orchard-an-open-framework-for-scalable-agentic-ai/ https://mlcommons.org/2026/08/endtoend-inference https://mlflow.org/articles/llm-application-architecture-a-2026-engineers-guide/ https://mnemoverse.com/docs/library/ai-memory-solutions-2026-q3 https://multimodalrag.github.io https://neo4j.com/videos/nodes-ai-2026-memmachine-agents-that-learn-memory-that-lasts https://neo4j.com/videos/nodes-ai-2026-the--agent-memory-landscape https://neo4j.com/videos/nodes-ai-2026-the-agent-memory-landscape https://neo4j.com/videos/nodes-ai-2026-the-ai-agent-memory-landscape https://noqta.tn/en/blog/harness-engineering-ai-agents-production-ready-2026 https://onyx.app/insights/enterprise-rag-platforms-2026 https://open-techstack.com/blog/pgvector-vs-qdrant-2026 https://open.substack.com/pub/alexewerlof/p/owasp-top-10-ai-llm-agents https://open.substack.com/pub/alexewerlof/p/owasp-top-10-ai-llm-agents?r=18prwo https://openreview.net/forum?id=GROUNDEDKG-RAG https://openreview.net/forum?id=ParliamentRAG https://papers.cool/arxiv/2609.05110 https://papers.cool/arxiv/2609.25991 https://pixle.blog.csdn.net/article/details/161263511 https://preuve.ai/blog/ai-memory-systems-statistics-2026 https://priyankavergadia.substack.com/p/10-types-of-rag-you-need-to-know https://proceedings.iclr.cc/paper_files/paper/2026/file/fd1eff9dd295df50a41f2521942fa31d-Paper-Conference.pdf https://promptfoo.dev/blog/rag-poisoning https://pub.towardsai.net/all-you-need-to-know-about-chunking-in-agentic-rag-bdc34c21a6e6 https://ragaboutit.com/7-rag-failure-modes-crippling-enterprise-deployments-in-2026 https://ranksquire.com/2026/02/26/best-vector-database-rag-applications-2026 https://rauljitechnologies.com/blog/agentic-rag-2026 https://research.google/blog/chain-of-agents-large-language-models-collaborating-on-long-context-tasks https://research.google/blog/unlocking-dependable-responses-with-gemini-enterprise-agent-platforms-agentic-rag/ https://semanticscholar.org/paper/SoK-Agentic-RAG-Mishra-Niroula/e23f86a3021bb728af4a2d7f70cccb50ff91f4eb https://shattered.io/hugging-face-agent-intrusion-anatomy-2026 https://simo-online.com/en/blog/rag-retrieval-augmented-generation-unternehmen-leitfaden-2026 https://skywork.ai/slide/en/harnessing-ai-agents-2056955859996663809 https://spheron.network/blog/vllm-vs-sglang-2026 https://splx.ai/blog/rag-poisoning-in-enterprise-knowledge-sources https://squirro.com/squirro-blog/state-of-rag-genai https://supermemory.ai/blog/hybrid-search-guide https://tech.yahoo.com/ai/article/openai-just-disclosed-more-concerning-ai-behavior-following-the-hugging-face-incident-heres-everything-you-need-to-know-154529895.html https://theaiengineer.substack.com/p/the-ai-agents-stack-2026-edition https://theaiengineer.substack.com/p/what-is-agent-memory https://theaiengineer.substack.com/p/what-is-rag-retrieval-augmented-generation https://theaiengineer.substack.com/p/why-ai-agents-keep-failing-in-production