ia-qa.com/mcp llm and RAG testing - Dev/QA toolbox
IA-QA as an MCP Server
目录 / 搜索检索 / rag-grounding-guard
--- name: rag-grounding-guard display_name: RAG事实溯源校验(Grounding Guard) display_name_en: "RAG Grounding Guard" description: "当用户说『RAG回答胡说八道/编造』『引用对不对』『回答有依据吗』『检索到的资料支撑不了结论』,或要给 RAG/检索增强回答做事实校验时使用。校验每条声明在检索来源里是否有支撑( grounding 覆盖率),未覆盖的声明标红为幻觉风险,并要求来源带出处(有籍)。可运行脚本(grounding_check 校验器)。理论根基:LGD 三律之有籍(引用溯源)+有证(防幻觉可核验)。触发词:RAG校验、grounding、事实溯源、防幻觉、引用核查、检索支撑、hallucination、回答有依据吗。" version: 1.1.1 agent_created: true author: 诺声(Logos)@SynomosAI license: MIT license_scope: 代码(.py 文件)MIT;本 SKILL.md 与其中理论文本不在 MIT 覆盖范围内(见文末「License & attribution」) allowed-tools: Read, Glob, Grep, WebSearch category: AI工程方法 platforms: [windows, macos, linux] platform_compatibility: [windows, macos, linux] read_when: - RAG/检索增强的回答出现编造或张冠李戴 - 要核对 AI 声明是否有检索来源支撑 - 上线一个知识库问答前做事实校验 - 用户问「你这话有依据吗」 tags: [RAG校验, grounding, 事实溯源, 防幻觉, 引用核查, 有籍] slug: rag-grounding-guard title: RAG事实溯源校验(Grounding Guard) copyright: SynomosAI description_zh: "当用户说『RAG回答胡说八道/编造』『引用对不对』『回答有依据吗』『检索到的资料支撑不了结论』,或要给 RAG/检索增强回答做事实校验时使用。校验每条声明在检索来源里是否有支撑( grounding 覆盖率),未覆盖的声明标红为幻觉风险,并要求来源带出处(有籍)。可运行脚本(grounding_check 校验器)。理论根基:LGD 三律之有籍(引用溯源)+有证(防幻觉可核验)。触发词:RAG校验、grounding、事实溯源、防幻觉、引用核查、检索支撑、hallucination、回答有依据吗。" description_en: "Verify RAG answers are grounded: per-claim support coverage against retrieved sources, with unsupported claims flagged before the answer ships." classification: skillhub: ["security-compliance"] iso_25010: ["Security"] risk_tier: medium risk_rationale: "门禁批量补丁默认 medium(内部约定:本 skill 设计为中等风险):调用 API/读写文件/本地执行,最小权限,密钥走 Secret;若实际涉对外代执行(push/发布/发送/上架)须人工复评,改 high 并走负责人批准。" ai_protocol: true spec: usk/1.0 interface: type: cli entry_point: scripts/grounding_check.py runtime: python3 call_pattern: args input_schema: claims: type: string description: "claims" required: true sources: type: string description: "sources" required: true output_schema: stdout: 结构化 JSON IR(含 result / rc / error_code / warnings) exit_code: rc=0 正常 / rc=1 告警或从严推定 / rc=2 输入不足或非法 capabilities: - rag - grounding - verification permissions: network: false filesystem: false subprocess: false requirements: python_packages: [] min_python: "3.8" changelog: "1.1.1: Bilingual rewrite (full EN+ZH, English first); 1.0.0: initial upload (USK v3)" --- > 凡自治之物 · LGD 治理学 · 创造术家族 | 评测协议:skill-quality-gate ## English ### When to use - When users question the truthfulness of RAG answers, or the input contains suspected hallucination keywords such as "fabricated" or "nonsense". - When concrete conclusions drawn from retrieval results must be fact-checked, in line with the LGD registered/evidenced principles. - In high-risk decision scenarios or when producing traceable reports, answers must carry supporting source citations. ### Usage Call