菜单
Answer questions from a document store with citations, then verify groundedness with a judge before accepting -- retrieval-augmented QA that refuses to hallucinate.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/KrakenNet/stargraph --skill rag-qa命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
Plan a coding task, generate and execute Python, and loop on reviewer feedback until the code passes -- a plan/code/judge fix-loop with governed execution.
Research a question with web search and page reads, judge the answer for completeness, and keep digging at the named gaps until nothing major is missing.
Draft an answer, judge it against a rubric, and refine on the judge's written feedback until it passes -- the generate/evaluate/optimize loop.
基于 SOC 职业分类
正在显示 SKILL.md
| name | rag-qa |
| description | Answer questions from a document store with citations, then verify groundedness with a judge before accepting -- retrieval-augmented QA that refuses to hallucinate. |
| subgraph | graph.yaml |
Retrieve → answer → groundedness check (benchmark T6). A rag node
retrieves from the bound store and answers with citations; a judge
node verifies every claim is grounded in the cited documents. Rejected
answers loop with the judge's feedback; grounded answers halt.
question.answer, citations (ranked retrieved doc ids), verdict,
score, rationale.Point the qa node's store binding at your corpus: edit
stores[0].path in graph.yaml (a SQLiteDocStore database; ingest
with SQLiteDocStore.put). A lancedb binding (FTS, stores extra)
can be added alongside for native ranking.
stargraph run graph.yaml (set the LM via
--lm-model/--lm-url or per-node config.model).kind: subgraph, spec: <path to graph.yaml>;
config.max_steps bounds the retry loop.