Tag
LocalLightChat (LLC) 推出重大更新,新增 Agents、MCP、RAG 和图像编辑功能,同时保持轻量级特性,可运行在老旧硬件上,作为 Open WebUI 的轻量替代方案。
The author describes building sift-light, an evidence-first local search layer for coding agents that prioritizes exact search and verifiable sources over plausible semantic guesses, with new visual retrieval for images and scanned PDFs on Apple Silicon.
提出 T2Mem 框架,将记忆构建延迟到推理阶段,随推理过程渐进式构建面向查询的记忆,在 LoCoMo 和 LongMemEval 上超越强基线并显著降低记忆构建开销。
本文提出 KV-Repurpose,一个基于 vLLM 的开源测试平台,用于评估超越精确前缀匹配的分段 KV-cache 复用;在 Zipfian RAG 负载下,其 token 级缓存命中率达到 63.9%,远高于前缀缓存的 3.3%,并显著降低 TTFT。
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.
We present SmartSalesBench, the first multi-turn conversational benchmark designed to evaluate LLM sales agents against psychologically grounded, resistant buyer personas, with nine evaluation metrics and experiments across four models totaling 13,680 sessions.
A practitioner asks enterprise AI deployers which requests from clients they still can't fulfill, citing poor internal adoption (e.g., one sysadmin reported only ~5 of 300 users actually using it), and flags gaps like company-specific knowledge, answer provenance, and access control.
The author argues that EmbeddingGemma 2 and embedding models in general are being underestimated, highlighting the open-source DigUp project — a Mac search app built on Google's EmbeddingGemma 2 that outperforms Spotlight with content, image, and audio search — and suggests embeddings could also improve RAG knowledge base search.
一个求职者分享了面试 AI Agent 实施团队时被 RAG 检索细节问题难住的经历,并总结出生产级 RAG 需结合向量检索与 BM25 关键词搜索的混合检索,再用 Cross-encoder Reranker 进行二次精排,以解决语义相近但实体错误的检索问题。
A 10,000-word guide sharing the learning path to becoming an Agent engineer, covering LLM fundamentals, Prompt Engineering, RAG, Agent applications, and backend engineering. Includes supporting learning materials, project requirements, and acceptance criteria to help readers systematically fill in their AI application development skills.
Garry Tan分享使用Capydotai驱动延迟与检索评估自动改进循环的体验,AI代理在过去16小时内自动运行并持续优化GBrain检索指标,无需人工干预。
本文介绍了 Uber 法律红线代理从失败的 RAG 演进到四代架构,实现 91% 决策准确率,将合同审查时间缩短 20%+,并赢得 2026 ALM Legalweek 年度最具创新力法务部门奖。文章详细分析了从 RAG 到反馈闭环、语气分层、再到最终代理架构的技术演进路径。
Onyx 推出团队知识与上下文层,通过索引 50+ 应用的内容、元数据与权限,为搜索、研究和 AI 代理响应提供支撑,并提供基于 MIT 许可的免费社区版。
KernelAI 上线 Product Hunt,支持在 iPhone 和 Android 上本地运行 53 个来自 15 家实验室的开源模型,提供聊天、文档 RAG、图像理解与代码编写功能,全程离线且免费无广告。
作者批评把整站页面粗暴灌入向量库的做法,主张先爬取发现 URL、再用简单路径过滤大幅缩减索引规模,因为智能体有任务和 token 预算,无法容忍大量噪声内容。
提出CogMem认知记忆架构,基于PEC²F图谱将对话记忆从被动检索转为主动重构,在LoCoMo和LongMemEval上取得多跳、时序和知识更新任务的良好表现。
The paper introduces PARCEL, a claim-level benchmark for detecting legal hallucinations in LLM output, using 3,396 parenthetical-style claims from New York State Court of Appeals decisions framed as three-way NLI. State-of-the-art LLMs reach up to 0.97 accuracy zero-shot but still mislabel unsupported claims as supported, especially with fabricated yet plausible citations.
介绍如何用 Postgres + pgvector、FastAPI、Gemini、Celery 和 Redis 构建生产级 RAG 流水线:答案引用来源、在不知道时拒绝回答,并在数据库层面强制执行行级安全。
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.
该论文研究了LLM智能体中的'记忆过度依赖'现象——即使记忆是良性的、正确存储且被恰当检索,部分证据迁移仍会误导推理。作者提出MemTrim框架,在写入时索引记忆证据、在读取时控制复用,无需重训即可减少过度依赖。