[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-libredb-studio-local-llm-agent":3,"topics-all":38,"news-related-dcd8b3e1-a3c7-4614-aba4-9002219ea5f6":57},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":24,"news_slug":31,"published_at":32,"created_at":33,"modified_at":34,"is_published":35,"publish_type":36,"image_url":14,"view_count":37},"dcd8b3e1-a3c7-4614-aba4-9002219ea5f6","LibreDB Studio 0.15 发布:本地 LLM 接管数据库交互","LibreDB Studio 0.15 把自托管 SQL IDE 与 LLM Agent 整合,支持 Gemini、OpenAI 或本地 Ollama。AI 基于数据库结果回答问题、对比执行计划;调用走只读受审计管线,写入与 DDL 在到达驱动前被拒。","把数据库工具直接跑在数据旁边,而不是装到每台工程师的笔记本上——这是开源 SQL IDE LibreDB Studio 0.15 最显眼的设计取向。MIT 协议的项目 9 月 8 日发布这个更新,把\"LLM Agent 当成 SQL 工具一等公民\"做成了一个可在生产环境使用的方案。\n\n## 一个浏览器入口覆盖十六种引擎\n\nLibreDB Studio 部署在容器、Helm Chart 或 OpenShift Operator 里,或用 `npx @libredb\u002Fstudio` 一行命令起在服务器;浏览器、手机、Windows \u002F macOS \u002F Linux 桌面打开同一个界面。它覆盖十六种数据库引擎(PostgreSQL、MySQL、Oracle、SQL Server、SQLite、libSQL、DuckDB、MongoDB、Redis、Couchbase、ClickHouse、Apache Druid、Elasticsearch、OpenSearch、Apache Trino、Apache Cassandra),共享同一套 schema 浏览器、ER 图、schema 对比与监控面板。PostgreSQL 项目、Redis、ClickHouse、Trino 与 Apache Cloudberry 官方文档都把它列入推荐列表——对开源工具是不常见的信誉信号。\n\n## Agent 走的是受审计的只读管线\n\n0.15 把 AI 从\"查询旁的解释按钮\"升级成\"编辑器旁的 Agent 面板\"。用户写下目标——\"哪个部门员工最多?\"\"这条查询为什么慢?\"——按 Start,Agent 起草 SQL、读取结果,最后写一份带引用支撑的报告。\n\nAgent 不靠解析 SQL 挡危险,而是把只读交还给数据库:PostgreSQL 走只读事务;SQLite 每条语句前重新声明 `PRAGMA query_only`;DuckDB 用 `READ_ONLY` 引擎句柄加 SQL 守卫。所有写入与 DDL 在到达驱动前即被拒;`EXPLAIN ANALYZE` 因会真执行而默认禁止。这条受审计管线是 Agent 独享的——用户在编辑器里直接跑 SQL 的路径不走它,也不产生审计事件。\n\n模型支持范围写得明确:默认 Gemini 2.5 Flash,同样支持 OpenAI、Ollama 或任何兼容 OpenAI 的端点;Agent 模式要求模型具备工具调用能力,且 README 警告 Ollama 上要靠一次真实探测来确认。\n\n## 只读档案只在三款数据库上存在\n\nAgent 模式的只读执行档案由数据库原生提供,只在 PostgreSQL、SQLite 与 DuckDB 上存在;其他十三种引擎跑 Agent 模式会以 `engine-unsupported` 收尾。Plan 模式作为兜底,完全不使用工具、不访问数据,只把目标翻译成 SQL 草稿给用户自己跑;它的 schema grounding 在所有十六种引擎上都可用。每次 Agent 运行都有显式上限:最多 20 条语句、60 秒数据库时间、单次读取 200 行、整轮 5 分钟 deadline。\n\n## 边界在哪里\n\n开源 SQL 工具接 LLM Agent 不是新想法,但这条路线有几个取舍值得留意:把\"工具去找数据\"和\"用户自己用工具\"切成两条独立路径;把 Agent 模式的边界用数据库原生能力来表达,而不是项目方自己写 SQL parser——后者在生产里几乎都漏过;把协议选成 MIT 而非更严格的源码可用许可,因为\"不能把按席位收费的工具放进每一个你拥有的环境\"——这是项目方在 README 里写出来的明确立场。对需要在自己服务器上跑数据库工具、且希望用本地 LLM 保持数据不外流的团队,这是一个值得评估的选项,前提是接受 Agent 模式只在三种数据库上可用,把其他引擎交给 Plan 模式。\n","https:\u002F\u002Fgithub.com\u002Flibredb\u002Flibredb-studio\u002Fblob\u002Fmain\u002FREADME_zh.md","998df6db-96e6-4b8e-8be1-cfa00a6cd177",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"6ad31a14-c0da-42df-81fd-564281f768db","agentic-ai",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"40269b40-7942-4650-9672-ed2e6524d37a","ai-technology",{"id":19,"name":20,"slug":20,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":22,"name":23,"slug":23,"description":14,"color":14},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"8a4f999e-8e39-4fe7-b48c-c6f95f07a7a1","en","LibreDB Studio 0.15: Bringing a local LLM to the database","LibreDB Studio 0.15 ties a self-hosted SQL IDE to an LLM Agent. Supports Gemini, OpenAI, or local Ollama. The Agent drafts SQL, reads results, writes a cited report. Every call goes through an audited read-only pipeline backed by database-native mechanisms.","Putting the database tool right next to the data, instead of installing it on every engineer's laptop — that's the most visible design posture of the 0.15 release of open-source SQL IDE LibreDB Studio. The MIT-licensed project published this update on September 8 and turned \"LLM Agent as a first-class SQL citizen\" into something usable in production.\n\n## Sixteen engines, one browser entry\n\nLibreDB Studio deploys as a container, a Helm Chart, or an OpenShift Operator; it also starts on a server with `npx @libredb\u002Fstudio`. The browser, a phone, and Windows \u002F macOS \u002F Linux desktops open the same interface. It covers sixteen database engines (PostgreSQL, MySQL, Oracle, SQL Server, SQLite, libSQL, DuckDB, MongoDB, Redis, Couchbase, ClickHouse, Apache Druid, Elasticsearch, OpenSearch, Apache Trino, Apache Cassandra), all sharing one schema browser, ER diagrams, schema diff, and monitoring panel. The PostgreSQL project, Redis, ClickHouse, Trino, and Apache Cloudberry all list it as recommended in their official docs — an unusual credibility signal for an open-source tool.\n\n## The Agent runs on an audited, read-only pipeline\n\nVersion 0.15 upgrades AI from \"an explain button next to the query\" to \"an Agent panel beside the editor.