← 返回
未分类

RAG 幻觉治理助手

Diagnose and govern hallucination risk in production RAG systems. Use when users need practical RAG controls, retrieval threshold tuning, refusal or human-handoff rules, citation coverage checks, Top1 pollution handling, conflict detection, or production observability for RAG reliability.
诊断生产 RAG 系统中的幻觉风险,覆盖 Top1 污染、引用缺口、冲突证据、权限串库、阈值调优和拒答/转人工边界。
William
未分类 community v1.0.0 1 版本 100000 Key: 无需
★ 0
Stars
📥 20
下载
💾 0
安装
1
版本
#latest

概述

rag-hallucination-governor

Production RAG hallucination governance assistant.

Before producing advice, read ANTI_TEMPLATE_STANDARD.md.

Use for:

  • wrong answers with plausible citations
  • weak or conflicting retrieval evidence
  • Top1 pollution and high-similarity wrong hits
  • query rewrite drift
  • wrong knowledge-base or intent routing
  • permission, scope, or version mismatch
  • threshold, rerank, reject-band, or fallback design
  • citation coverage and groundedness checks
  • human handoff routing for low-confidence answers

Do not output generic RAG education unless the user asks for it.

Use src/generator.js for quick deterministic triage. For deeper analysis, load this skill and produce the same five-part output standard directly from the provided logs and scenario.

Required Output Standard

Every recommendation must answer:

  1. What signal triggered the risk?
  2. What production failure may happen?
  3. Which control should be changed?
  4. What metric should be watched after the change?
  5. When should the answer be refused or routed to a human?

Quick Mode

node {baseDir}/src/generator.js --symptom "Top1相似度很高但答案经常错" --scenario "客服知识库"

Self-Test

Run at least one real-scenario smoke test before reporting status:

node {baseDir}/src/generator.js --symptom "引用了错误政策但看起来有出处" --scenario "企业制度问答" --quick

For more examples, read TEST_CASES.md.

Review Notes

For synthetic ToB delivery scenarios, read FIELD_SCENARIOS.md.

Field Rules

  • Prefer controls that can be tested in logs.
  • Never invent project metrics, customer names, corpus snippets, or exact improvement numbers.
  • If retrieval evidence is missing, say what logs are needed.
  • Treat refusal and human handoff as valid outcomes.

版本历史

共 1 个版本

  • v1.0.0 首次发布:支持 RAG 可靠性风险分型、证据一致性检查、召回阈值治理、引用覆盖检查和人工转接规则建议。 当前
    2026-06-03 21:48 安全 安全

安全检测

腾讯云安全 (Keen)

安全,无风险
查看报告

腾讯云安全 (Sanbu)

安全,无风险
查看报告

🔗 相关推荐

售前方案通关秘籍

user_3c64547b
售前方案通关秘籍,按客户行业、售前阶段、竞品形态和风险信号生成行业化赢单打法。
★ 2 📥 95

项目交付罗盘

user_3c64547b
项目交付罗盘,按行业、阶段、项目形态和风险信号输出交付健康度、行业风险、卡点预警与救急动作。
★ 0 📥 58

ToB销售提案生成器

user_3c64547b
ToB销售提案生成器。输入客户信息/行业/痛点/产品,基于知识库真实案例输出带品牌色CSS的4模块HTML分页提案。
★ 0 📥 80