# Retrieve Kb Answer

> Search the project KB and return either a direct answer with citations or "no high-confidence match".

- Skill: `jcaiagent7143-ui/retrieve-kb-answer` (Agent Skill)
- Install (CLI): `npx skillmds@latest add jcaiagent7143-ui/retrieve-kb-answer`
- Raw SKILL.md: https://api.skillmd.com/api/skills/jcaiagent7143-ui/retrieve-kb-answer/raw
- Safety review: pending (external: skill-scanner PASS, skillspector PASS)
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: jcaiagent7143-ui (https://skillmd.com/u/jcaiagent7143-ui)
- Updated: 2026-09-10
- Page: https://skillmd.com/skills/jcaiagent7143-ui/retrieve-kb-answer

---


# Retrieve KB Answer

## Steps

1. Run vector + keyword hybrid search over the project KB (likely under
   `docs/`, `kb/`, or the README + CONTRIBUTING).
2. Compute reranker confidence. If ≥ `confidence_threshold`, draft a
   reply that cites every claim by source filename + section anchor.
3. If below threshold, return `{reply: null, reason: "low_confidence",
   topics: [...]}` so the next step files a ticket instead.
4. **Never** guess: if the KB doesn't cover it, that's a ticket, not a wing-it.

## Output shape

```jsonc
{
  "reply": "...",                          // or null
  "citations": [{"source": "docs/x.md#hdr", "snippet": "..."}],
  "confidence": 0.84,
  "reason": null                            // or "low_confidence"
}
```

## Failure modes

- Citing a doc you didn't actually retrieve.
- Suppressing low confidence so the answer ships unticketed (looks
  productive; produces angry follow-ups).
- Replying with the user's question rephrased — say nothing rather than
  hallucinating an answer.