\" The user writes down a goal — \"which department has the most employees?\" \"Why is this query slow?\" — presses Start, and the Agent drafts SQL, reads the results, and finally writes a report whose every claim cites the source row it came from.\n\nBut this Agent does not block danger by parsing SQL; it hands read-only enforcement back to the database itself. PostgreSQL uses a read-only transaction. SQLite re-asserts `PRAGMA query_only` before every statement. DuckDB uses a `READ_ONLY` engine handle plus an SQL-level guard. All writes and DDL are rejected before they reach the driver. `EXPLAIN ANALYZE` is denied by default because it would actually run the statement. This audited pipeline belongs to the Agent alone — SQL the user runs directly in the editor does not go through it, and produces no audit event.\n\nSupported models are spelled out clearly: Gemini 2.5 Flash is the default; OpenAI, Ollama, or any OpenAI-compatible endpoint is supported the same way. Agent mode requires tool-calling capability, and the README warns that on Ollama, a live probe — not the vendor page — is what establishes that.\n\n## The read-only profile only exists on three databases\n\nThe read-only execution profile is database-native, so it exists only where a provider implements it — PostgreSQL, SQLite, and DuckDB. On the other thirteen engines, Agent mode runs end with `engine-unsupported`.\n\nPlan mode is the fallback: it uses no tools, accesses no data, and only translates a goal into a SQL draft for the user to run themselves; its schema grounding works on all sixteen engines. Every Agent run has explicit caps: at most 20 statements, 60 seconds of database time, 200 rows per read, a 5-minute deadline for the whole run.\n\n## Where the boundary sits\n\nPlugging an LLM Agent into an open-source SQL tool isn't new. But this line has a few choices worth noting: \"the tool goes to the data\" and \"the user uses the tool themselves\" are two distinct paths; the Agent-mode boundary is expressed in terms of native database capabilities, rather than the project writing its own SQL parser — the latter tends to leak in production; the license is MIT rather than a stricter source-available license, because \"you cannot place a per-seat licensed, feature-gated tool into every environment you own\" — a position the project writes out plainly in the README. For teams that need to run a database tool on their own servers and keep data from leaving via a local LLM, this is an option worth evaluating, on the condition that Agent mode is only available on three databases and the rest are handed to Plan mode.\n","libredb-studio-local-llm-agent","2026-09-15T00:00:00Z","2026-09-15T09:15:23.520283Z","2026-09-15T09:15:23.520290Z",true,"agent",34,[39,48],{"slug":40,"tag_slug":40,"title_zh":41,"title_en":42,"intro_zh":43,"intro_en":44,"id":45,"is_active":35,"created_at":46,"modified_at":47},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":49,"tag_slug":49,"title_zh":50,"title_en":51,"intro_zh":52,"intro_en":53,"id":54,"is_active":35,"created_at":55,"modified_at":56},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":58},[59,64,69,74,79,84],{"id":60,"title":61,"news_slug":62,"published_at":63},"777afb24-262f-45cc-961f-d5d49ad42883","AgentOPSD 用递归贝叶斯信念破解多轮 Agent 强化学习的信用分配：清华\u002F浙大\u002F美团让 GRPO 学会看哪个 turn 决定胜负","agentopsd-recursive-belief-credit-assignment","2026-08-07T02:00:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"af056e63-5622-48ae-8629-5226aed64767","PalmClaw 把端侧 Agent 拉进「原生」时代:94.9% 完成时间压缩 + 11.5% 成功率提升","palmclaw-on-device-agent","2026-07-15T20:10:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"3ebeac99-ddd6-432d-a97f-aab8ec609baa","OPID 把\"已完成轨迹\"变成训练信号：Agentic RL 第一次有了\"事后诸葛亮\"式的密集监督","opid-agentic-rl-hindsight-skill","2026-06-27T20:10:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"98f60b3b-5001-4df9-bd5a-a661b1575724","OpenEnv 升格为多机构共治：HF、NVIDIA、Meta 等 9 家共建 agentic RL 的「通用插座」","openenv-hf-nvidia-meta-9-socket","2026-06-14T16:00:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"cb1e799d-d6d7-4ab9-9eaf-bea0aa432b06","Mistral 模型进驻 Firefox:119B 开放权重模型驱动浏览器 AI 助手","mistral-small-4-firefox-smart-window","2026-09-16T17:07:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"5535b4e4-21de-4ed6-a9bb-2b4d824e6568","F-Droid 一次更新的 102 款应用,72.5% 主要是 AI 写的","f-droid-72-percent-ai-written-audit","2026-09-15T15:12:16+00:00"]