Tag
The author shares a learning roadmap for RAG and memory systems geared toward Agent development and interviews, covering chunking, hybrid retrieval, reranking, vector database selection, and short-term/long-term memory management, along with reference materials including the original RAG paper by Lewis et al.
The author explores how RAG evaluations should distinguish three different reasons a system answers "I don't know" — retrieval failure, missing evidence, and conflicting sources — and proposes evals that label the expected unresolved state rather than just scoring answers as correct, incorrect, or refused.
Qwen's 10-million-document RAG pipeline uses four lines of defense — hybrid retrieval, citation-constrained generation, claim-level verification, and calibrated abstention — to bring the hallucination rate on unanswerable questions down to 2%, while keeping p95 latency at 18ms. It's a production-grade anti-hallucination architecture worth studying.
本文介绍并审计了一个针对 Bengali 幻觉检测挑战赛（Olikbochon）的推理系统，该系统结合响应条件化查询重写、混合检索与任务特化的语言模型裁决，实现了 0.956 的私有宏 F1，同时指出语料重叠与运行不可复现性限制了其泛化性声明。
Presents an evidence-gated bilingual RAG framework for consumer banking search combining hybrid dense-lexical retrieval, cross-encoder reranking, and a three-outcome answer mechanism, improving Top-1 hit rate from 50.48% to 79.05% on 210 Arabic and English queries.
该论文研究不确定性度量何时能有效衡量 RAG 幻觉风险，发现当检索证据充分约束答案时，困惑度能达到 0.81–0.91 AUROC，而检索覆盖不足时同一指标会失效，提出不确定性是条件性的幻觉风险度量。
We present AuditKIT, a source-available evaluation library that unifies benchmark accuracy, judge-based assessment, runtime performance, tool-use behavior, and retrieval quality in a single result representation with direction-aware metric contracts and extensible backends. Seven case studies show how answer extraction, runtime behavior, and metric selection can change evaluation outcomes for LLMs and agentic systems.
该论文提出 ExecutionRAG，一种无需训练的需求寄存器机制，用于分析多跳 RAG 中 LLM 未能检索到证据的原因，发现许多检索失败实为搜索漂移，并指出剩余瓶颈从搜索引导转向证据充分性验证。
该论文提出 MeddeX 系统，结合 QLoRA 微调的 LLaMA-3 (8B)、报告级检索增强生成 (RAG) 以及 Gemini Vision 多模态层，实现对医疗报告的证据支撑式摘要与分析，并在 ROUGE、BLEU 和 BERTScore 上相对基线模型取得一致提升。
该论文提出 GraphRAG-LTM 分层威胁建模框架，测试三种攻击者访问层级（黑盒查询、白盒查询、图感知白盒），发现检索排序操纵与关键三元组删除能显著削弱 GraphRAG 的答案质量与支撑事实检索。
svpino 推介开源工具 XERJ Corpus Hub，它可作为代理的"参考记忆"引擎接入任意 AI 模型，通过本地索引的代码、文档与安全知识库进行检索，让代理无需再训练即可提升表现。
The paper introduces SEEDS, a preview-guided evidence admission framework for search agents that selectively decides which retrieved candidates should enter the agent's reasoning context, improving search effectiveness across nine knowledge-intensive QA benchmarks while reducing full-evidence consumption.
Researchers introduce GeopoliticalQA, a benchmark of 5,525 QA pairs over 415 geopolitical and strategic intelligence reports, evaluating 15 retrieval systems and six open-weight LLMs. They find mid-size encoder bge-m3 matches or beats much larger and commercial embedding models at chunk-level retrieval, while answer generation remains hard even with gold evidence for inferential and risk-oriented questions.
提出 KnowForge，一种按角色分工的多智能体检索增强问答管线（Locator、bridge judge、Deep Reader、Synthesizer），在开放域多模态文档 QA 基准上显著超越单智能体 RAG 与已发布系统。
SignRAG 提出一个统一框架，结合分层预训练、目标域检索增强与检索效用引导的强化微调（RUG-RFT），将 decoder-only LLM 用于无词汇（gloss-free）手语翻译，在多个基准上达到 SOTA，并首次在 CSL-Daily 上全面超越有词汇监督的方法。
RAG-Stress 提出一个受控诊断协议，通过编辑检索证据中的单一断言来诱导错误答案，量化模型在检索增强生成中被误导证据覆盖正确答案的比率（misleading rate），并发现严格遵循文档的指令会系统性提高误导率。
RIT-RAG 提出一种结合内容检索与结构导航的检索增强生成方法：离线为每个文档从目录或站点地图构建树结构，查询时通过检索到的 chunk 位置归纳出可管理的子树，由 LLM 智能体导航并选择性阅读节点。在多个基准上超越现有方法，在其提出的 EntQABench 上较最强基线提升 6.8–11.4 个百分点。
A production-grade RAG checklist covering citations on every answer, refusing to answer when context doesn't support it, non-bypassable access control, and non-blocking ingestion, shared as a YouTube tutorial video.
作者分享对 GraphRAG 的兴趣，并尝试构建一个轻量本体论、无需向量搜索的 Jev-based GraphRAG 变体。
This paper investigates whether forcing LLMs to reason in the same language as the query and retrieved documents improves monolingual non-English RAG performance. Using a German RAG testbed built over the fictional world of The Dark Eye, the authors find that aligning the reasoning language with the evidence helps (German reasoning beats French despite the model's stronger French proficiency), but native unconstrained English reasoning still matches the aligned setting, highlighting the need for truly multilingual reasoning models.