the core script `grounding_check.py` and specify the check target and sources via command-line arguments: ```bash python grounding_check.py --target <claim-to-check> --source <retrieval-store ID or URL> --mode lgd_grounding ``` - `--target`: the specific factual statement or answer segment to verify. - `--source`: the retrieved knowledge index, URL or related context data. - `--mode`: validation mode; `lgd_grounding` recommended (LGD Three-Law grounding). ### Workflow 1. Extract the key factual claims from the current RAG answer into a list of statements to check. 2. Semantically match claims against retrieved sources and compute the grounding coverage. 3. Run `grounding_check.py` to produce the internal verification result per LGD rules. 4. The script flags unsupported claims as high risk, warning "hallucination risk" in red. 5. The system automatically requires unsupported conclusions to carry source links, ensuring every answer is traceable. 6. Aggregate the analysis into a final credibility assessment with evidence-based correction suggestions. ### Evidence & safety statement - Evidence: TBD (no evaluation / badge link yet; update here once produced) - Version: 1.1.1 - Safety: no independent security scan has been performed (TBD); the conclusion is pending and no endorsement is given until it is completed. ## 中文 ### 何时用 - 当用户质疑 RAG 回答内容的真实性,或输入包含“编造”、“胡说八道”等疑似幻觉关键词时。 - 需要对检索结果中的具体结论进行事实核验,确保符合 LGD 三律之“有籍”与“有证”。 - 在涉及高风险决策或需要生成可追溯报告的场景中,要求答案附带出处链接支撑。 ### 用法 调用核心脚本文件 `grounding_check.py`,通过命令行参数指定校验对象与来源: ```bash python grounding_check.py --target <待校验的文本声明> --source <检索库 ID 或 URL> --mode lgd_grounding ``` - `--target`: 输入需要核查的具体事实陈述或回答片段。 - `--source`: 提供检索到的知识文档索引、URL 或相关上下文数据。 - `--mode`: 指定验证模式,推荐使用 `lgd_grounding`(LGD 三律 grounding)。 ### 流程 1. 提取当前 RAG 回答中的关键事实声明,形成待校验文本列表。 2. 将待校验声明与检索来源进行语义匹配,计算接地覆盖率(Grounding Coverage)。 3. 执行 `grounding_check.py` 脚本,依据 LGD 三律规则生成内部校验结果。 4. 脚本输出中标记未覆盖的声明为高风险,并以红色文本警示“幻觉风险”。 5. 系统自动要求未支撑结论附带来源链接,确保每一条回答“有籍可查”。 6. 汇总分析结果,输出最终的可信度评估报告及基于证据的修正建议。 ### 证据与安全声明 - 证据:TBD(暂无评测/徽章链接;产生后在此更新实际链接) - 版本:1.1.1 - 安全:未做独立安全扫描(TBD),结论待补;补全前不作推荐背书。
| 版本 | 存档时间 | 内容哈希 | 内容 |
|---|---|---|---|
| v1 | 2026-09-29 11:46 | 2a353d12 | 已过保留期 |
| v2 | 2026-10-17 15:59 | e58e5cae | 可取 |
版本索引永久保留;内容副本只保留最新 1 份,更早版本仅留索引与哈希(存档时间线的证据链不会因此断裂)。
IA-QA as an MCP Server
DACH B2B opportunity intelligence with FIT/NEED/NOW scoring and MCP access
Chemical intelligence for compound structure search, fragment analysis, and ADMET property prediction
Provide a comprehensive spellbook for Dungeons & Dragons 5th Edition to enhance your gameplay experience
Calculate and compare global shipping costs including detention, demurrage, and local port charges
Calculate detention, demurrage, and local shipping charges for ports worldwide
侵权举报经核验成立后,我们会即时下线该条目并删除已存的内容副本。