Communitygithub.com

evaluate-ai-release

Evaluate a RAG system or AI-agent release against an accepted baseline, diagnose regressions, and produce explainable PASS, WARN, or BLOCK evidence. Use for candidate traces, RAG evaluation scenarios, model or prompt changes, release gates, citation regressions, abstention checks, and pre-release AI quality reviews.

evaluate-ai-release 是什麼?

evaluate-ai-release is a Claude Code agent skill that evaluate a RAG system or AI-agent release against an accepted baseline, diagnose regressions, and produce explainable PASS, WARN, or BLOCK evidence. Use for candidate traces, RAG evaluation scenarios, model or prompt changes, release gates, citation regressions, abstention checks, and pre-release AI quality reviews.

相容平台~Claude Code~Codex CLI~Cursor
npx skills add https://github.com/thangldw/ragops/tree/main/skills/evaluate-ai-release

在你喜歡的 AI 中提問

開啟一個已預先載入此 Agent Skill 的新對話。

說明文件

Evaluate an AI release

Use RAGOps as a deterministic evaluator. Treat its result as release evidence, not as proof that a system is universally correct or safe.

Workflow

  1. Identify the scenario, candidate responses or traces, accepted baseline, and policy thresholds. Do not invent missing fixtures.
  2. Read references/commands.md when selecting the RAGOps command or interpreting exit codes.
  3. Inspect inputs for secrets or personal data before running an evaluation. Keep generated reports private unless the user approves sharing.
  4. Run the narrowest applicable command through python3 scripts/run_ragops.py ... from this skill directory.
  5. Interpret exit code 0 as PASS and exit code 2 as an evaluated BLOCK, not an execution failure. Treat other nonzero codes as operational errors.
  6. Report the decision, failed gates, material metric deltas, provenance limitations, and the smallest next action.

Guardrails

  • Do not weaken thresholds, replace the accepted baseline, or omit failing cases to obtain a pass.
  • Do not claim semantic correctness when only lexical or citation checks were run.
  • Do not run provider-backed evaluation unless the user explicitly supplies and authorizes the provider configuration.
  • Do not publish, merge, tag, or deploy based only on this skill's output.
  • Preserve input artifacts and command parameters in the final evidence so another reviewer can reproduce the result.

Output

Return a concise release recommendation with: decision, evidence, failed_gates, limitations, and next_action.

相關技能