https://thenewstack.io/graphrag-multi-hop-reasoning-python https://tianpan.co/guides/rag-vs-long-context-decision-framework https://turingpost.com/p/ragtypes https://usewire.io https://varunpratap.com/blog/superlocalmemory-vs-mem0-benchmark-2026 https://vectorize.io/articles/owasp-asi06 https://wavect.io/blog/rag-vs-finetune-vs-longcontext-2026 https://www.actian.com/blog/databases/how-to-evaluate-vector-databases-in-2026 https://www.alphaxiv.org/abs/2604.04936 https://www.alphaxiv.org/abs/2604.09666 https://www.alphaxiv.org/abs/2607.26497 https://www.amazon.science/publications/keyword-search-is-all-you-need-achieving-rag-level-performance-without-vector-databases-using-agentic-tool-use https://www.articsledge.com/post/lazygraphrag-retrieval-augmented-generation https://www.codecon.sk/en/whats-new-in-rag https://www.databricks.com/blog/long-context-rag-performance-llms https://www.digitalapplied.com/blog/hybrid-search-bm25-vector-reranking-reference-2026 https://www.falkordb.com/blog/how-graph-architecture-grounds-rag-and-prevents-hallucinations https://www.firecrawl.dev/blog/best-chunking-strategies-rag https://www.getfeather.store/theory/agentic-search-vs-vector-database-2026 https://www.glean.com/perspectives/top-10-enterprise-use-cases-for-rag-models-in-2026 https://www.harness.io/blog/ai-deployment-in-production-orchestrate-llms-rag-agents https://www.lancedb.com/blog/hybrid-search-rag-for-real-life-production-grade-applications-e1e727b3965a https://www.linkedin.com/posts/muhammad-arslan-b591b1285_research-aiagents-agenticai-activity-7474154906019438592-erLD https://www.meilisearch.com/blog/graph-rag https://www.microsoft.com/en-us/research/blog/lazygraphrag-setting-a-new-standard-for-quality-and-cost https://www.paperclipped.de/en/blog/graph-rag-production https://www.percona.com/blog/vector-distance-vector-similarity-search-in-percona-server-for-mysql https://www.sphereinc.com/blogs/hybrid-search-enterprise-rag https://www.spheron.network/blog/vllm-vs-sglang-2026 https://www.strike48.com/post/agentic-ai-security-risks https://www.turingpost.com/p/ragtypes https://www.varunpratap.com/blog/superlocalmemory-vs-mem0-benchmark-2026 https://www.youtube.com/watch?v=9BpXVodhALk https://www.youtube.com/watch?v=ZNFFuzuwSBc https://www.youtube.com/watch?v=rEITYxTJggU https://www.zartis.com/rag-in-production-what-nobody-tells-you-until-youre-debugging-it https://xingyun3d.csdn.net/69f019380a2f6a37c5a67fba.html https://developer.jdcloud.com/article/4382 https://adg.csdn.net/695331d35b9f5f31781bc374.html https://edu.csdn.net/course/detail/37529 https://gradientflow.com/eight-security-assumptions-ai-agents-quietly-break/ https://gradientflow.com/the-part-of-the-ai-security-story-i-keep-thinking-about/ https://gradientflow.com/the-ai-data-problem-moved-downstream/ https://gradientflow.com/your-model-is-easy-to-buy-your-data-pipeline-is-not/ https://gradientflow.com/six-reasons-i-think-open-ai-models-will-win/ https://blog.csdn.net/a2875255254060/article/details/156463704 https://blog.csdn.net/2301_79319949/article/details/160134772 https://blog.csdn.net/2301_80117363/article/details/163054080 https://blog.csdn.net/2401_84494441/article/details/160311466 https://blog.csdn.net/2401_85300269/article/details/164812712 https://blog.csdn.net/2401_87746054/article/details/162119032 https://blog.csdn.net/2403_87740391/article/details/160400735 https://blog.csdn.net/2501_92406411/article/details/158391326 https://blog.csdn.net/2611_95335967/article/details/164082693 https://blog.csdn.net/AgentInsight/article/details/162166516 https://blog.csdn.net/CYYYCDZ/article/details/163500980 https://blog.csdn.net/Maolei_NyaRu_/article/details/163861971 https://blog.csdn.net/Python_cocola/article/details/161121337 https://blog.csdn.net/enheng1238/article/details/163363550 https://blog.csdn.net/huangmingleiluo/article/details/149409536 https://blog.csdn.net/keshi_curry/article/details/160235298 https://blog.csdn.net/m0_59164520/article/details/160314591 https://blog.csdn.net/m0_59235245/article/details/162430133 https://blog.csdn.net/m0_59235945/article/details/162176846 https://blog.csdn.net/m0_69581581/article/details/162674291 https://blog.csdn.net/m0_69581581/article/details/162706987 https://blog.csdn.net/m0_69581581/article/details/163142070 https://blog.csdn.net/m0_69581581/article/details/163481543 https://blog.csdn.net/qcx23/article/details/160820786 https://blog.csdn.net/qcx23/article/details/160959849 https://blog.csdn.net/qq_27504375/article/details/163024278 https://blog.csdn.net/qq_31429225/article/details/163084256 https://blog.csdn.net/qq_35160742/article/details/163271979 https://blog.csdn.net/qq_45657541/article/details/163280072 https://blog.csdn.net/qq_56999332/article/details/161400644 https://blog.csdn.net/qq_73472828/article/details/161128801 https://blog.csdn.net/smallym1/article/details/163407419 https://blog.csdn.net/srk950606/article/details/156913534 https://blog.csdn.net/weixin_27310985/article/details/164650777 https://blog.csdn.net/weixin_29045585/article/details/163284405 https://blog.csdn.net/weixin_29046611/article/details/163992955 https://blog.csdn.net/weixin_33212799/article/details/164766449 https://blog.csdn.net/weixin_33806859/article/details/165183806 https://blog.csdn.net/weixin_42509513/article/details/160697415 https://blog.csdn.net/weixin_42561143/article/details/160044073 https://blog.csdn.net/weixin_43705457/article/details/163801743 https://blog.csdn.net/weixin_47547625/article/details/161535116 https://blog.csdn.net/weixin_52208686/article/details/162926745 https://blog.csdn.net/weixin_58753619/article/details/161024790 https://blog.csdn.net/xx_nm98/article/details/162153537 https://blog.csdn.net/xx_nm98/article/details/163927291 https://blog.csdn.net/yonggeit/article/details/163207114 https://blog.csdn.net/zxsxj/article/details/161504705 