Agentic RAG Fallback

What this Claude Code skill does

Designs a Retrieve-Check-Route fallback pattern for agentic (multi-tool) RAG systems so that when every retriever tool returns empty or irrelevant results, execution routes to a base-LLM parametric fallback or a controlled no-answer instead of failing silently or hallucinating. Use when a LangGraph agent that selects among multiple retriever tools has no defined behavior on out-of-corpus queries,

Agentic RAG Fallback is a packaged skill for Claude Code: a SKILL.md plus references and a runnable test that drops into .claude/skills/, so your agent can run this workflow reliably instead of you re-explaining it each time. Install it via MCP, run it as a CLI, or fork it into your own.

Inside the SKILL.md

<h1>Agentic RAG Fallback (Retrieve-Check-Route)</h1>
<h2>Why This Is Hard</h2>
<p>An agentic RAG system reasons over a set of retriever tools and decides which to
call. That design has a hole: there is <strong>no defined behavior when every tool
returns nothing useful</strong>. The agent was built to retrieve-then-generate, and
retrieval just failed. Without an explicit fallback path, one of three bad
things happens:</p>
<ul>
<li><strong>Silent empty response</strong> — the agent returns nothing (or raises) because no
  retriever tool found relevant documents, discarding the base LLM's parametric
  knowledge entirely.</li>
<li><strong>Hallucinated retrieval</strong> — the agent fabricates a context and passes it to
  the LLM as if it were retrieved, producing a confident but ungrounded answer
  with no signal that no real documents were used.</li>
<li><strong>Low-relevance context pollution</strong> — with no relevance check, the
  lowest-scoring chunks are injected anyway, and the LLM blends retrieved noise
  with its own generation unpredictably.</li>
</ul>
<p>The distinguishing factor from single-store fallback: here the <em>agent itself</em>
chose the tools, so a failure is ambiguous — bad query, wrong tool, or genuinely
out-of-corpus. The fallback must be defined at the agent's decision layer, not
just after a single retriever.</p>
…