RAG Pipeline Vector Store

What this Claude Code skill does

Designs a baseline single-hop RAG pipeline that grounds LLM answers in private or domain-specific documents via vector store retrieval, with emphasis on the ingestion and chunking stage that drives retrieval quality. Use when building a question-answering or chatbot system over documents the base LLM never saw, or when "RAG answers are wrong" turns out to be a chunking or embedding problem. Common

RAG Pipeline Vector Store 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 Pipeline — Vector Store Retrieval (Chunking-First)</h1>
<h2>Why This Is Hard</h2>
<p>A base LLM has no knowledge of private or domain-specific data and cannot answer
questions about content that postdates its training or belongs to a specific
organization. Without a retrieval layer, every domain-specific query reaches an
LLM that is epistemically blind to the answer. Three failure modes recur:</p>
<ul>
<li><strong>Ungrounded hallucination.</strong> Domain-specific queries sent straight to the LLM
  produce confidently wrong answers — the model fills the knowledge gap with
  plausible fiction.</li>
<li><strong>No semantic search without vectorization.</strong> Storing raw text without
  embedding makes semantic search impossible; keyword search cannot match the
  <em>intent</em> of a query, only its surface tokens.</li>
<li><strong>Garbage chunks, garbage answers.</strong> The single biggest quality lever in a
  traditional RAG pipeline is chunking. Bad chunk boundaries split an answer
  across two chunks (neither retrievable alone) or bury the answer in noise.
  Most "RAG doesn't work" reports are chunking problems, not model problems.</li>
</ul>
<hr />
<h2>Core Framework: The Three-Stage RAG Pipeline</h2>
<p>Applicable when domain-specific or private data exists that the base LLM was not
trained on, and answers must be grounded in it. The LLM is invoked <strong>exactly
…