RAG Query Routing

What this Claude Code skill does

Designs and implements LLM-backed query routing across multiple domain-specific vector stores using the retriever-tool pattern, so each query is dispatched to the right store instead of fanning out or hitting one bloated merged store that wrecks the signal-to-noise ratio. Use when building a multi-domain RAG pipeline that must pick the correct vector store from query semantics, and you want the im

RAG Query Routing 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>RAG Query Routing Across Multiple Vector Stores</h1>
<h2>Why This Is Hard</h2>
<p>When an application holds information in multiple domain-specific vector stores,
a static retrieval pipeline cannot determine which store to query. That forces an
unpleasant choice: merge everything into one bloated store, or write brittle
manual routing code. Three failure modes make naive approaches unworkable:</p>
<ul>
<li><strong>Merged-store precision collapse.</strong> Combining all domains into one vector
  store wrecks the <strong>signal-to-noise ratio</strong> of top-k retrieval — unrelated
  chunks from other domains compete for the same slots and pollute the context
  window.</li>
<li><strong>Hard-coded routing fragility.</strong> If/else conditions keyed on query keywords
  break whenever domains are renamed or added, or when users paraphrase or use
  synonyms.</li>
<li><strong>Fan-out latency waste.</strong> Querying every vector store on every request
  multiplies retrieval latency and floods the LLM with irrelevant context from
  the wrong domains — degrading the signal-to-noise ratio a different way.</li>
</ul>
<hr />
<h2>Core Framework: The Retriever-Tool Routing Pattern</h2>
<p>Each vector store is wrapped as a <strong>named retriever tool</strong> with a descriptive
docstring and registered with an LLM agent. The agent inspects the query and
…