https://csdn.net/2301_77803292/article/details/162705936 https://csdn.net/article/2026-07-30/163342366 https://csdn.net/huang9537638381/article/details/157699556 https://csdn.net/kaka0722ww/article/details/161592550 https://csdn.net/likuolei/article/details/156766693 https://csdn.net/m0_59164520/article/details/161121236 https://csdn.net/m0_69581581/article/details/163542988 https://csdn.net/m0_69581581/article/details/163842030 https://csdn.net/weixin_33269743/article/details/162160811 https://csdn.net/weixin_35774598/article/details/162381479 https://csdn.net/weixin_37763484/article/details/148407621 https://csdn.net/weixin_42118737/article/details/149105217 https://csdn.net/yangshangwei/article/details/155892912 https://www.csdn.net/2301_77803292/article/details/162705936 https://www.csdn.net/a2875254060/article/details/156463704 https://www.csdn.net/huang9537638381/article/details/157699556 https://www.csdn.net/kaka0722ww/article/details/161592550 https://www.csdn.net/likuolei/article/details/156766693 https://www.csdn.net/m0_59164520/article/details/161121236 https://www.csdn.net/m0_69581581/article/details/163542988 https://www.csdn.net/weixin_33269743/article/details/162160811 https://www.csdn.net/weixin_35774598/article/details/162381479 https://www.csdn.net/weixin_37763484/article/details/148407621 https://www.csdn.net/weixin_42118737/article/details/149105217 https://www.csdn.net/yangshangwei/article/details/155892912 https://www.csdn.net/yonggeit/article/details/162179765

本次变更

R113 第 113 轮（2026-10-07 13:55 CST）承接 R112-R82。 R113 增量 / 2 件 RAG 主分类 net-new + 1 件强邻接 + 4 件邻接级（Oct 5 提交 · Oct 6 建卡）：① Bounded Provisional Visibility 2610.05826v1 paper_card 1687（持续 ingestion 时间窗口攻击面防御 · fail-closed provisional-visibility 协议 · 主分类 rag · ⭐⭐⭐⭐）+ ② AI-Decision Checkpoints 2610.06207v1 paper_card 1686（RAG 作为 BPM 一等设计要素 · 主分类 rag · ⭐⭐⭐）+ ③ Source Learning 2610.02150v1 paper_card 1684（从检索到能力范式跃迁 · 主分类 agent · 副分类 rag · ⭐⭐⭐⭐ · DOI 10.48550/arxiv.2610.02150）+ ④ Agentic-ZTA 2610.05782v1 paper_card 1688（RAG 嵌入零信任 · NIST SP 800-207 · ⭐⭐⭐⭐）+ ⑤ When Does Selection Replace Extraction 2609.34227 paper_card 1698（Jev typed decision · 预注册 · ⭐⭐⭐⭐）+ ⑥ UndoBench 2610.05622（任务能力 vs 故障恢复解耦 · 5760 执行 · ⭐⭐⭐⭐）+ ⑦ jay 10-07 CSDN 工业视角 + R112 全部锚点续立 + R111 7 件锚点续立。 R113 八角张力 58 → 60 维（新增 2 维 + 强化 6 维：持续 ingestion 时间窗口攻击面 + BPM 一等设计要素 + 工业部署多源印证 + Defense 轴+协议级门控 + Agent Memory 范式 + 范式对比 + 工业部署 + 评测体系）。共识 212-214 + 争议 188-189 + 开放 257-258 + 趋势 322-324（R113 新预备）

R114 第 114 轮（2026-10-08 13:55 CST）承接 R113-R82。 