Agentic RAG pipeline using Claude agents for query analysis, hybrid retrieval, evaluation, and answer synthesis with CRAG validation.
npx claudepluginhub flashwade03/fablers-ragAgentic RAG pipeline using Claude agents for query analysis, hybrid retrieval, evaluation, and answer synthesis with CRAG validation.
flashwade03Ask your documents. Get a cited answer.
A Claude Code plugin that runs an agentic RAG pipeline — query analysis, hybrid retrieval, evaluation with CRAG validation, and cited answer synthesis — all orchestrated by Claude agents. Supports PDF, plain text, and Markdown.
You have documents. You have questions. But keyword search is fragile and LLMs hallucinate without sources.
fablers-agentic-rag bridges the gap: it chunks your document (PDF, TXT, or Markdown), indexes it with vector + BM25, and deploys a 3-agent pipeline that retrieves, validates, and synthesizes answers with page-level citations — all inside Claude Code.
There are many ways to do RAG with Claude Code. You could wire up Obsidian + an MCP server + a vector DB + a separate AI for summarization. It works — but now you're managing four tools, each with its own setup, updates, and quirks.
Or you could just feed the whole PDF to Claude. But a 600-page book blows past the context window, and even if it fits, you'd burn tokens re-reading it with every question.
This plugin takes a different approach: one tool, one workflow. Ingest, search, validate, answer — all inside Claude Code. The only reason OpenAI is involved at all is that Claude doesn't offer an embedding API (yet). Everything else runs on the Claude you're already using.
| Typical RAG MCP | This Plugin | |
|---|---|---|
| Workflow | Obsidian / vector DB / external AI — multiple tools to manage | Claude Code only — ingest to answer in one place |
| Brain | External LLM API calls (OpenAI, etc.) for reasoning | Claude Code agents ARE the brain — no external LLM |
| Architecture | Single retrieve → paste | Multi-agent pipeline with validation |
| Quality check | None — returns whatever vector search finds | CRAG validation scores every passage, retries with rewritten queries |
| Complex questions | Same path for all queries | Complexity routing — 1 agent for simple, 3 for multi-part |
| Citations | Raw chunk dump or none | Every claim gets [Source N] inline + sources section |
| Search method | Vector-only (misses exact terms) | Hybrid vector + BM25 (catches both semantics and keywords) |
| Infrastructure | Often requires Docker, vector DB server | Zero infra — pure Python files + Claude agents |
| Self-correction | One-shot, no retry | CRAG loop rewrites queries up to 2x when results are poor |
| Whole PDF in context | This Plugin | |
|---|---|---|
| ~50 pages | Works fine. Just read it. | Overkill |
| ~150+ pages | Exceeds context window or costs explode | Index once, query cheaply forever |
| Repeated questions | Full re-read every time (10 questions = 10x cost) | One-time index, ~5K tokens per query |
| Citation accuracy | May hallucinate page numbers | Chunk metadata has exact pages/headings |
The only external API call is OpenAI text-embedding-3-small for query embedding. Everything else — query analysis, reranking, validation, answer synthesis — runs on Claude Code's own agent system. No extra LLM costs.
TL;DR: An MCP gives you search results. This plugin gives you a validated, cited answer — powered by the Claude you're already paying for. Put those tokens to good use.
This plugin was born from a real need: ideating on game design theory from Jesse Schell's The Art of Game Design — a 600-page book full of interconnected concepts, lenses, and frameworks. The goal was never to index a million documents. It was to deeply understand one.
Think of it as a compact, self-contained reactor you drop into a project:
numpy arrays + in-memory BM25 — no vector DB, no server, no Docker. If your data fits in a few files, this is all you need. For enterprise-scale knowledge bases with millions of records, use GraphRAG or a dedicated vector DB solution instead.
plugin/ directory)/search for direct search, /ingest for document indexing