R114 增量 / 4 件 RAG 主分类 net-new + 1 件强邻接（δ-mem arXiv 补齐）+ 1 件 RAG 综述重锚 + 4 件邻接级（Oct 6-7 提交 · Oct 8 建卡）：① RAG-PIBench 2610.08571v1 paper_card 1700（RAG 提示注入检测基准 · 4,876 三档 leakage-aware · DistilBERT F1=0.896 / PR-AUC=0.968 · 主分类 rag · Defense 轴五节点新节点 · ⭐⭐⭐⭐）+ ② UNREAL 2610.08463v1 paper_card 1701（model-native 统一 RAG 与 Long-Context · <500K · 3B-token Wikipedia · 主分类 rag · Efficiency 轴新节点 · ⭐⭐⭐⭐）+ ③ EC-RAG 2610.08674v1 paper_card 1699（视频事件链 · training-free · 主分类 rag · Reasoning 轴新节点 · ⭐⭐⭐⭐）+ ④ Agentic AutoRAG 2610.08452v1 paper_card 1706（推理驱动 RAG 超参优化 · failure attribution · 前 10 次 vs 30 次 · 主分类 rag · Efficiency+Interactivity 轴双轨 · ⭐⭐⭐⭐）+ ⑤ δ-mem 2605.12357v2 NeurIPS 2026（Efficient Online Memory · Mind Lab NTU+复旦+上交+CUHK+HKUST-GZ · 8×8 OSAM · 4.87M/0.12% Qwen3-4B · 5 benchmark 46.79%→51.66% · MemoryAgentBench 1.31× / LoCoMo 1.20× · 主分类 agent · 强邻接 · ⭐⭐⭐⭐ 升级自 R110 ⭐⭐⭐ Substack · DOI 10.48550/arxiv.2605.12357）+ ⑥ RAG Foundation Survey 2312.10997 paper_card 438（Gao et al. · S2被引 4207 / 影响力被引 264 · OpenAlex W4389984066 · Naive/Advanced/Modular 三阶段范式 · 主分类 rag · ⭐⭐⭐⭐⭐ · 综述 1+1 双锚 2023 三阶段 + R109 2026 四轴）+ ⑦ 4 件邻接级（DeCoPrune 2609.39096 + Structuring MoE 2610.07332 + Sensor-Language-Action 2610.08244 + DMAD 2610.02188）+ R113-R111 全部锚点续立。 R114 八角张力 60 → 62 维（新增 2 维 + 强化 8 维：① RAG 表征层统一证据选取维度 ② RAG 提示注入检测标准化维度 ③ 续立 持续 ingestion 时间窗口攻击面 ④ 续立 BPM 一等设计要素 ⑤ 续立 工业部署失败率多源印证 ⑥ 续立 Defense 轴+协议级门控 ⑦ 续立 Agent Memory 范式 ⑧ 续立 范式对比 ⑨ 续立 工业部署 ⑩ 续立 评测体系）。共识 215-218 + 争议 190 + 开放 259-260 + 趋势 325-327（R114 新预备）

R115 第 115 轮（2026-10-09 13:55 CST）承接 R114-R82。 R115 增量 / 3 件 RAG 主分类 net-new + 1 件 Defense 轴邻接级正式入档 + 1 件 Jay 工业视角 + 1 件 Substack 续立（Oct 7-9 提交 · Oct 8-9 建卡）：① RECAST 2610.10507v1 paper_card 1717（Google DM · 自适应证据路由 · "检索文档"→"推导证据" · 主分类 rag · 副分类 agent · ⭐⭐⭐⭐ · Reasoning 轴结构化推理新节点）+ ② ExperienceIndex 2610.10091v1 paper_card 1716（MIT/Microsoft/UPenn · Jacob Andreas · artifact-grounded 经验记忆 · 在线质量提升 + 成本下降 · 主分类 agent · 副分类 rag · 强邻接 · ⭐⭐⭐⭐ · 交互性轴新节点）+ ③ Memory Portability 2609.05339v1 paper_card 1238（受控研究 · 3 关键发现：方向特定迁移测试 + embedding space 隔离 + 源历史保留 · 主分类 rag · 副分类 agent · ⭐⭐⭐⭐ · 交互性轴新维度 · RAG 生产模型升级流程标准化）+ ④ EAL-Bench 2609.01836 paper_card 1187（5+2 LLM × 3 场景授权漂移评测 · 主分类 agent · 副分类 rag · S2被引 3 · DOI 10.48550/arxiv.2609.01836 · ⭐⭐⭐⭐ · Defense 轴九节点标准化体系新增第 9 节点）+ ⑤ Jay CSDN Oct 9 工程视角（手写 RAG 完整流水线 + 企业 RAG 成本分析 + Agentic RAG ADK 框架 + 多模态 RAG Dify Pipeline + 12-Factor Agents + 七层企业 AI 架构 · Top-1 58.3%→82.5% +24.2pp / 跨模态召回 61.7%→89.2% +27.5pp / 成本 ↓33% · ⚠️ 召回率 +17% / 通义灵码 +23% 待原论文核验）+ ⑥ Substack theaiengineer 2026 Memory 三层架构续立（in-context + agentic + shared infrastructure · Mem0/Letta/Zep + LongMemEval 新基准）+ ⑦ R115 fresh web_search 5 源（Progress Software + Lyzr + Squirro + Atlan + Turingpost · 60+ 格式 · 八模式 · 2026 主导模式 Agentic RAG）+ R114 全部锚点续立 + R113 全部锚点续立 + R112 全部锚点续立 + R111 7 件锚点续立。 R115 八角张力 62 → 63 维（新增 1 维 + 强化 9 维：① RAG 记忆系统可移植性维度 Memory Portability ②-⑪ 续立 8 维）。共识 219-222 + 争议 191 + 开放 261-263 + 趋势 328-330（R115 新预备）

Tom · 2026-10-09 13:55 CST · R115 Wave3 E1 · 3 RAG 主分类 net-new（RECAST 2610.10507 Google DM + ExperienceIndex 2610.10091 MIT/Microsoft/UPenn + Memory Portability 2609.05339）+ 1 Defense 轴 EAL-Bench 2609.01836 + 1 Jay CSDN 工程（Top-1 +24.2pp / 跨模态召回 +27.5pp / 成本 ↓33%）+ 1 Substack theaiengineer Memory 三层架构 + 1 fresh web_search（Progress+Lyzr+Squirro+Atlan+Turingpost）+ R114-R111 全部锚点续立 + 63 维张力